INCIDENCE AND TREATMENT OF BRAIN METASTASES ARISING FROM

LUNG, BREAST, OR SKIN CANCERS: REAL-WORLD EVIDENCE FROM

PRIMARY CANCER REGISTRIES AND MEDICARE CLAIMS

by

MUSTAFA STEVEN ASCHA

Submitted in partial fulfillment of the requirements for the degree of

Doctor of Philosophy

Clinical Translational Science

CASE WESTERN RESERVE UNIVERSITY

May 2019

CASE WESTERN RESERVE UNIVERSITY

SCHOOL OF GRADUATE STUDIES

We hereby approve the thesis/dissertation of

MUSTAFA STEVEN ASCHA, MS

Candidate for the degree of Doctor of Philosophy

Committee Chair

Fredrick R. Schumacher, Ph.D.

Faculty Advisor

Jill S. Barnholtz-Sloan, Ph.D.

Committee Member

Jeremy S. Bordeaux, M.D.

Committee Member

Andrew E. Sloan, M.D.

Date of Defense

March 18, 2019

* We also certify that written approval has been obtained for all proprietary material

contained therein.

2 Dedication

This work is dedicated to: the women in my family who brave, have braved, or pre-empted breast cancer,

whose strength and selflessness inspires and motivates me; to

Nizar N. Zein, MD,

who ten years ago began mentoring me in the discipline I study with this

dissertation; and to

Jill S. Barnholtz-Sloan, PhD

who three years ago helped me begin and later complete this work.

3 Contents

Dedication ...... 3

List of Tables ...... 7

List of Figures ...... 9

List of Abbreviations ...... 10

Acknowledgements ...... 12

Abstract ...... 13

Chapter 1 – Introduction ...... 14

1.1 The changing nature of brain metastases epidemiology ...... 14

1.2 Brain metastases treatment ...... 17

1.2.1 Localized treatment ...... 17

1.2.2 Systemic treatment ...... 20

1.3 Data sources ...... 22

1.3.1 Data descriptions and considerations ...... 22

1.3.2 Data source-specific selection criteria ...... 24

1.4 Methodological rigor in computational and secondary data research ...... 25

1.4.1 Openness and version control ...... 26

1.5 Specific Aims ...... 27

Chapter 2 - Identifying incidence of brain metastases ...... 28

2.1 Abstract ...... 28

2.2 Introduction ...... 30

2.3 Materials and methods ...... 31

2.3.1 Data description ...... 31

4 2.3.2 Identifying patients with brain metastases ...... 32

2.3.3 Study Population ...... 33

2.3.4 Study Variables ...... 33

2.3.5 Analysis ...... 34

2.3.6 Reproducibility ...... 35

2.4 Results ...... 36

2.5 Discussion ...... 38

Chapter 3 - Bevacizumab for the treatment of brain metastases ...... 45

3.1 Abstract ...... 45

3.2 Introduction ...... 47

3.3 Materials and methods ...... 48

3.3.1 Dataset ...... 48

3.3.2 Data-derived definitions ...... 49

3.3.3 Population ...... 50

3.3.4 Treatment patterns ...... 51

3.3.5 Treatment efficacy ...... 51

3.3.6 Rigor and Reproducibility ...... 53

3.4 Results ...... 53

3.4.1 Clinical and demographic characteristics ...... 53

3.4.2 Treatment patterns ...... 53

3.4.3 Survival analysis...... 54

3.5 Discussion ...... 55

3.6 Conclusions...... 58

Chapter 4 - Stereotactic radiosurgery for the treatment of brain metastases . 60

5 4.1 Abstract ...... 59

4.2 Introduction ...... 62

4.2 Materials and Methods ...... 63

4.2.1 Dataset ...... 63

4.2.2 Population ...... 64

4.2.3 Outcomes ...... 64

4.2.4 Definitions and covariates ...... 65

4.2.5 Statistical Analysis ...... 65

4.2.6 Rigor and reproducibility ...... 67

4.3 Results ...... 67

4.3.1 Treatment patterns ...... 67

4.4.3 Survival analysis...... 68

4.5 Discussion ...... 69

Chapter 5 - Conclusions and future directions ...... 73

5.1 Summary and translational impact ...... 73

5.2 Claims data limitations and future directions...... 74

5.3 Brain metastases incidence and treatment ...... 76

Bibliography ...... 145

6 List of Tables

Table 1-1 Frequencies and median survival for patients with synchronous brain metastases ...... 80 Table 1-2 Diagnosis and procedure codes used to identify BM and its treatment... 81 Table 2-1 Medicare claims brain metastasis classification accuracy, SEER- Medicare population 2010-2012 ...... 84 Table 2-2 Diagnosis code-only algorithms performance, SEER-Medicare population 2010-2012 ...... 88 Table 2-3 Incidence of primary cancer associated with brain metastases, SEER- Medicare population 2008-2012 ...... 92 Table 2-4 Lung cancer demographics and clinical characteristics, SEER-Medicare population 2008-2012 ...... 95 Table 2-5 Breast cancer demographics and clinical characteristics, SEER-Medicare population 2008-2012 ...... 98 Table 2-6 Melanoma demographics and clinical characteristics, SEER-Medicare population 2008-2012 ...... 102 Table 2-7 Stage-specific incidence proportions of brain metastases, SEER- Medicare population 2010-2012 ...... 105 Table 3-1 NSCLC patient characteristics by SBM and bevacizumab treatment status, SEER-Medicare population 2010-2012 ...... 106 Table 3-2 Hazard ratios of mortality associated with the use of Bevacizumab, SEER-Medicare population 2010-2012 ...... 110 Table 3-3 Odds of bevacizumab prescription among non small-cell carcinoma patients with brain metastases, SEER-Medicare population 2010-2012 ...... 111 Table 3-4 Odds of bevacizumab prescription among all non small-cell carcinoma patients, SEER-Medicare population 2010-2012 ...... 112 Table 3-5 Hazard ratios of mortality among the NSCLC synchronous brain metastases patients, SEER-Medicare population 2010-2012 ...... 115 Table 3-6 Hazard ratios of mortality in the overall NSCLC population, SEER- Medicare population 2010-2012 ...... 117 Table 3-7 Missing covariate values in the overall NSCLC population, SEER- Medicare population 2010-2012 ...... 120 Table 4-1 Clinical and demographic characteristics by SRS treatment and presence of BM, SEER-Medicare population 2010-2012 ...... 121 Table 4-2 Odds of stereotactic radiosurgery among lung cancer patients with brain metastases, SEER-Medicare ...... 124 Table 4-3 Odds of stereotactic radiosurgery in the overall lung cancer patient population, SEER-Medicare population 2010-2012 ...... 125 Table 4-4 Hazards of mortality for lung cancer patients with synchronous brain metastases, SEER-Medicare population 2010-2012 ...... 127 Table 4-5 Hazards of mortality for the overall lung cancer patient population, SEER-

7 Medicare population 2010-2012 ...... 128

8 List of Figures

Figure 2.1 Cancer registry and Medicare claims data selection, SEER-Medicare population 2008-2012 ...... 130 Figure 2.2 Lung and breast cancer incidence proportions of brain metastases by race, SEER-Medicare population 2008-2012 ...... 131 Figure 2.3 Lung and melanoma cancer brain metastases incidence proportions by sex, SEER-Medicare population 2008-2012 ...... 133 Figure 2.4 Population selection diagram, SEER-Medicare population 2008-2012 134 Figure 3.1 Kaplan-Meier estimates of survival among non small-cell carcinoma lung cancer patients with brain metastases, SEER-Medicare population 2010-2012 ... 135 Figure 3.2 Kaplan-Meier Survival estimates, overall population, SEER-Medicare population 2010-2012 ...... 137 Figure 3.3 Population selection diagram, SEER-Medicare population 2010-2012 139 Figure 3.4 Propensity score distributions among patients with brain metastases, SEER-Medicare population 2010-2012 ...... 140 Figure 3.5 Propensity score distributions among all patients, SEER-Medicare population 2010-2012 ...... 141 Figure 3.6 Standardized mean differences before and after matching for both the overall and brain metastases populations, SEER-Medicare population 2010- 2012 ...... 142 Figure 4.1 Survival following synchronous brain metastasis diagnoses in lung cancer cases from SEER-Medicare 2010-2012 ...... 143 Figure 4.2 Time to stereotactic radiosurgery across race among lung cancer patients diagnosed with synchronous brain metastases, SEER-Medicare 2010- 2012 ...... 144

9 List of Abbreviations

Abbreviation Meaning

AAIR Average Annual Age-Adjusted Incidence Rate

BM Brain Metastases

Car Carcinoma

CI Confidence Interval

CNS Central Nervous System

CPT Current Procedural Terminology

DME Durable Medical Equipment

ER Estrogen Receptor

Her2 Human Epidermal Growth Factor Receptor 2

HMO Healthcare Management Organization

HR Hormone Receptor

ICD-9-CM International Classification of Disease, Ninth Revision,

Clinical Modification

ICD-O-3 International Classification of Diseases, Oncology, 3rd

Revision

IP Incidence Proportion

LBM Lifetime Brain Metastases

MEDPAR Medicare Part A

NCCN National Comprehensive Cancer Network

10 NCH National Carrier Claims

NSCLC Non-Small Cell Lung Cancer

OUTSAF Outpatient standard analysis files

PCI Prophylactic Cranial Irradiation

PPV Positive Predictive Value

PR Progesterone Receptor

SBM Synchronous Brain Metastases

SCLC Small Cell Lung Cancer

SEER Surveillance, Epidemiology, and End Results

SRS Stereotactic Radiosurgery

US United States

WH White Hispanic

WNH White Non-Hispanic

11 Acknowledgements

Funding for CBTRUS has been provided by the Centers for Disease Control and

Prevention (CDC) under Contract No. 2016-M-9030, the American Brain Tumor

Association, The Sontag Foundation, Novocure, AbbVie, the Musella Foundation,

National Brain Tumor Society, the National Cancer Institute (NCI) under Contract

No. HHSN261201800176P, the Zelda Dorin Tetenbaum Memorial Fund, the Uncle

Kory Foundation, and from private and in kind donations. The findings and conclusions in this report are those of the authors and do not necessarily represent the official position of the CDC or NCI.

This work made use of the High Performance Computing Resource in the Core

Facility for Advanced Research Computing at Case Western Reserve University.

This study used the linked SEER-Medicare database. The interpretation and reporting of these data are the sole responsibility of the authors. The authors acknowledge the efforts of the National Cancer Institute; the Office of Research,

Development and Information, CMS; Information Management Services (IMS), Inc.; and the Surveillance, Epidemiology, and End Results (SEER) Program tumor registries in the creation of the SEER-Medicare database.

This work would not have been possible without the gracious support of my dissertation committee. Their efforts enabled my own, and I am forever grateful for their contributions.

12

Incidence and Treatment of Brain Metastases Arising from Lung, Breast, or

Skin Cancers: Real-World Evidence from Primary Cancer Registries and

Medicare Claims

By

MUSTAFA STEVEN ASCHA, MS

Abstract

Brain metastases (BM) are the most frequent type of brain tumor, and cause significant morbidity and mortality. As imaging modalities improve, treatments improve and diagnoses increase in frequency, epidemiological studies of BM and its clinical outcomes become increasingly relevant and potentially beneficial to cancer patients. Here, we use large population-level data from the Surveillance,

Epidemiology, and End-Results (SEER) program combined with Medicare claims to identify measures of BM incidence, claims algorithm concordance with SEER data, treatment patterns, and treatment efficacy for patients with BM arising from primary lung, breast, or skin cancers. We make use of both traditional epidemiological approaches to minimize confounding, such as age adjustment to a standard population, and more advanced methods, such as propensity score matching. By examining the Medicare claims of patients diagnosed with BM, the following chapters shed light on which patients develop BM, how BM patients benefit from an antineoplastic drug that may simultaneously relieve intracranial hypertension, and which patient populations are underserved with respect to BM treatment.

13 Chapter 1 - Introduction

Brain metastases (BM) represent a late stage of disease and reflect a poor prognosis, often indicating only a matter of months may remain for a patient

(Cagney 2017). Though significant advances have been made in understanding metastatic seeding and colonization, such discoveries have not yet been translated into metastasis-targeted therapies, and the effects of many primary cancer treatments on BM remain the subject of investigation (Kienast 2010, Sofietti 2017).

This dearth of evidence is underscored when considering that BM may arise from any primary cancer and are therefore frequent. Breast cancer BM alone may be ten times more common than primary brain tumors, and estimates suggest 168,000 women will have metastatic breast cancer in the US by the year 2020 (Barnholtz-

Sloan 2004, Patchell 2003, Gavrilovic 2005, Mariotto 2017).

As a result of the increasing relevance of BM to practice, in 2010 the North

American Association of Central Cancer Registries (NAACCR) began large-scale data collection on BM diagnosed during primary cancer staging workup (called

“synchronous” BM, or “SBM”). These data were recently released, spurring many new population-level studies of BM (Kromer 2017, Park 2016, Voog 2017, Johnson

2018, Robin 2018). Large-scale studies of BM patients and their treatment outcomes shed light on critical aspects of the epidemiology of these tumors as they relate to diagnosis, treatment and clinical outcomes.

1.1 The changing nature of brain metastases epidemiology

At the population level, the incidence of BM diagnosis is increasing (Frisk

14 2012). Lung, breast, and skin cancers are the most frequently-metastasizing primary cancers, with SBM incidence proportions at 10.8%, 0.4%, and 1.2% for primary lung, breast, and melanoma diagnoses, respectively (Davis 2012, Kromer 2017,

Table 1-1). Similarly, among patients diagnosed with distant metastases from primary lung, breast, and melanoma cancer, the Metropolitan Detroit Cancer

Surveillance System reports that the incidence proportion percentages of BM diagnoses were 28.1%, 14.2%, and 36.8%, respectively (Barnholtz-Sloan 2004).

The trend of increasing incidence proportions with later stages of disease continues; in non-small-cell lung cancer (NSCLC), small-cell lung cancer (SCLC), breast cancer, and melanoma, BM incidence proportions on autopsy are up to 40%, 52%,

42%, and up to 75% compared to 10.7%, 15.1%, 0.4%, and 1.2% for SBM, respectively (Hirsch 1979, Hirsch 1983, Bafaloukos 2004, DiSibio 2008, Dasgupta

1964, Patel 1978, de la Monte 1983, Sampson 1998, Kromer 2017). Table 1-1 demonstrates how these rates compare to those of other primary cancers, illustrating that 67-80% of BM result from lung, breast, and skin cancers and hence why those cancers were selected for investigation in our work presented here

(Palmieri 2012).

BM diagnosis in particular is more frequent due to improved imaging technology and increased resolution, allowing clinicians to identify metastases sooner in the disease process. The feasibility of MRI as a diagnostic tool was first demonstrated at 3.0 Tesla in 1994 by Mansfield et al, shortly after which ultrahigh field strengths and a plethora of other methods to improve cranial imaging emerged

(Mansfield et al 1994, Robitaille 1998). MRI studies to identify BM arose soon thereafter; Peretti-Viton (1998), for example, reported discovering BM in 23% more

15 patients when using magnetization transfer compared to gradient-echo T1-weighted imaging (Peretti-Viton 1998). Since then, cranial imaging has progressed to a point that subtler differences are now detectable both with respect to image resolution and contrast agent tissue distributions. Novel radiotracers have been investigated for earlier detection of BM or to identify pseudoprogression of BM, along with a host of other meaningful advances that may contribute to increased BM diagnosis incidence (Ben-Ami 2010, Kebir 2016).

In addition to rising incidence associated with greater detection, better primary cancer control itself increases the time-at-risk of brain metastases or even physiologically reveals metastatic colonies that would otherwise not have been clinically recognized. This physiological effect is most relevant to therapies that do not penetrate the blood-brain barrier (BBB) and therefore do not reach CNS parenchyma, allowing medication to reach only disease that lies at the margin of the

BBB or near sites of hemorrhage. Trastuzumab is one such large-molecule breast cancer treatment that is not thought to cross the BBB, and thus the effect on extracranial systemic disease is thought to ‘unmask’ BM that would otherwise not have been detected. One meta-analysis found 1.6 times the odds of BM diagnosis among breast cancer patients treated with adjuvant trastuzumab compared to those without adjuvant trastuzumab, potentially suggesting that BM incidence may be a moving target that will require frequent re-assessment as primary cancer treatments progress (Yin 2011). Interactions with new treatments, as well, will be relevant for future work: Stemmler et al (2007) report that the ratio of serum concentrations of trastuzumab to those of cerebrospinal fluid decreases from 420:1 before stereotactic radiotherapy to 76:1 afterwards, suggesting the use of trastuzumab during

16 radiotherapy may in fact better address metastatic disease (Stemmler 2007).

1.2 Brain metastasis treatment

Heterogeneity in BM count, volume, and aggressiveness leads to similarly heterogenous treatment decisions and outcomes, but treatment often involves one or a combination of surgery, stereotactic radiosurgery (SRS), or whole-brain radiation therapy (WBRT), in addition to primary cancer systemic therapy that may help manage cranial disease. Current debate focuses on which combinations and sequences of therapy benefit patients most, unfortunately, clinical trial criteria often exclude patients with BM and the effects of novel therapeutics on BM are therefore often lacking in evidence (Zatloukal 2004, Senan 2016, Tsimberidou 2011).

BM treatment pathways recommended by the National Comprehensive

Cancer Network are stratified into two categories: those for patients with limited BM and those for extensive BM. Whether limited or extensive, however, surgical resection is among the first BM treatments outlined in NCCN guidelines (NCCN

Guidelines V. 1.2018).

1.2.1 Localized treatment

Tumor resection depends on several clinical and patient characteristics that may preclude neurosurgery even among patients with limited BM; these factors include, among others, previous cancer diagnosis, systemic disease progression, or a lack of systemic treatment options that penetrate the BBB (NCCN Guidelines V.

1.2018). Especially for patients whose surgery would offer some extra benefit, such as those with tumors of uncertain origin or who have large-volume tumors imposing a mass effect, surgical management is favored.

Because surgery necessarily varies with anatomic manifestation, surgical

17 team expertise, and institutional practice, a precise population-level understanding of which patients stand to benefit the most from surgery is unfortunately not yet possible (Adams 2016, Flieger 2014). With greater intraoperative data collection allowing for more precise comparisons across institutions, such variability can be minimized and reveal what characteristics predict greater benefit from BM resection.

One example of a characteristic that is easier to reason about is the anatomical accessibility of a tumor; for example, smaller and deeper BM are considered better candidates for stereotactic radiosurgery whereas larger and more superficial BM may be better-managed by neurosurgery (Ewend 2008). In an era when stereotactic procedures or pressure readings from surgical instruments may be recorded, a large-scale description of surgical practices and outcomes may resolve many outstanding questions, and may not be distant.

NCCN guidelines suggest considering surgery in patients experiencing tumor mass effect, with no previous cancer diagnosis, stable systemic disease, or appropriate systemic treatment options. Further, for the population of patients whose systemic treatment options may minimize BM disease burden, NCCN guidelines state it is reasonable to consider systemic treatment as a trial and to wait until BM progression is observed before moving on to surgical management. These systemic treatments would presumably have good CNS penetration and would apply to select patients whose disease may be disseminated or has poorer prognosis, for example melanoma patients with asymptomatic BM but active extracranial disease, or patients with anaplastic lymphoma kinase (ALK)- rearrangement positive NSCLC (Davies 2011, Shaw 2011).

Because up to 60% of patients undergoing surgical resection of BM may

18 experience recurrence at the surgical site, postoperative radiation is used to minimize odds of BM recurrence (Patchell 1998, Kocher 2011, Mahajan 2017).

Cranial radiotherapy can be broadly classified as whole brain radiation therapy

(WBRT) or stereotactic radiosurgery (SRS). Compared to WBRT, the noninvasive nature and highly localized radiation delivery afforded by SRS have led to increased usage trends over time (Moraes 2016, Park 2016). Indeed, NCCN guidelines report a preference for SRS over WBRT except in rare circumstances such as cases with violation of ventricles or suspected meningeal disease. Because SRS treatment is associated with greater risk of neurotoxicity or local recurrence when used for larger tumors, its use as a monotherapy is considered more appropriate than surgery or

WBRT only for cases of small (less than 3cm), solitary, surgically inaccessible BM.

Standard risk factors for radiation necrosis include volume treated, prior treatment history, systemic therapy, and SRS dose, among others (Chao 2013). There are no randomized clinical trials comparing SRS alone compared to SRS plus neurosurgery and observational studies of these treatment approaches are limited both in number and strength of evidence, significantly restricting conclusive recommendations in this context (Loeffler and Wen 2018).

One of the more concerning side effects associated with WBRT is cognitive impairment, which is not as frequently associated with SRS as a result of more focused and specific radiation dosing. Because greater hippocampal radiation dosage is associated with neurocognitive impairment, there is particular interest in

SRS as a way to minimize hippocampal tissue irradiation, though randomized controlled trials of hippocampal avoidance using WBRT for BM are currently underway (Gondi 2013, Awad 2013 , Zhang 2017, Martinage 2018). In contrast to

19 certainty regarding the benefit of hippocampal avoidance, insufficient evidence obfuscates the benefit of SRS versus SRS with WBRT. A Cochrane systematic review of literature through 2017 found three randomized controlled trials of BM patients treated with SRS monotherapy compared to SRS with adjuvant WBRT, and concluded that there is no evidence of an overall survival benefit to SRS with adjuvant WBRT except among patients with high functional status or a single BM, though this evidence is to be interpreted with caution and more randomized controlled studies are necessary to identify a difference in the two treatment regimens (Patil 2017).

1.2.2 Systemic treatment

Systemic therapy for patients with BM is more likely to be a result of primary cancer treatment considerations rather than BM itself, both to manage the primary cancer but also because the blood brain barrier prevents drug delivery to parenchymal tissue from circulation as a result of tight junctions, several cellular layers, and drug efflux transporters that all serve to ensure the only large or hydrophilic molecules entering the brain are those that are transcytosed (Newelt

2004, Abbott 2006, Eichler 2011).

Some chemotherapeutic agents may have non-negligible partition coefficients, however: prospective clinical studies have found detectable levels of capecitabine and lapatinib in resected BM tissue, one-tenth the concentration of vinorelbine has been observed in BM compared to extracranial metastases in mouse models of BM, and BBB penetration of paclitaxel and doxorubicin only reaches cytotoxic concentrations in about 10% of metastases (Hudacheck 2013,

Morikawa 2015, Samala 2016, Lockman 2010). Most notably, temozolomide

20 crosses the BBB and has been the most frequently-used systemic chemotherapy for patients with melanoma BM, and is currently under investigation in combination with novel pharmacologic interventions (Agarwala 2000, Margolin 2012, Hong 2015,

McQuade 2016).

Large BM (those with a diameter greater than 3 centimeters) often disrupt the BBB and allow contrast to leak into the BM on MRI studies, and therefore some patients may have chemotherapeutic benefit despite the BBB (Zhang 1992, Stewart

1994, Arvold 2016). This potential benefit in patients with significant disease burden is complicated by conflicting reports about the relationship between different systemic therapies and radiation necrosis following SRS treatment (DuFour 2012,

Parker 2015). Among treatment-naive patients, BM response rates follow those of extracranial disease, enough that primary cancer chemosensitivity may be as important as BBB penetration to halting intracranial disease (Lesser 1996, Postmus

1999, Lassman 2003, Arvold 2016).

Different mechanisms of action also afford intracranial disease response.

Ipilimumab is an anti-CTLA-4 receptor monoclonal antibody for which the BBB may not be relevant, this is a result of its mechanism of action disinhibiting cytotoxic T lymphocytes and thus stimulating their activity rather than acting directly on cancer tissue. Recently, Long et al (2018) report on their randomized phase II study of ipilimumab plus nivolumab compared to nivolumab, finding 60% (95% CI: 44, 83) 6- month intracranial progression-free survival in 27 patients treated with ipilimumab plus nivolumab compared to 21% (95% CI: 9, 50) in 19 patients treated only with nivolumab. Similarly, Margolin et al (2012) report an open-label phase II clinical trial of ipilimumab in which 24% of patients with asymptomatic BM achieved disease

21 control in the brain over a 32-month follow-up period.

The third chapter of this work focuses on the systemic therapy called

Bevacizumab, an anti-vascular endothelial growth factor (VEGF) monoclonal antibody whose use is associated with reduced vasogenic edema at the same time that its antiangiogenic effects counter perpetually-starved tumor signals to increase perfusion (Levin 2011). This dual mechanism of action is also thought to help reduce radionecrosis following SRS for BM, and so Phase II trials of bevacizumab treatment following SRS are underway (Alliance for Clinical Trials in Oncology

2018). Further, it may not be strictly necessary for Bevacizumab to cross the BBB to exert its effects within the lumen of microvessels or capillaries (Arvold 2016).

1.3 Data sources

1.3.1 Data descriptions and considerations

The Surveillance, Epidemiology and End-Results (SEER) program of the

National Cancer Institute released data regarding the presence or absence of brain metastases at primary cancer diagnosis (called “synchronous” brain metastases), but these data and associated work lack information regarding lifetime brain metastases frequency and metastasis-specific treatments. Though not available to the public, claims data from Centers for Medicare and Medicaid Services (CMS) can be linked to SEER, allowing for the characterization of long-term brain metastases incidence and treatment patterns.

The SEER-Medicare dataset is comprised of SEER cancer registry data linked with CMS administrative claims data, which contains diagnosis, procedure, medical equipment, and prescription drug codes in several formats. To link SEER data to these claims, the SEER program obtains individual identifiers and matches

22 these identifiers with a federal Medicare master enrollment list (Warren 2002).

Medicare claims, however, have low sensitivity and specificity in identifying metastases, and this has traditionally precluded the use of Medicare claims to identify patients with metastases. However, the works cited by the Healthcare

Delivery Research Program (HDRP) in support of this conclusion are limited in that they consider all potential metastasis sites and only using International

Classification of Disease, Ninth Revision (ICD-9) diagnosis codes in cohort selection algorithms (Whyte 2013, Cooper 1999). On the other hand, Eichler and Lamont provided conflicting evidence for brain-specific metastases identification (Eichler and

Lamont 2009), for whom diagnosis code presence had sensitivity of 100%, specificity of 97%, and positive predictive value of 91%. With three or more diagnosis codes, these values were 97%, 99%, and 98%, respectively.

As a result of the issues associated with using claims data to identify BM cases, one of the aims of the following projects was to replicate claims concordance at the population level. By calculating concordance between Medicare claims, the population-level BM classification performance of which is not well-characterized, with the gold standard diagnosis provided in SEER data, we simultaneously established a measure of accuracy for our own estimates of incidence and created a generalizable context for future research that uses SEER synchronous data. This approach allowed us to extrapolate lifetime BM incidence proportions from the SBM incidence proportions that are increasingly reported due to the recent release of

SBM status in cancer registry data.

To identify synchronous treatments, procedure codes must have been observed within a window of 60 days before or after primary cancer diagnosis. Not

23 all treatments were common enough to report in our population, in particular, there were very few elderly patients who have administrative claims indicating neurosurgery procedures. The majority of the Medicare population is over the age of

65, and previous work shows that increased age is associated with worse enough outcomes that age should be considered when deciding whether to perform surgery

(Noordijk 1994).

Medicare data arrive as Medicare-centered arrays of files rather than patient- centric tables. Patient-centric tables might be classified as ‘observations’ or ‘events’, for example, to distinguish the relationship of that record with the patient in that record. Medicare data that arrive with SEER-Medicare materials are instead organized as sets of the following files: (1) Medicare Provider Analysis and Review

(MEDPAR) files for Medicare Part A (inpatient claims), (2) National Claims History

(NCH) files for Part B Medicare data, (3) hospital outpatient files (OUTSAF) for outpatient procedures, (4) the Chronic Condition Flags (CCW) file, (5) Part D Event

(PDE) and (6) Durable Medical Equipment file (DME).

MEDPAR, NCH, and OUTSAF files were used to identify diagnoses or treatment as presented. The Part D Event (PDE) and DME files will be used to identify any non-IV chemotherapies or other drugs used by the diagnosis set.

Intravenous chemotherapy CPT, ICD-9-CM, and HCPCS codes are provided in

Table 2. This work used non-Intravenous chemotherapy drug and procedure codes that are listed in Table 1-2, in addition to chemotherapy look-up tables provided by the National Cancer Institute.

1.3.2 Data source-specific selection criteria

Investigators routinely rely on SEER and SEER-Medicare data, but these

24 data do have limitations (Nattinger 2004, Gold 2007, Koroukian 2003). The use of

Medicare administrative claims requires several exclusion criteria and careful consideration of any codes used to characterize patients. We used the following as criteria to choose our population for study:

 Medicare entitlement must be due solely to age (ie age 65 and older);

 Patients with End-Stage Renal Disease (ERSD) or disability must

also be excluded due to changes in Medicare qualification associated

with those conditions

 Patients must also not have been part of a Health Maintenance

Organization (HMO) within the six months before and twelve months

after primary diagnosis, to avoid the impact of competing insurance

coverage on administrative claims

 Observations where race, sex, follow-up time, or date of death is

missing or unknown were excluded from analysis, and patients who

have more than one primary diagnosis of any cancer or were

diagnosed only on autopsy were excluded from these studies.

1.4 Methodological rigor in computational and secondary data research

Given the development of open source tools to support reproducible research, this work was performed so that all code may be re-run by individuals with approved access to similarly-formatted data and the same versions of software as reported. Such an endeavor only became available in the past few years because free and open source packages to automate statistical reporting (e.g. rmarkdown and knitr, or Jupyter notebooks) have reached a point of maturity and wide adoption,

25 and are particularly relevant to reproducibility in studies making use of claims data.

1.4.1 Openness and version control

Maintaining a clear record of the version of a project greatly improves both transparency and ease of reversion. First started by Linus Torvalds for the purposes of software development, git is a version control system with a rich set of commands that allow a research developer to manage different versions of a project directory and its contents. The most frequently-used git commands in the projects herein were ‘add’, ‘rm’, and ‘commit’; these allow a research developer to ‘commit’ to add/change or remove project files, where the ‘commit’ itself becomes a reference point for future iterations of the project. For the present work, this approach served a dual purpose: ‘commits’ were further uploaded to a remote server such as GitHub

(with a configuration file to specify what files are to be ignored), creating a remote backup of the project. Second, in the event a critical file is mistakenly removed from the project, git may be used to restore that file from a specific commit. With informative messages to label each commit, the researcher may inspect git logs to identify noteworthy time-stamped commits and restore project files from them as- needed. Git is far more complex than this brief description and is grounded in a field of mathematics called graph theory, refer to its extensive documentation for more information.

Besides serving as a remote backup, GitHub acts as a way to disseminate research projects for code or methods reuse. For each of the three manuscripts presented here, the project folder used to produce results for that manuscript are openly available online and also minted with a Document Object Identifier (DOI) for unambiguous reference in the manuscript (e.g.

26 https://doi.org/10.5281/zenodo.1300052 for Chapter 2 of this dissertation). Ideally, researchers who have approved access to SEER-Medicare data will be able to reproduce our work or modify code to suit their own needs.

1.5 Specific Aims

The specific aims for this work, then, are to:

1. Determine lifetime incidence of brain metastases according to the presence of

Medicare diagnosis codes for metastasis to the central nervous system in combination with a brain or head procedure diagnostic imaging or biopsy test, as stratified by primary cancer. There is significant disagreement regarding the frequency of brain metastases, where autopsy studies consistently report greater frequency than studies using other methods.

2. Validate Medicare claims identification of BM against SEER-reported BM.

Considering the proportion of the US population whose cancer diagnoses are reported to SEER, about 29%, along with 97% Medicare coverage for people over

65 years of age, SEER-Medicare studies are highly regarded for strength of evidence. We will thus provide a context for future research regarding the adequacy of Medicare claims for identifying brain metastases by using SEER-reported synchronous brain metastases as a gold-standard and the proportions of patients whose brain metastases are not identified at primary cancer diagnosis.

3. Investigate BM treatment patterns, efficacy and disparities. Understanding current patterns of care from a top-down perspective rather than through the lens of institutional guidelines reveals variation and decision-making that occurs at the population-level. Here, we examine one systemic therapy (bevacizumab) and one localized therapy (stereotactic radiosurgery) with respect to SBM status.

27 2. Lifetime occurrence of brain metastases arising from lung, breast, and skin cancers in the elderly: A SEER-Medicare study

Accepted for publication: Ascha MS, Ostrom QT, Wright J, Kumthekar P, Bordeaux

JS, Sloan AE, Schumacher FR, Kruchko C, Barnholtz-Sloan JS. “Lifetime

Occurrence of Brain Metastases Arising from Lung, Breast, and Skin Cancers in the

Elderly: A SEER-Medicare Study”. Cancer Epidemiology, Biomarkers, and

Prevention, May 2019. doi: 10.1158/1055-9965.EPI-18-1116. In press.

2.1 Abstract

Background: The SEER program recently released data on brain metastases (BM) diagnosed during primary cancer staging workup (“synchronous” BM, or SBM); this study examines the incidence of SBM compared to that of lifetime BM (LBM) identified using Medicare claims for patients diagnosed with lung cancer, breast cancer, or melanoma.

Methods: Incidence proportions (IP) and age-adjusted rates for each of SEER SBM and Medicare LBM are presented along with measures of concordance between the two sources of data, where Medicare LBM were defined by several combinations of diagnosis and putative diagnostic imaging procedure codes.

Results: The SBM IP in lung, breast, and melanoma cancers were 9.6%, 0.3%, and

1.1%, respectively; the corresponding LBM IP were 13.5%, 1.8%, 3.6%. The greatest SBM IP among lung cancer patients was 13.4% for non-small cell lung cancer, and among breast cancer patients was 0.7% for triple-negative breast cancer. The greatest LBM IP among lung cancers was 23.1% in small-cell lung

28 cancer, and among breast cancers was 4.2% for cases of the Triple Negative subtype.

Conclusions: Using a large dataset that is representative of the elderly population in the US, these analyses estimate synchronous and lifetime incidence of BM in lung cancers, breast cancers, and melanomas. These and other population-based estimates may be used to guide development of BM screening policy and evaluation of real-world data sources.

29 2.1 Introduction

Brain metastases (BM) are associated with significant morbidity and may impart a median survival of two to fourteen months following simultaneous diagnosis of BM and various primary cancers (Cagney et al 2017). Estimates of its frequency in the US vary by orders of magnitude depending on study-specific characteristics, but some of this variation in epidemiologic measures of BM may also reflect medical advances in primary cancer control or metastasis detection. This, combined with an aging population and subsequent increase in absolute numbers of patients with brain metastases, suggests that detailed population-level evidence of BM occurrence in the elderly is more and more relevant to clinical practice and the decision to screen for cranial disease (Werner CA 2010, Loeffler 2018).

In 2016, the Surveillance, Epidemiology, and End-Results (SEER) program helped address the need for such studies by releasing data on BM diagnosed during staging workup for primary cancer, called “synchronous brain metastases” (SBM)

(CS Data Collection System 2017). Synchronous metastasis evaluation is helpful to many treatment decisions, and SEER SBM studies performed with these recently- available data contribute valuable information to the pressing need for such work

(Kromer 2017, Martin 2017).

Metastasis following primary cancer staging may also be studied using

Medicare claims, which offer long-term evidence of patient characteristics and may be linked to SEER data. The SEER-Medicare linkage further allows SEER data to serve as a referent to compare with Medicare claims, the classification performance of which may inform Medicare claims-based estimates of lifetime brain metastases

(LBM).

30 The present work examines cases of lung cancer, breast cancer, and melanoma due to their increased incidence of BM (Fox 2011). Motivated by the availability of new data and a need for estimates over the patient’s lifetime, our aim was twofold: to (1) evaluate Medicare claims BM identification algorithms against

SEER SBM data, and (2) report incidences of SEER SBM and Medicare claims- identified LBM for lung cancers, breast cancers, and melanoma.

2.2 Materials and Methods

This study was approved as exempt of review by the University Hospitals

Cleveland Medical Center Institutional Review Board, and assigned the study number “EM-17-05.”

2.2.1 Data description

The National Cancer Institute SEER program offers cancer registry data covering approximately 28% of the US population; these data may be linked to

Medicare claims data to further investigate patient characteristics. Because

Medicare is the primary insurer for the vast majority of patients age 65 years or older, the results of SEER-Medicare studies are highly generalizable to the elderly population (Warren 2002).

Data for primary cancers diagnosed in the years 2008 through 2012 were linked to Medicare data from 2007 through 2014. SEER SBM data are only available for the year 2010 and onwards, therefore only those years were used to compare

Medicare claims identification algorithms. This yielded a total of three diagnosis years for SEER SBM comparison to Medicare identification algorithms and five diagnosis years for Medicare LBM estimates (Figure 2-1).

Four types of claims files offered as part of SEER-Medicare were used: Part

31 A inpatient, carrier, outpatient, and durable medical equipment files. Each record in these files contains a date of service, International Classification of Diseases, ninth revision, Clinical Modification (ICD-9-CM) diagnosis codes, and Current Procedural

Terminology (CPT) procedure codes.

2.2.2 Identifying patients with brain metastases

The 2010 - 2013 SEER Collaborative Staging brain metastases variable uses clinical or pathologic evidence from staging workup (but not thereafter) to identify “distant metastatic involvement of the brain at the time of diagnosis…This includes only the brain, not spinal cord or other parts of the central nervous system

(CNS) (SEER 2017).” In contrast, the closest ICD-9-CM code (198.3X) refers to metastases to any part of the CNS and is not specific to the brain. Therefore,

Medicare LBM was defined as the presence of diagnosis codes for secondary cancer of the CNS (ICD-9 code: 198.3X) and procedure codes for a brain or head imaging study (CPT codes 70450 – 70470, 70551 – 70553, 78607 – 78608) each within 60 days of the other, at any time throughout Medicare claims data.

To inform the reliability of Medicare LBM estimates, classification performance metrics (Cohen’s Kappa, predicted and true counts, sensitivity, specificity, and positive predictive value) are presented for both the Medicare LBM algorithm and a Medicare SBM algorithm based on the same codes with the additional requirement that codes occurred within 60 days of primary cancer diagnosis. This duration of 60 days was chosen as an approximation of the time required to clinically establish extent of disease, after which registry data regarding metastases are not updated (CS Data Collection System 2017). Classification performance measures for a modified Medicare LBM algorithm that relies only on

32 diagnosis codes are also provided for comparison.

2.2.3 Study Population

The use of Medicare data has several implications for population selection.

Exclusion criteria included age less than 65 years, Medicare qualification not due to age, lack of insurance or unknown insurance status, any record of Healthcare

Management Organization (HMO) use, or presence of other primary cancer diagnosis in a different site. Before exclusion, there were 198,730 lung, 208,909 breast, and 98,299 melanoma cases. After exclusion, 120,405 lung, 110,983 breast, and 35,268 melanoma cases remained (Figure 2-4). In the cohort of patients diagnosed in 2010 through 2012, there were 70,974 lung, 67,096 breast, and

21,860 melanoma cases.

2.2.4 Study Variables

Age and race were described across BM status. Age was summarized both as a continuous and categorical variable, where categories were age groups 65 -

69, 70 - 79, 80 - 89, and 90+ years. Race was categorized as Asian/Pacific Islander,

African American, Native American, White, or other. “White” was further categorized as “White Non-Hispanic” (WNH) or “White Hispanic” (WH) using the North American

Association of Central Cancer Registries’ Hispanic Identification Algorithm. For breast cancers, Native American and Asian/Pacific Islander race were collapsed into an “other” category; for melanoma, Native American/American Indian, Black/African

American, and Asian race categories were collapsed.

Histology was categorized according to the International Classification of

Diseases, Oncology, 3, (ICD-O-3) (NCI 2018). The five most frequent histology labels among each cancer site were identified, and less frequent labels were

33 considered “other” categories. For lung cancer, the most frequent histology was adenocarcinoma (8140-8141, 8144), followed by squamous cell carcinoma (8070-

8075, 8078), small cell carcinoma (SCLC, 8041-8045), and non-small cell carcinoma (NSCLC, 8046), and carcinoma (8010-8014). Among breast cancers, the most frequent histology was duct carcinoma (8500-8504, 8507), followed by lobular and other ductal carcinoma (8520-8525), mucinous adenocarcinoma (8480-8481), and adenoid cystic & cribriform carcinoma (8200-8201). Melanoma histology was categorized as nevi & melanomas (8720-8723) and malignant melanoma in junctional nevus (8740-8745) due to a lack of observations of other histology groups.

The derived SEER Summary Stage 2000 data element was used to classify extent of disease at primary diagnosis as in situ, localized, regional with direct extension, regional with extension to lymph nodes only, regional with direct and lymph node extension, regional not otherwise specified (NOS), distant, or unknown.

Breast cancer subtypes were categorized using the 2010+ breast subtype

SEER data element, which is a combination of estrogen receptor (ER) status, progesterone receptor (PR) status, and HER2 (also called CD340 or neu) values. It has six categories: Her2+/HR+, Her2+/HR-, Her2-/HR+, Triple Negative, unknown, or not 2010+ Breast, where HR represents ER and PR status. For measures of incidence, categories were consolidated into Her2+, Her2-/HR+, Triple Negative, and unknown.

For privacy and confidentiality purposes, statistics are suppressed or levels collapsed for categorical values representing data from 1 to 11 subjects (NCI 2018).

2.2.5 Analysis

34 Concordance was measured using Cohen’s Kappa. Sensitivity, specificity, and positive predictive value (PPV) are presented for further detail, along with predicted and true counts of BM cases.

Incidence proportion (IP) was defined as the ratio of brain metastases case counts to the total number of cases, presented for each primary cancer and its associated strata. Strata were histological for lung cancer and melanoma, and based on molecular subtype for breast cancers. Because molecular subtype was only available from cases in 2010 onwards, stratified estimates of Medicare LBM were restricted to the 2010-2012 diagnosis years (Figure 2-1). Additionally, analysis included estimates of the IP of primary cancer later associated with Medicare LBM diagnosed in 2008 through 2012, stratified by SEER Summary Stage 2000.

Descriptive statistics of demographic and clinical characteristics are presented for the overall population, patients diagnosed with SEER SBM, and patients classified as having Medicare LBM.

The average annual age-adjusted incidence rate (AAIR) was calculated as an age-weighted sum over each year using census values from 2010 for each five- year-interval of ages starting at sixty-five and combining ages 90 or greater into a single interval (Keyfitz 1966). Crude incidence rates are presented for breast cancer strata where fewer than 11 cases per annum were observed.

2.2.6 Reproducibility

This study was conducted with the goal of providing open, reproducible, and replicable results (Stodden 2013). Analyses reported here were performed using R version 3.4.3 (2017-11-30) and managed using GNU make (R Core Team 2017,

Stallman 2004); reproducibility materials are available online at

35 http://doi.org/10.5281/zenodo.1336604 (Ascha 2018), and two authors independently used these materials to arrive at the same conclusions presented herein.

2.3 Results

There were 6,789 cases of SBM among lung cancer subjects, 203 among breast cancers, and 230 melanoma SBM cases using the SEER gold standard data, while 5,100, 196, and 152 subjects were found to have Medicare SBM, respectively.

Using the Medicare claims algorithm intended to identify SBM based on both diagnosis and cranial imaging codes, Kappa concordance with SEER-reported SBM was found to be 0.63 (95% CI: 0.62 - 0.64) for lung cancer, 0.46 (95% CI: 0.40 -

0.53) for breast cancers, and 0.66 (95% CI: 0.60 - 0.71) for melanomas (Table 2-1).

The least restrictive algorithm, relying on the presence of a diagnosis code for central nervous system metastases at any time throughout the patients’

Medicare claims history, yielded concordance of 0.73 (95% CI: 0.72-0.74) for lung cancers, 0.49 (95% CI: 0.43 – 0.54) for breast cancers, and 0.78 (95% CI: 0.73 –

0.82) for melanomas (Table 2-2).

120,405 lung cancer cases diagnosed between 2008 and 2012 were selected for inclusion (Figure 2-1), 70,974 of which were diagnosed in 2010 through 2012.

For cases of primary lung cancer diagnosed in 2010 through 2012, the AAIR of SBM was 9,422 per 100,000 cases (95% CI: 9,034 - 9,825) with an IP of 9.6%

(Table 2-3). Medicare LBM was calculated for the 120,405 patients with primary diagnosis from 2008 through 2012. the LBM AAIR was 13,255 per 100,000 cases

(95% CI: 12,798 - 13,727) and the IP of Medicare LBM in lung cancer overall was

13.5%.

36 The IP of SEER SBM among adenocarcinoma lung cancer patients was

11.8%, while the corresponding Medicare LBM value was 15.5%. For patients with carcinoma histology lung cancer, the SEER SBM IP was 11.1%, compared to

Medicare LBM at 11.9%. The proportions of SEER SBM and Medicare LBM among squamous cell lung cancer patients were 4.6% and 8.1%, respectively.

Patients with NSCLC had a SEER SBM IP of 13.4%, while the corresponding LBM proportion was 15.3%. Lung cancer cases with SCLC histology had an SBM IP of 13.5%, with the IP of LBM in these cases being 23.1%.

When stratified by stage at primary cancer diagnosis, the IP of Medicare

LBM was 18.8% among lung cancer cases with distant metastasis; 10.5% among cases with metastasis to lymph nodes only; and 4.8% among cases that were restricted to tissue local to the disease (Table 2-4).

110,983 subjects with breast cancer were selected for inclusion (Figure 2-1), of whom 67,362 were diagnosed in 2010 through 2012. The AAIR of SEER SBM was 309 per 100,000 cases (95% CI: 239- 397), and the IP of cases with SBM diagnosis was 0.3% (Table 2-3).

A total of 110,983 patients were diagnosed with primary breast cancer from

2008 through 2012. The Medicare LBM AAIR for primary breast cancer diagnoses in this population was 1,790 per 100,000 cases (95% CI: 1,618- 1,979), while the overall IP of primary cancer diagnosis that is associated with Medicare LBM is 1.8%.

The IP of SEER SBM was 0.2% among subjects with Her2-/HR+ breast cancer, compared to 0.5% for Her2+/HR(+/-) and 0.7% for Triple Negative cancers.

The IPs of Medicare LBM were greater, at 1.1%, 3.1%, and 4.2% for each of Her2-

/HR+, Her2+/HR(+/-), and Triple Negative breast cancer, respectively.

37 The incidence proportion of Medicare LBM among patients with distant disease at primary diagnosis was 13.6%, compared to 4.0% in patients with lymph node or regional involvement; 2.0%, with only regional tissue involvement; 2.4% with only lymph node involvement; and 0.9% with localized disease (Table 2-7).

35,268 subjects diagnosed with melanoma in 2008 through 2012 met inclusion criteria (Figure 2-1), 21,860 of these subjects were diagnosed in 2010 through 2012 (Table 2-6).

The AAIR of SBM in the 2010-2012 cohort was 1,049 per 100,000 cases

(95% CI: 826- 1,318), and the IP of cases with SBM diagnosis was 1.1% (Table 2-

3). The Medicare LBM AAIR in the 2008-2012 melanoma cohort was 3,583 per

100,000 cases (95% CI: 3,154- 4,059), while the IP of primary cancer diagnosis later associated with Medicare LBM was 3.6%.

When stratified by stage at primary diagnosis, 30.4% of cases with distant disease at primary diagnosis were later associated with Medicare LBM; 15.2%, regional and lymph node involvement; 13.2%, lymph node involvement only; 7.8%, regional tissue involvement; and 2.5% among cases with localized disease at primary diagnosis (Table 2-7).

2.4 Discussion

This study offers population-based estimates of BM incidence at two intervals during the course of disease, while also describing the reliability of the estimates and providing full reproducibility thereof. Given continued advances in primary site control coupled with the use of therapies that may affect metastatic behavior, such a snapshot may prove a valuable reference (Shao 2011, Paex-Ribes

2009).

38 In addition, population-based estimates of BM occurrence among Medicare recipients are useful because the 9.7% increase in the US population 65 years of age or older between 2000 and 2010 may continue with concomitant increases in absolute measures of cancer and therefore BM incidence (Werner 2010). The convergence of biomedical advances with an aging population underscores the need for studies of secondary cancer.

Using the SEER population whose primary cancer diagnosis was made in the years 2010 through 2013, Cagney et al report that the IP of SEER SBM in SCLC was 15.8%; adenocarcinoma, 14.4%; and NSCLC, 12.8% (Cagney 2017). Similarly, the corresponding SEER SBM IP for this study’s 2010 - 2012 Medicare population were 13.5%, 11.8%, and 13.4%, for SCLC, adenocarcinoma, and NSCLC, respectively. Though these differences in IP may be attributable to population selection procedure, the most notable of which is our restriction to the elderly population due to the use of Medicare claims, their approximate similarity lends credence to each estimate.

The greatest difference between synchronous and lifetime BM incidence proportions was observed in SCLC, for which the Medicare LBM IP was 21.7%. In contrast, the Medicare LBM IP among NSCLC patients was found to be 15.3%; relative to each other, these proportions are consistent with literature describing the occurrence of SCLC and NSCLC metastases (Little 2007). Further, these increased proportions support National Comprehensive Cancer Network (NCCN) guidelines that suggest brain magnetic resonance imaging (MRI) with contrast should be a part of pretreatment evaluation for people diagnosed with Stage II-IV and high-risk Stage

1B NSCLC (NCCN 2018). Moreover, 4.8% of cases considered localized and 10.5%

39 of cases with lymph node involvement were found to be associated with Medicare

LBM, supporting the use of MRI to evaluate patients prior to treatment (Table 2-7).

For patients not diagnosed with SBM, Nussbaum et al (1996) report that the median durations from NSCLC and SCLC diagnosis to presentation with BM were 3 and 6 months, respectively, with a median 10 months from BM presentation to death for each histology (Nussbaum 1996). This progression is not untreatable, however, and prophylactic cranial irradiation following lung cancer diagnosis has been shown to reduce later risk of BM (Gore 2011, Slotman 2007).

Still, at present, deferred occurrence of BM does not address concern regarding the benefit to survival of PCI versus close monitoring and quick response

(Hochstenbag 2000, Gregor 1997, Takahashi 2017). Ideally, improved detection of potentially cancerous tissue combined with precision non-surgical treatment such as stereotactic radiosurgery will lower the threshold to benefit from prophylactic cranial irradiation.

Previous work has also shown that ethnic minorities have been more likely to present with late-stage lung cancer (Halpern 2008, Haiman 2006). In NSCLC and

SCLC, Black/African American cases had greater IPs of SEER SBM compared to

WNH or WH, though this did not extend to Medicare LBM (Figure 2-2a). Males had a greater IP of SEER SBM in both NSCLC and SCLC (Figure 2-3a), which is notable given the median survival following brain metastasis diagnosis has been reported to be decreased among males (Videtic 2009).

Previous work shows that IP of SEER SBM in the overall population of breast cancer patients is estimated to be 0.41%, which is higher than the observed

IP of SEER SBM in the Medicare breast cancer population here (0.3%) (Kromer

40 2017, Martin 2017). This difference is likely multifactorial, however one explanation could be that it reflects differences in study populations. Excluding patients younger than 65 years of age removes a population that is at greater risk of BM, as

Barnholtz-Sloan et al note: “the highest IP% for brain metastases in individuals with primary breast cancer occurred in the youngest age category (IP%, 10%; age at diagnosis, 20 to 39 years) (Barnholtz-Sloan 2004).”

Though not restricted to the Medicare population, Martin et al report that 361 out of 162,078 (0.22%) Her2-negative breast cancer cases had SBM (Martin 2017), which is very similar to the corresponding proportion observed here, 0.2% (Table 2-

3). Combining their reported Her2+/HR+ and Her2+/HR- categories, Martin et al report 242 out of 32,095 (0.75%) Her2-positive subjects developed brain metastases in the overall SEER population, which is slightly greater than the corresponding 31 out of 5,900 (0.5%) found in the SEER-Medicare Her2+ population examined here (6) (Table 2-5).

Such associations with hormonal receptor expression status are well- documented (Hicks 2006, Garcia 1992). In our elderly population, 13.4% of cases were ER negative; in contrast, ER-negative breast cancers comprised 28.4% of cases in the SEER SBM and 34.6% in the Medicare LBM populations. Extending this effect seen in cancers lacking hormone receptor expression, the greatest LBM rates were observed among patients with the Triple Negative (Her2-/ER-/PR-) molecular subtype: the SEER SBM IP among Triple Negative patients was 0.7%, while the Medicare LBM IP was about 5 times greater at 4.2% (Figure 2-2b).

Multiple sources report Triple Negative cancer patients diagnosed with SBM have median survival of about 6 months (95% CI: 2.0 - 20.0), in contrast to median

41 survival following HER2-/HR+ and HER2+/HR+ SEER SBM at 14.0 and 21.0 months, respectively (Martin 2017, Lin 2008, Dawood 2009).

Previous studies have demonstrated that male sex is associated with a greater IP of melanoma BM, which was the case for IP of both SEER SBM and

Medicare LBM in the population studied here (Bedikian 2011) (Figure 2-3b). Kromer et al showed that the IP of SEER SBM in melanoma cases was 1.2% for the general population, which is slightly greater than the IP observed in the elderly population here (1.1%) (Kromer 2017).

Whereas NCCN guidelines recommend cranial MRI for stage III (regional) and IV (metastatic) melanoma (NCCN 2018), our results suggest cranial evaluation may be warranted even for localized disease. Of Medicare LBM cases, 34.6% were considered localized at the time of primary cancer diagnosis (Table 2-6).

Considering the broader population, 2.6% of localized melanoma cases later had evidence of Medicare LBM, a proportion that increases to 30.4% among patients with distant disease (Table 2-7).

Because the ICD code 198.3 describes metastasis to the CNS rather than the brain, we used this CNS metastasis code coupled with an intracranial imaging procedure to help ensure localization to the brain. The use of claims data is also problematic because BM is not itself tied to reimbursement, though some investigators report excellent classification performance compared to retrospective chart review (Eichler 2009). The present work reports several BM identification algorithms that yield either sensitivity or PPV exceeding 0.8, but none with both sensitivity and PPV that simultaneously exceeding 0.8 (Table 2-1, Table 2-2).

Though insufficient for precise, individual-level prediction, estimates of frequency

42 appeared consistent or comparable across data sources, suggesting claims data may reasonably be used to evaluate incidence of metastatic disease.

Two significant limitations to this study are (1) generalizability of the

Medicare population to the population at large and (2) accuracy of CNS metastasis diagnosis and imaging procedure codes to detecting BM.

Because BM IP peaks for cancers diagnosed at around 60 years of age, our estimates in the population that was age-eligible for Medicare must be cautiously applied to future work (Bedikian 2011). Further, the occurrence of BM arising from breast cancers can greatly exceed the average four years of Medicare claims follow- up available here, though the interval from lung cancer primary diagnosis to BM is well-covered by available claims follow-up. Despite these potential issues, confirmation of the decreased IP of SBM in elderly patients is itself a generalizable conclusion because Medicare covers approximately 97% of the population 65 years or older (Warren 2002, Cooper 1999).

Careful consideration must be given to the use of claims data for metastasis research (Cooper 1999, Whyte 2015, Nordstrom 2012). The present work addresses this need by purposefully omitting statistical tests of difference, and by providing concordance estimates of Medicare algorithms identifying BM. This is the first study to examine Medicare concordance with a SEER gold standard for BM, thus providing context for many future Medicare claims-driven studies of cranial disease following cancer diagnosis, but such highly-reliable data as SEER are not always available. Studies of Medicare claims (part of what is now widely-agreed to be ‘real-world’ data (Sherman 2016)) may benefit from similar evaluation of codes used to derive patient characteristics.

43 One significant strength of this work is the availability of documented analysis code, ideally enabling study replication with a single command (“make”)

(Stallman 2004). To the best of our knowledge, this is the first such open and reproducible SEER-Medicare work.

The results of this study approximately agree with previous population-based estimates of BM incidence, and rigorously shed light on which disease may warrant closer monitoring. Despite increases in the incidence of brain metastases, the majority of primary cancer diagnoses have no standard of care for brain metastasis screening (CS Data Collection System 2017). The development of screening guidelines based on such population-level evidence could help to identify many patients’ cranial disease sooner in its progression, ultimately allowing earlier treatment and improved patient outcomes.

44

3. Bevacizumab for the treatment of patients with synchronous brain metastases arising from non-small cell lung cancer: a SEER-Medicare study

3.1 Abstract

Introduction: Bevacizumab is FDA approved in the treatment of primary brain tumors, but its effects in patients with brain metastases is unclear. This study examines a population of non-small cell lung cancer (NSCLC) patients overall and with synchronous brain metastases to identify predictors for the decision to use bevacizumab and survival following bevacizumab treatment.

Methods: Primary cancer registry data were used to determine which NSCLC patients diagnosed in the years 2010 through 2012 had synchronous brain metastases at the time of diagnosis, and Medicare claims from 2007 through 2014 to identify a population of patients treated with bevacizumab. Multivariable model covariates were selected according to the least absolute shrinkage selection operator, and 1:1 propensity score matching was used to identify predictors of the decision to use bevacizumab and its effect on survival.

Results: After adjusting for clinical and demographic characteristics, bevacizumab was associated with 0.89 times the hazard of mortality in the elderly NSCLC population (95% CI: 0.81 – 0.96, p: 0.003) and a corresponding hazard ratio of 0.75 in the population of elderly NSCLC patients with synchronous brain metastases

(95% CI: 0.59-0.96, p: 0.020).

Conclusions: Bevacizumab may benefit NSCLC patients with synchronous brain metastases more than it does patients without such intracranial disease, this may be

45 a result of its mechanism of action in inhibiting neoangeogeneis and minimizing vasogenic edema which is far more deleterious in the brain than in other organ systems

46 3.2 Introduction

Bevacizumab, a monoclonal antibody that inhibits angiogenesis by neutralizing Vascular Endothelial Growth Factor A (VEGF-A), improves progression- free survival for Glioblastoma Multiforme and is approved by the Food and Drug

Administration for this primary brain tumor (Cloughesy 2008, Gilbert 2014). Whereas

VEGF-A expression normally directs angiogenesis in response to trauma, exercise, or for vasculogenesis in embryonic development, increased expression in tumors is pathological and simply increases the nutrient supply to tissue that will use it to grow exponentially. VEGF-A is the primary VEGF in its sub-family of platelet-derived growth factors, therefore this therapy directly impedes tumor growth by blocking a crucial oncogenic signal.

Bevacizumab reduces the increase in vascular permeability that is associated with VEGF expression and consequently helps relieve patients of the potentially serious morbidity and symptoms that accompany peritumoral edema (Gil-

Gil 2013). This is particularly important in the CNS as edema can lead to increased intracranial pressure in the fixed volume of the cranial vault, which has potentially fatal consequences. Intracranial edema is frequently managed using corticosteroids, but bevacizumab both targets tumor growth mechanisms and simultaneously relieves symptoms, potentially sparing a patient of the effects of corticosteroids while also addressing the underlying disease (Socinski 2009, Dietrich 2012, Besse

2010, De Braganca 2010).

Though several studies have addressed the safety of bevacizumab treatment for brain metastases (BM), its efficacy for this purpose is less well- explored: one meta-analysis reports that, of 57 anti-VEGF treatment studies, 76%

47 explicitly stated the presence of central nervous system metastases was among exclusion criteria, and only four studies reported on its use treating patients with BM

(Carden 2008, Tamaskar 2006, Heymach 2006, Amaravadi 2007, Henderson 2007).

As a result, researchers of BM in NSCLC suggest caution when considering bevacizumab for patients with active BM (Cedrych 2016) until ongoing clinical trials of this subject yield more conclusive evidence (ClinicalTrials.gov 2019).

For research that relies on analyses of healthcare claims, the dearth of studies regarding bevacizumab for BM can be explained by the limited accuracy of secondary cancer diagnosis codes. In 2016, the Surveillance, Epidemiology, and

End-Results (SEER) program released its own data regarding diagnosis of BM during primary cancer staging workup, affording a new level of accuracy in population-level studies of synchronous BM (SBM). High-fidelity cancer registry data regarding BM may then be related to healthcare claims, further opening the door to large-scale analysis of BM treatment and outcomes.

This study identifies NSCLC patients with and without SBM treated with bevacizumab using Medicare claims data and evaluates the survival benefit of treatment with respect to primary cancer characteristics available from SEER, while further adjusting for treatment with several commonly-used chemotherapeutic agents. The resulting analysis offers insight into the treatment patterns and efficacy of bevacizumab among Medicare patients with NSCLC SBM.

3.3 Materials and Methods

3.3.1 Dataset

The SEER program of the National Cancer Institute collects cancer data

48 from 18 sites throughout the United States, representing about 27% of the population. SEER data may be linked to Medicare claims for further investigation, thus enabling us to identify the use of monoclonal antibodies in subjects aged 65 years or older. SEER data include data on BM diagnoses made at the same time as primary cancer diagnosis, and is directly abstracted by cancer registrars from medical records. Such a diagnosis of metastases during initial primary cancer staging is referred to as “synchronous” BM (SBM).

Five types of claim files offered as part of SEER-Medicare were used for this project: Part A inpatient claims (MEDPAR), carrier claims (NCH), outpatient

(OUTSAF), durable medical equipment (DME), and Part D drug prescription files.

Each record in these files contains a date of service, International Classification of

Diseases, Ninth revision, Clinical Modification (ICD-9-CM) diagnosis codes, Current

Procedural Terminology (CPT), and Healthcare Common Procedure Coding System

(HCPCS) procedure codes that were used to identify treatment and BM diagnoses.

Several variables from the joined dataset were considered. Age at diagnosis was reported as age groups 65-70, 71 to 75, 76 to 80, and over 80 years. Race was examined in terms of two categories: White Non-Hispanic and Non-White/Other.

Histology of lung cancer was categorized into non-adenocarcinoma and adenocarcinoma histologies. The derived American Joint Committee on Cancer staging data element was used to identify cancer diagnosed at stages I through IV, with A and B subcategories for stages I through III.

3.3.2 Data-derived definitions

Healthcare Common Procedural Coding System (HCPCS) codes indicating bevacizumab use in non-small-cell lung cancer patients were identified in Medicare

49 claims spanning 2007 through 2014 for patients whose primary cancer was diagnosed in the years 2010 through 2012, and included codes S0116, J9035,

C9257, C9214, and Q2024 (NCI Cancer Therapy Lookup Tables); this range of years of claims was selected to account for potential errors or delays in processing claims. Because bevacizumab has a half-life ranging on the order of weeks, only one infusion was required for a patient to have been considered as treated with bevacizumab.

To further evaluate the effect of bevacizumab, medications listed in the NCI’s chemotherapy lookup tables were also reported. The most frequently-used medications were selected for investigation, where potential medications must have been used in at least 11 patients diagnosed with brain metastases.

3.3.3 Population

Primary cancer cases were included only if it was the first cancer diagnosis for that patient, and lung cancers were selected on the basis of WHO site recoding

(code: 22030, NCI SEER Site Recode ICD-O-3). Because bevacizumab is used for non-small-cell lung cancers, patients with small-cell lung cancer according to ICD-O-

3 histology codes were excluded from this analysis (codes: 8002, 8041-8045) (NCI

SEER Training Modules, Lung Equivalent Terms and Definitions). Additionally, survival analysis was restricted to the population of patients diagnosed in 2010 through 2012 due to the unavailability of synchronous BM data prior to this timeframe.

Clinical and demographic characteristics are presented for the four categories of patients produced by a cross of bevacizumab treatment and SBM status. However, regression models utilized two populations of patients: one

50 population with brain metastases, and another representing the overall population of

NSCLC patients. Patients are therefore described as SBM+ or SBM-, indicating presence or absence of SBM, and BEV+ or BEV-, indicating presence of a record of bevacizumab treatment versus the absence thereof.

3.3.4 Treatment patterns

Differences in the use of bevacizumab were investigated across the NSCLC and the NSCLC with SBM populations using descriptive statistics and logistic regression. Descriptive statistics are presented for clinical and demographic characteristics across each of the four populations of patients produced by a cross of SBM and BEV statuses. Univariable logistic regression predicting the use of bevacizumab on the basis of clinical and demographic characteristics is presented to further describe treatment patterns.

Candidate predictors for a multivariable model of bevacizumab prescription included clinical and demographic characteristics, along with prescriptions of other medications that had been used in at least 11 SBM patients who were treated with bevacizumab. Predictors for a multivariable model were selected using the least angle shrinkage and selection operator (LASSO) as implemented by Friedman et al in their software package glmnet. The effect of regularization across 81 lambda values from ten thousand to one ten-thousandth were examined using 200-fold validation, and predictors minimizing mean square error across these parameters were selected for inclusion.

3.3.5 Treatment efficacy

Treatment efficacy in the population of patients diagnosed with brain metastases was evaluated using survival analysis predicting mortality with respect

51 to bevacizumab treatment and patient characteristics. Because several patient characteristics affect the decision to use bevacizumab, propensity score matching was used to account for potential confounding due to treatment decisions, where covariates were selected in the same manner as for multivariable models. Patients were matched on the basis of propensity to receive bevacizumab treatment using a caliper of 0.1 applied by a greedy nearest-neighbors matching algorithm and a 1:1 ratio of control to treated subjects without replacement. The success of propensity score matching was evaluated by fitting a multivariable logistic regression model to each of the matched and unmatched populations, then visually comparing the distributions of scores and standardized mean differences (Figures 3-4, 3-5, and 3-

6).

For both the propensity matched and unmatched samples, univariable Cox proportional hazards modeling was used to estimate the survival benefit of bevacizumab use with respect to candidate predictors. Weighted estimation of average hazard ratios was performed where non-proportional survival distributions were found across predictors using methods reported and implemented by Dunkler et al (2018).

A multivariable Cox proportional hazards model predicting survival was constructed using the same parameters and criteria as those of logistic regression, performed using the coxnet implementation provided by Simon et al (2011).

Univariable and adjusted Kaplan-Meier estimates of survival distributions were also included, where adjustment accounted for covariates presented as part of tables in results. The conditional method to balance a population across a set of covariates, described by Therneau et al (2015) and implemented by Kassambra and Kosinski

52 (2018), was used to visualize adjusted curves.

3.3.6 Rigor and reproducibility

All code written to perform analysis is available online as a series of scripts, along with a Makefile to make the entirety of analysis explicit and enable others to replicate this work, assuming proper data permissions are obtained by the user

(Ascha 2019).

3.4 Results

3.4.1 Clinical and demographic characteristics

Eighty-one SBM+ patients had a record of bevacizumab treatment, while

4,343 patients with SBM did not have such a record. Among those cases without

SBM, claims indicating bevacizumab administration were found for 666, and no such evidence was found for 35,197 patients (Table 3-1). The majority of cases were of adenocarcinoma histology among patients treated with bevacizumab, at

75.3% and 70.7% for SBM+/BEV+ and SBM-/BEV+ patients, respectively, but only

50.5% and 38.2% of SBM+/BEV- and SBM-/BEV- patients had adenocarcinoma histology. Except for the SBM+/BEV+ group, which was comprised of 43.2% males, each of the four populations had an approximately even distribution of males versus females. A greater proportion of SBM+/BEV- patients were of nonwhite race

(24.4%) than SBM+/BEV+ patients (18.5%), and the age of patients across all four categories was approximately 70 years.

3.4.2 Treatment Patterns

More than 90% of BEV+ cases were found to have some record of radiotherapy treatment, whereas the proportion of BEV- patients undergoing

53 radiotherapy was lower, at about 65%. Pemetrexed treatment was consistently associated with increased odds of bevacizumab treatment; for each of the SBM+ and overall NSCLC patient populations, patients receiving pemetrexed had 6.06

(95% CI: 3.45-10.85, p < 0.001) and 8.93 (95% CI: 7.3-11.0, p < 0.001) times the odds of bevacizumab prescription (Table 3-3, Table 3-4).

A great majority of BEV+ patients were treated with dexamethasone, ranging from 87.2% to 28.3% in the SBM-/BEV+ and SBM-/BEV- populations.

Dexamethasone treatment was associated with greater odds of BEV treatment in the SBM+ population (OR: 2.9, 95% CI: 1.4-6.4, p = 0.005) and overall patient population (OR: 2.0, 95% CI: 1.5-2.6, p < 0.001), after adjusting for clinical and demographic characteristics.

In both the SBM population and the overall population, treatment with paclitaxel was associated with bevacizumab prescription: adjusting for a variety of clinical and demographic characteristics, SBM+ patients treated with paclitaxel had

4.5 times the odds of bevacizumab treatment than their counterparts who were not

(95% CI: 2.7-7.6, p < 0.001), and the overall NSCLC population had a corresponding odds ratio of 2.5 (95% CI: 2.1-3.0, p < 0.001).

3.4.3 Survival analysis

Adjusting for clinical and demographic characteristics, bevacizumab treatment was associated with improved survival in the SBM+ (HR: 0.75, 95% CI:

0.59 – 0.96, p: 0.020) and overall population (HR: 0.88, 95% CI: 0.81 – 0.96, p =

0.003) (Figure 3-1, Figure 3-2); this effect persisted after propensity score matching in both the SBM+ (HR: 0.66, 95% CI: 0.47 – 0.91, p: 0.012) and overall population

(HR: 0.71, 95% CI: 0.63 – 0.80, p < 0.001) (Table 3-5, Table 3-6). Notably, the

54 distributions of survival times across treatment and control groups in the overall population was non-proportional (p < 0.001), and thus weighted estimation of an average hazard ratio associated with bevacizumab use (AHR: 0.73, 95% CI: 0.66-

0.80, p < 0.001) (Table 3-2) is reported. Among data elements selected for regression analysis, none had more than 5% of values missing (Table 3-7).

3.5 Discussion

Bevacizumab was first approved by the FDA for use in 2004, and its targeted mechanism of action was efficacious enough to merit development of biosimilars as soon as 2017 (Melosky 2018). Initially approved for metastatic colorectal cancer, non-small cell lung cancer and glioblastoma were added to its indications in 2006 and 2009, respectively (Kurzrock 2017, Diaz 2017). Intense research efforts have demonstrated that bevacizumab benefits patients affected by a variety of primary cancers, though its cost poses a barrier to many patients who might otherwise benefit from its use (Geynisman 2014). A precise understanding of its effects and those patients who stand to benefit the most, therefore, is distinctly important to informing treatment decisions.

Bevacizumab blocks the activity of VEGF-A upregulation in response to e.g. hypoxia-inducible factor 1, thereby directly counteracting the inappropriate angiogenesis that arises from oncometabolic stress (Ferrara 2004). This mechanism has led to its classification as a cytostatic agent rather than a cytotoxic agent; consequently, it has been suggested that survival endpoints address bevacizumab efficacy more appropriately than objective endpoints that are defined by measurements such as tumor size or spread (Ellis 2006). Furthermore, the extent of

55 tumor angiogenesis has strong prognostic implications in itself, so the potential benefit of bevacizumab may depend on the extent of disease with respect to angiogenic signalling (Hollingsworth 1995). Such a hypothesis would be supported by our observations of decreased hazard of mortality among SBM+ patients treated with bevacizumab compared to the overall population, and could meaningfully contribute to treatment decisions.

Studies to evaluate the efficacy of bevacizumab most often come in the form of randomized controlled trials and retrospective reviews, but the rise of real-world evidence has recently opened the door to such research using population-level data. Li et al (2019) report on a cohort of 22,258 patients with NSCLC from the US

Flatiron Electronic Health Record ddatabase, of whom 961 met inclusion criteria (Li

2019). Their examination of the overall population of NSCLC patients treated with bevacizumab revealed no significant differences in overall survival with respect to treatment type, similar to our finding of modest survival benefit resulting from the use of bevacizumab in the Medicare population of NSCLC patients. However, we found that after adjusting for clinical and demographic characteristics among

NSCLC patients with synchronous brain metastases, there is a significant survival benefit of bevacizumab such that SBM+/BEV+ patients had 0.75 times the hazard of mortality as compared to their SBM+/BEV- counterparts at any given time following primary cancer diagnosis (95% CI: 0.59 – 0.95, p = 0.020).

Concerns regarding a potential association with risk of intracranial hemorrhage have discouraged the use of bevacizumab (Caffo 2013). Yet, there have been studies of patients with primary brain tumors reporting clinical benefit and no significant increase in intracranial hemorrhage (Seoane 2014), and a meta-

56 analysis of eight studies covering 8,713 patients demonstrated no increase in the risk of intracranial hemorrhage in patients with brain metastases (Yang 2018).

Potential benefits to cranial pressure and hemorrhage, however, must be carefully weighed against the beneficial intracranial effects of anti-VEGF treatments.

VEGF-A is a ‘vascular permeability factor’, and bevacizumab reduces vasogenic edema by decreasing blood-brain barrier permeability. Similarly, dexamethasone is used to suppress vascular response to VEGF while also reducing VEGF production

(Heiss, 1996). In our population, dexamethasone was used at some point for a majority of SEER-Medicare patients diagnosed with synchronous BM, however, like in previous work, its use was not associated with differences in survival (Kural

2018). Consistent with an expected relationship between two drugs used for similar indications, the odds of bevacizumab prescription among patients who receive dexamethasone were greater for both the synchronous BM and overall populations, even after adjusting for clinical and demographic characteristics.

In addition to potentially minimizing cerebral edema, bevacizumab has significant systemic effects in patients with advanced stage NSCLC. Bernardo et al

2002 reported that patients whose brain metastases respond to this combination therapy also had response in extracerebral sites (Kim 2015, Matsuoka 2015,

Bernardo 2002). At the same time, large molecules such as bevacizumab may not cross the blood-brain barrier, so its delivery to brain parenchyma and chemotherapeutic effects on BM may not fully explain the increased survival benefit to patients with BM (Lampson 2011). The cost of bevacizumab should therefore be considered in light of both systemic disease management and potential alleviation of brain tumor sequelae.

57 While it reflects a great number of patients, the present study is limited with respect to population selection and detail available in these data. The SEER-

Medicare linkage allows researchers to examine high-fidelity cancer registry data as it relates to Medicare claims, however, this necessitates a restriction to the senior and elderly population. Though studies using SEER-Medicare are highly generalizable to an older population, conclusions do not reflect the US population at large. Furthermore, these data represent treatment patterns in the US, though there is significant variation in practice patterns across countries. In these claims data, the record of a test that may guide treatment decisions would be present, but the results of such a test would not be available; for example, in NSCLC, where EGFR tests are regularly performed in order to inform decisions to use EGFR-targeted drugs.

Despite these limitations, the availability of population-level cancer and mortality data from a well-established source is unique to the SEER-Medicare dataset, in contrast to other real-world evidence that relies solely on claims to identify death. The present work examines a high volume of data representing more than 40,000 NSCLC patients. This great number of cases was also not subject to usual constraints of high-dimensional data, due to its focus on one medication, with appropriate adjustment for potential confounders.

3.6 Conclusions

Bevacizumab is associated with modest survival benefit among the overall elderly population of NSCLC patients, and is associated with even greater such benefit when used for patients with synchronous BM. Understanding the outcomes and usage trends of bevacizumab in large datasets like SEER-Medicare both informs and reflects clinical practice, particularly with respect to the survival benefit

58 of bevacizumab for patients diagnosed with synchronous BM.

59 4. Disparities in stereotactic radiosurgery for the treatment of lung cancer brain metastases: A SEER-Medicare study

4.1 Abstract

Background: Stereotactic radiosurgery (SRS) is a costly procedure used to irradiate intracranial disease while sparing tissue that would be irradiated using whole-brain radiotherapy, ideally limiting the extent of cognitive impairment following treatment. Because brain metastases (BM) are the most frequent intracranial cancer and occur especially frequently among lung cancer patients, it is often used to treat

BM, though its cost may lead to potentially inequitable use across patient populations.

Materials and Methods: Surveillance, Epidemiology, and End-Results (SEER) cancer registry data for patients diagnosed between the years 2010 and 2012 were examined to identify lung cancer patients diagnosed with BM at the same time as their primary cancer (SBM). Medicare claims were used to identify patients treated with SRS; the odds of SRS therapy and hazards of mortality were examined with respect to various clinical and demographic characteristics were examined.

Results: Of 74,142 Medicare-enrolled patients diagnosed with lung cancer, 9,192 were diagnosed with SBM and 3,259 of those patients were treated with SRS.

Among patients diagnosed with SBM, after adjusting for clinical and demographic characteristics, males had 0.85 times the odds of SRS that females had. Despite having no significant difference in odds of SRS compared to White Non-Hispanic

(WNH) patients, patients of Asian/Pacific Islander background had improved survival. Black and patients of other race had significantly lower odds of SRS compared to WNH patients.

60 Conclusions: Differences in the use of SRS across patient populations may exist, but may not necessarily lead to differences in survival in the case of Asian/Pacific

Islander patients. Males and minority patients may have decreased odds of SRS and worse survival compared to WNH patients.

61 4.1 Introduction

Stereotactic radiosurgery (SRS) precisely irradiates tumors while sparing surrounding healthy tissue and has thus been increasingly adopted as an approach to treat brain metastases (BM). Because SRS treatment is recorded in claims data and its high cost encourages accurate billing codes, population-level analysis of claims data to identify SRS usage is both feasible and reliable. Combined with recently-released cancer registry data concerning the presence of brain metastases found during primary cancer staging workup (called “synchronous”, SBM), SRS treatment patterns can be examined both with respect to SBM and also SRS treatment for BM after primary cancer staging workup.

Though screening and early detection are ideal, it remains the case that up to 40% of NSCLC patients are diagnosed with brain metastases over their lifetime, with potentially greater incidence among patients with ALK-rearrangement or EGFR mutation. After a diagnosis of brain metastases, patients are presented with an increasing array of options for therapy, though decisions regarding such therapy remain largely in the hands of patient and provider due to a lack of strong evidence in favor of any given treatment regimen for each patient.

SRS treatment for BM is, however, a costly procedure whose equitable delivery has become a significant concern after several recent studies of the

National Cancer Database (NCDB) revealed potential disparities across several strata including distance to treatment center, insurance provider, and race (Haque

2018, Kann 2018, Trifletti 2018). Though these studies have confirmed disparities within the NCDB dataset, there is less work examining SRS treatment disparities as found in other data sources, such as Medicare claims, or primary cancer registry

62 data specifically made to be geographically representative, such as that offered by the Surveillance, Epidemiology, and End-Results (SEER) program.

The present study examines Medicare-reported SRS treatment among elderly US patients diagnosed with lung cancer and SBM, in particular whether there exist differences in survival or times to SRS treatment across strata previously noted to reflect disparate treatments or outcomes in studies of the NCDB.

4.2 Methods

4.2.1 Dataset

The SEER program reports cancer registry data on about 29% of the population in the United States. Cancer registry reporting sites and data elements are carefully selected to ensure a geographically representative sample of cancer diagnoses, therefore studies using SEER data are highly generalizable to the population at large. To maintain its high fidelity, cancer registry data is often reviewed several times prior to reporting. As a result, data elements in cancer registries are strictly defined according to standards, for example the Collaborative

Stage Data Collection System provides a single uniform set of codes and rules to define extent of disease.

These well-defined data may be supplemented by data from other sources, the most notable of which being Medicare claims data. SEER makes available a set of Medicare claims files with which investigators can examine a rich variety of patient characteristics and treatments by selecting relevant diagnosis or procedure billing codes. In this study, claims reflecting Part A (Inpatient), Part B (Outpatient), and carrier claims files from the years 2007 through 2014 were examined for presence of SRS treatment codes.

63 4.2.2 Population

Data from patients with primary lung cancer were examined for evidence of

SRS treatment due to the greater rate at which these primary cancers metastasize.

Several exclusion criteria were required to ensure appropriate comparisons, these can be broadly considered as general criteria of cancer registry studies and specific criteria required for the use of Medicare data. General criteria included a restriction to include only the population of patients with lung cancer as their first cancer diagnosis, only patients with World Health Organization codes for lung cancer

(22030), and patients diagnosed with cancer during their lifetime rather than on autopsy. Because Medicare claims are used to identify SRS treatment, this population was further reduced by excluding patients under 65 years of age, those with unknown insurance status or with a record of HMO use, and patients who qualified for Medicare due to end-stage renal disease or disability.

In 2016, SEER released data regarding brain, lung, liver, and bone metastases diagnosed during staging workup for primary cancers. Though they all contribute to the definition of stage of disease, only metastases to the brain were examined because only cranial SRS codes were used to identify the treated population. These cases are referred to as “synchronous” BM. Stage itself was classified according to the derived American Joint Committee on Cancer (AJCC),

Sixth edition, and SBM according to the Collaborative Staging data element on which the derived AJCC stage is defined.

4.2.3 Outcomes

Diagnosis codes for stereotactic radiosurgery were identified using previous

64 studies and by selection from the 2014 American Medical Association CPT codebook; these included the codes G0173, G0242, G0243, G0251, G0338, G0339,

G0340, 61793, 61796, 61797, 61798, 61799, 61800, 77371, 77372, and 77432.

Patients were considered as having had SRS if at least two records of SRS were found for that patient, and the first record of SRS was taken to be the treatment date for analysis of time to SRS treatment.

4.2.4 Definitions and covariates

Race was defined by two SEER data elements, race and the NHIA definition of Hispanic ethnicity. The race recode identifies patients as either White, Black,

American Indian (AI), Asian or Pacific Islander (API) or other, the NHIA classification was used to further categorize patients into White Non-Hispanic (WNH) and White

Hispanic (WH).

Rural-urban classification is also reported according to 2013 rural-urban continuum codes released by the USDA Economic Research Service, in which

1,167 counties are considered metropolitan and 1,976 non-metropolitan. The five most frequent histology codes were used to define histology, and others considered

‘other’ histology. For regression modeling, histology categories were further reduced into non small-cell lung cancer (NSCLC), small-cell lung cancer (SCLC), and other.

4.2.5 Statistical Analysis

Descriptive statistics of clinical and demographic characteristics were calculated across SRS use for each group of patients diagnosed with or without

SBM, and tests of difference (t-tests, Wilcoxon rank-sum tests, and chi-square tests) between characteristics of patients who did and did not undergo SRS treatment are presented for the overall population. Because race was the primary predictor of

65 interest, it was retained wherever predictors for multivariable models were selected using statistical methods. For each analysis, two cohorts of patients were considered: the overall population of elderly lung cancer patients, and the population of patients diagnosed with SBM.

Univariable and multivariable logistic regression predicting the use of SRS were performed to identify the odds of SRS associated with race. Predictors for the multivariable logistic regression model were selected using the least angle shrinkage and selection operator (LASSO) as implemented by Friedman et al in their R package glmnet. Regularization parameters were calculated for 81 lambda values ranging from ten thousand to one ten thousandth, and predictors minimizing root mean square error on 200-fold cross-validation were selected for inclusion.

SRS was also examined using Cox proportional hazards regression models of time to the first Medicare claim for SRS as an outcome. By examining differences in SRS use across time in addition to SRS as a binary measure, the instantaneous hazard of SRS is estimated with respect to time, supplementing analysis of odds over all available follow-up time.

Mortality following primary cancer diagnosis was also examined using Cox proportional hazards regression, where the primary predictors of interest were the interaction between SRS use and race. To minimize selection bias associated with the decision to use SRS, propensity score matching on the basis of propensity to undergo SRS was performed using a caliper of 0.1, greedy nearest-neighbors matching algorithm, 1:1 ratio of treated to control subjects, without replacement. The success of matching was evaluated using multivariable logistic regression predicting a propensity of SRS among each of the treated and control groups, and comparing

66 the distributions of these propensity scores across populations.

Both univariable and multivariable Cox regression models were fit, where multivariable covariates were selected using the LASSO implemented by Simon et al (2011) using the same parameters as those described for logistic regression variable selection.

Multivariable analysis candidate predictors included age at diagnosis, derived AJCC stage, rural-urban classification, race, histology, sex, and record of any surgical treatment.

4.2.6 Rigor and reproducibility

All code written to perform analysis contained herein is available as a series of scripts and a Makefile to ensure the entirety of methods are explicit and can be re-used by others to replicate this work, assuming appropriate data access is obtained (Ascha 2019).

4.3 Results

4.3.1 Treatment patterns

Of 97,175 patients diagnosed with lung cancer between 2010 and 2012,

74,142 met inclusion criteria. Among those meeting inclusion criteria, the 19%

(14,317) of patients who received SRS were somewhat younger than untreated patients (74.6 versus 75.1 years) (Table 4-1). Differences in proportions across SRS treatment were observed in both race and sex, 82.0% of SRS+ patients were WNH compared to 78.8% among SRS- patients, and 54.6% of the SRS+ group were women compared to 48.0% of SRS- patients.

In the SBM population, patients over 75 years of age had 0.63 times the odds of SRS treatment that those between 65 - 70 years of age had (95% CI: 0.57-

67 0.70, p < 0.001), Black patients had 0.69 times the odds of SRS compared to WNH patients (95% CI: 0.59-0.80, p < 0.001), race, and sex all remained statistically significant after covariate adjustment for clinical and demographic characteristics in multivariable logistic regression models, except in the case of the odds ratio of A/PI race compared to WNH (p = 0.859, Table 4-2).

9,192 patients were found to have SBM, amounting to 12.4% of the overall population. 35% of patients diagnosed with SBM (n = 3,259) were treated with SRS, compared to about half that proportion (17%, n = 11,058) among patients not diagnosed with SBM. Patients with SBM were a mean 72.9 years old, compared to

75.2 years among those without SBM (Table 4-1).

Black patients and those of other (American Indian or White Hispanic) race had 0.70 (95% CI: 0.65, 0.75) and 0.86 (95% CI: 0.79, 0.93) times the odds of SRS treatment that the WNH population had, respectively. The odds of SRS among A/PI patients, however, were not significantly different from those of WNH patients (OR:

0.92, 95% CI: 0.77-1.11, p: 0.392) (Table 4-3).

4.3.2 Survival

In the overall population, adjusting for clinical and demographic characteristics, males were found to have 1.21 times the hazard of mortality that females did at any given time (Table 4-5). Adjusted analysis also revealed that A/PI patients had 0.76 times the hazard of mortality than WNH patients at any given time

(95% CI: 0.73-0.78, p < 0.001) and Black patients had 1.04 times the hazard of mortality compared to WNH patients (95% CI: 1.01, 1.07, p: 0.010). Male patients who received SRS had greater hazard of mortality than WNH patients who received

SRS (HR: 1.09, 95% CI: 1.00, 1.18, p: 0.044), but A/PI patients who received SRS

68 had 0.89 times the hazard of mortality that WNH patients receiving SRS (95% CI:

0.81, 0.98, p: 0.019) (Table 4-4).

In the population of patients with SBM, A/PI had significantly decreased hazard of mortality compared to WNH patients on both crude (HR: 0.68, 95% CI:

0.62-0.74, p < 0.001) and adjusted (HR: 0.60, 95% CI: 0.55-0.66, p < 0.001) Cox proportional hazards modeling (Table 4-4). In crude analysis, Black patients with

SBM had 1.18 times the hazard of mortality that WNH had (95% CI: 1.10, 1.27, p <

0.001), but this relationship did not hold on multivariable analysis (HR: 1.04, 95% CI:

0.97, 1.12, p: 0.2338). Male patients and those living in non-Metro areas had greater hazard of mortality even after adjustment, where males had 1.13 times the hazard compared to females (95% CI: 1.09, 1.18, p < 0.001) and non-Metro patients had

1.10 times the hazard of mortality that Metro-area patients had (95% CI: 1.03, 1.18, p = 0.007).

4.4 Discussion

The use of SRS to treat SBM in elderly patients is often preferred because it is non-invasive and delivers radiation to a tumor while sparing surrounding healthy tissue. Elderly patients are frequently less able to tolerate neurosurgery, and whole- brain radiation therapy (WBRT) irradiates non-cancerous tissue, so the use of SRS avoids several of the pitfalls associated with alternative BM treatments that may be contraindicated or have greater side effects in the elderly population. Because several studies of the NCDB have reported disparities across different patient characteristics, validation of these findings using SEER-Medicare data may confirm meaningful insight into which populations of patients are in the greatest need of access to care.

69 In their study of NCDB data from the years 2004 to 2013, Trifletti et al report that female patients had 1.11 times the odds of radiosurgery compared to male patients (95% CI: 1.04, 1.17, p: 0.001), even after adjusting for a variety of clinical and demographic characteristics. Similarly, the odds of SRS treatment were greater for females both in the general population and among patients with SBM in SEER-

Medicare data, with a greater effect observed in the SBM population (Tables 4-2, 4-

3).

Male lung cancer patients may have poorer prognosis and more aggressive disease than those of female patients. This was the case in the Medicare population presented here, among whom female patients in this population had decreased hazard of mortality (Tables 4-4, 4-5), similar to findings reported by Robin et al

(2018), in which female patients had 0.88 times the hazard of mortality at any given time (95% CI: 0.84, 0.93, p < 0.001). Indeed, previous work describes a statistically significant association between sex and survival even with small sample sizes, pointing to differences in healthcare outcomes that could meaningfully affect treatment decisions for males considering SRS. Given evidence that men may also experience greater pain and discomfort following SRS treatment for SBM, sex should serve some role in discussions about SRS and SBM treatment decisions.

In contrast to the lack of a statistically significant effect associated with Black race in the 42,148 patients studied by Trifletti et al, Kann et al (2017) report that non-white patients in the NCDB have decreased odds of SRS. The effect of black race in SEER-Medicare, however, was clear: on multivariable analysis, the odds of

SRS for black patients was 0.66 times that of WNH patients (95% CI: 0.57-0.78, p <

0.001). By confirming NCDB conclusions using a different data source, SEER-

70 Medicare, the present work suggests that the black population seems medically under-served across data sources and even after adjustment for clinical and demographic characteristics. Moreover, adjusted analysis showed the hazard of mortality among black patients was greater than that of WNH patients (1.10, 95%

CI: 1.07, 1.13, p < 0.001), and analysis of the interaction between race and SRS use revealed that black patients who received SRS also had greater hazard of mortality than did WNH patients who received SRS (1.09, 95% CI: 1.00, 1.18, p:

0.0437).

Torre et al (2016) report that Asian Americans, Native Hawaiians, and

Pacific Islanders have 30% to 40% lower incidence and mortality rates than WNH patients, the present study supports their findings: for those with BM, Asian patients had 0.61 (95% CI: 0.55-0.66) times the hazard of mortality that WNH patients had.

Further, in the overall population, Asian patients who received SRS had 0.89 times the hazard of mortality compared to WNH patients who received SRS (95% CI:

0.81, 0.98, p = 0.0185, Table N). Improved survival among A/PI compared to WNH is also observed among breast cancer patients, suggesting A/PI populations may have better survival across primary cancers, as well (Thompson, et al. 2016).

Though we found no statistically significant relationship between Asian race and odds of SRS compared to WNH (Figure 4-1), there did appear to be differences in time to SRS (Figure 4-2). Trifiletti, et al. (2017) confirm this finding with NCDB data, further reporting that Asian Americans received SRS as the sole course of treatment less frequently than WNH and Black patients.

The present study is limited in several respects, largely as a result of the granularity available in real-world evidence such as Medicare claims. Since there

71 were no cranium-specific radiotherapy codes available during the study period, this work cannot comment on the efficacy of SRS with or without WBRT. Previous work shows that, among patients whose cranial disease burden is limited to four or fewer

BM of less than 4cm in size, SRS as a standard of care is associated with improved survival when compared to more generalized therapy such as WBRT. Other work suggests that SRS plus WBRT leads to better patient outcomes than WBRT alone.

This limitation of claims coding, therefore, prevents the use of real-world evidence to investigate such an important subject as which therapeutic regimens offer the greatest benefit.

72 Chapter 5 - Conclusions and future directions

5.1 Summary and translational impact

The lifetime incidence of BM appears closely related to that observed during the synchronous period surrounding primary cancer diagnosis, where cases that are typically more aggressive (Triple Negative breast cancer or SCLC) demonstrated greater incidences of LBM. In the case of lung cancers, our finding that there is increased clinical recognition of brain metastases in SCLC patients following primary cancer diagnosis suggests support for the use of prophylactic cranial irradiation (PCI), particularly if quesstions regarding the survival benefit of PCI versus close monitoring and response are resolved (Hochstenbag et al. 2000,

Gregor et al. (1997), Takahashi et al. (2017)). In addition to evidence to support PCI in some cases, the decision to screen for BM may draw from estimates of SBM versus LBM incidence.

Beyond the pathology imposed by cancer generally, BM sequelae may include increased BBB permeability. By reducing BBB permeability and thus alleviating intracranial hypertension, bevacizumab may provide extra benefit to patients with BM. Greater survival benefit was observed with the use of bevacizumab in patients with BM, however, despite efforts to control for confounding, a number of other mechanisms may account for this observation.

Last, our investigation of SRS treatment and its equitable use revealed that, despite decreased use among patients of Asian/Pacific Islander background, this population seems to have improved survival compared to other races. Such observations are consistent with the literature, though the precise mechanisms by which this occurs are still unclear (Thompson et al 2016, Trifiletti, et al. 2017).

73 5.2 Claims data limitations and future directions

The results of each investigation presented herein are limited by the use of

Medicare claims as a result of the assumption that each code meaningfully reflects the occurrence of what it is intended to signify (Sarrazin 2012). The first manuscript in this series is limited by its reliance on diagnostic and diagnostic procedure billing codes, the documentation of which is often systematically underreported due to a lack of financial incentive (Haut 2012). Using inpatient data from 1997, Dismuke

(2005) reported that 33% and 43% of claims containing CT and MRI revenue codes, respectively, were associated with a corresponding ICD-9-CM code indicating CT or

MRI. More recently, inpatient data from a consortium of 8 hospitals was compared to clinician-verified surgical data from the National Surgical Quality Improvement

Program (NSQIP), and found concordance ranging from 0.79 for diabetes to 0.02 for dyspnea identification (Etzioni 2019). A claims-NSQIP concordance demonstrated significant variation across hospitals (p < 0.001), indicating that the predictive value of every data source must be considered in itself for study design and analysis purposes.

Diagnosis codes for acute conditions are particularly unreliable; Grams

(2014) report dismal billing code sensitivity (17.2%, 95% CI: 13.2, 21.2) and excellent (> 98%) specificity to detecting acute kidney injury (AKI) in their 1996-2008 claims data. AKI is not likely to be as frequently billed (both within each patient and across patients) compared to chronic conditions such as BM, which helps explain poor performance as a predictor of the condition it is meant to signify. Even in cases where code reliability is likely greater, as in procedures, the purpose or motivation of a procedure may remain unclear. In the case of endoscopy, greater than 90%

74 agreement between claims and medical records has been observed, but this reliability did not extend to whether an endoscopic procedure was meant for screening or diagnostic purposes (Schenck 2007).

Further, the study of claims diagnosis codes is often greatly limited by the common practice of relying on sensitivity and specificity rather than measures unaffected by class imbalance. In the example of AKI prediction presented above, a negative prediction for every case would yield > 98% sensitivity because so few patients had a diagnosis of AKI. One possible solution is resampling to artificially produce a balanced dataset, but this approach would simultaneously limit the generalizability of those findings to other populations that have been resampled. Our work used kappa concordance to overcome statistical issues associated with such severe class imbalance, as this measure of concordance accounts for the probability of random agreement given the proportions of an outcome ascertained using a gold standard.

On the other hand, both general advances and characteristics specific to BM suggest that BM may be studied using claims. First, a decade’s worth of progress in health information technology, such as more precise claims codes and federal mandates to use electronic medical records, would ideally narrow such wide discrepancies. Perhaps more importantly, the significant cost associated with cranial procedures such as neurosurgery or SRS suggests that their coding, at least in the context of primary cancer outside of the head and neck, would be associated with a diagnosis code that is local to the head. Also, BM are a chronic condition that may require many visits, and therefore bills, with which its identification in claims data may be performed. Future research on claims data phenotyping, which is the

75 identification of a patient condition from their claims data, could focus on these aspects of disease and treatment to improve phenotype accuracy and enable broader studies of more conditions than currently possible.

Beyond the study of a single code as it predicts what it nominally signifies, significant advances in predictive modeling may open the door to identifying patient characteristics using the whole of a patients’ electronic data (Banda 2017).

Treatment characteristics such as a patient’s primary care physician can be predicted with up to 85% concordance solely using Medicare data, so the highly accurate identification of disease or treatments from administrative data is not an idea of the future, despite warnings issued by data holders such as SEER-Medicare

(DuGoff 2018, SEER 2019). Claims data have been used to identify Parkinson’s disease using regularized regression, which has gained significant traction as a computationally inexpensive algorithm, and the evaluation of a wide array of algorithms has led to sensitivity, specificity, negative predictive value, and positive predictive value to detect multiple sclerosis all exceeding 95% (Nielson 2017,

Culpepper 2019). In addition, unstructured data may be leveraged to perform ‘High-

Throughput Phenotyping’, and are sufficiently accurate to identify differences in outcomes such as readmission or length of stay (McCoy 2018). Such sources are usually in the form of notes on a patient, and so classification accuracy becomes a non-issue.

5.3 Brain metastases incidence and treatment future directions

Though the frequency of late-stage lung cancer diagnosis is partially a result of the lack of innervation to signal pathological pain from within the lungs, it may be possible to better identify the subset of patients who are at risk of lung cancer and

76 thus find more diagnoses. Specifically, with greater information available in each patient’s electronic medical records, advances in predictive modeling could better identify at-risk patients when compared to e.g. the straightforward age and pack- year criteria offered by the American Cancer Society (ACS) (Hinton 2015). This line of research could be integrated into existing information technology systems and deployed at a massive scale, with practice in the distant future potentially informed by novel research in real-time rather than relying on dissemination via publication and the ideal that every practitioner is wholly up-to-date.

Identifying at-risk patients using such predictive models could, therefore, accurately inform screening decisions. For example, lung cancer patients are frequently diagnosed at a later stage of disease than patients with other cancers, yet the quality of evidence in summary recommendations on screening practice has been deemed very low. In the US, the American Cancer Society recommends lung cancer screening for moderately healthy patients between the ages of 55 and 74 who have a smoking history of at least 30 pack-years (Wender 2013, Smith 2017,

Smith 2018); Japuntich et al report that among 134 such at-risk patients, only 21% had undergone lung cancer screening, and that this proportion decreased when considering only minority patients (Japuntich 2018). Because up to forty percent of patients present with intrathoracic spread and ten percent present with clinically symptomatic BM, earlier detection of lung cancer would affect a significant proportion of patients and perhaps improve survival following SRS at the population- level (Collins 2007). For undiagnosed lung cancer patients whose hospital EMR integrates alerts, screening recommendation model deployment could simultaneously reduce the public burden of paying for screening based on

77 generalized recommendations while also diagnosing disease before the development of distant metastases.

Other next steps for the use of observational data to study BM could also include studies of HER2+ metastatic breast cancer. As mentioned in the introduction, increased rates of BM have been observed among HER2+ breast cancer patients treated with trastuzumab; the “lapatinib plus capecitabine in patients with previously untreated brain metastases from HER2-positive metastatic breast cancer” (LANDSCAPE) trial enrolled 45 patients, 90% of whom had been exposed to trastuzumab (Bachelot 2013). In that trial examining the HER1 and HER2 tryrosine kinase inhibitor lapatinib combined with the capecitibine, both of which are thought to cross the BBB (though there is evidence lapatinib may not reaceh meaningful concentrations), 66% of patients had partial response over a median 21 months of follow-up (Petrelli 2017, Zhang 2015). Because lapatinib for patients already using capecitabine recevied FDA approval in 2007, recent data sources could feasibly investigate the benefit of this combination of medications in a great number of patients (Higa 2007, Womens Health 2010).

With more (in any of volume, velocity, and variety) data generated for the purposes of patient care or healthcare administration, that data will increasingly be used for secondary purposes, primarily in the form of medical research. For any data source, the determinants of accuracy for new purposes must be comprehensively evaluated and dealt with so as to maximize usefulness for research while minimizing the possibility of errant conclusions. One approach to achieve this is by comparison with a gold standard, like cancer registries or NSQIP whose data are collected specifically for research, or by the study of procedures that

78 we can reasonably assume will be well-documented in that data, such as costly drugs or services in billing claims. Though systematic error is minimized by manual chart review, the scale of research now demands greater attention to generalized biases in the data generation process, in addition to bias specifically pertaining to an effect of interest. “Big data” may be amorphous in its definition, but the need for replicable research whose statistical methods go to greater lengths to ensure accuracy is clear, and the present work demonstrates requisite considerations to ensure accuracy in recently-emerged clinical research methods.

79 Table 1-1 - Frequency and median survival for patients with synchronous brain metastases

Frequency and median survival for patients with synchronous brain metastases,

SEER 18, 2010 through 2013. A star (*) indicates data from (Cagney 2017), otherwise data is from (Kromer 2017). “NR” indicates that a quantity that was not reported.

Site and Total Brain Brain Brain Median

Histology Cases metastases not metastases metastases survival*

present present data missing months (IQR)

Lung/Bronchus 203,915 164,479 22,032 17,404 NR

(80.7%) (10.8%) (8.5%)

Small-cell lung 23,280 18,471 (79.3%) 3,518 (15.1%) 1,291 (5.6%) 6.0 (2.0 – 11.0)

Non-small cell 153,650 128,930 16,483 8,237 (5.4%) 4.0 (2.0 – 9.0) lung (83.9%) (10.7%)

Melanoma 79,785 75,714 (94.9%) 980 (1.2%) 3,091 (3.9%) 6.0 (2.0 – 13.0)

Breast 246,763 239,256 1,001 (0.4%) 6,506 (2.6%) 10.0 (3.0 –

(97.0%) 30.0)

Esophagus 15,739 14,142 (89.9%) 238 (1.5%) 1,359 (8.6^) 4.0 (2.0 – 9.0)

Colon/Rectum 147,592 139,469 373 (0.3%) 7,750 (5.3%) 6.0 (2.0 – 15.0)

(94.5%)

Kidney/Renal 57,372 54,429 (94.9%) 814 (1.4%) 2,129 (3.7%) 5.0 (2.0 – 12.0)

Pelvis

80 Table 1-2 Diagnosis and procedure codes used to identify brain metastases and its treatment

Treatment Type Code Code

Classification

Biopsy CPT 61140, 61750, 61751

Diagnostic Imaging CPT 70450 – 70470, 70551 –

70553, 78607 – 78608

Radiation Therapy ICD-9-CM 92.23–92.24, 92.30–92.33, and

92.39.

CPT 77261–77263, 77280, 77285,

77290, 77295, 77299–77301,

77305, 77310, 77315, 77321,

77332- 77334, 77336–77337,

77370–77372, 77399, 77402–

77414, 77416, 77418–77420,

77425, 77427, 77430, 77432,

and 0073T.

HCPCS G0173-G0174, G0242-G0243,

G0251, and G0338-G0340.

Stereotactic radiosurgery ICD-9-CM 92.30–92.33 and 92.39.

CPT 61793, 61796–61800, 77371–

77372, and 77432.

81 HCPCS G0173, G0242, G0243,

G0251, G0338, G0339, and

G0340.

Neurosurgical resection ICD-9-CM 01.21–01.25, 01.31, 01.51, and

01.59.

CPT 61304–61305, 61312–61315,

61320–61321, 61330, 61332–

61334, 61340, 61343, 61345,

61440, 61450, 61458, 61460,

61470, 61500–61501, 61510,

61512, 61514, 61516, 61518,

61519- 61522, 61524, 61526,

61530–61531, 61533–61536,

61538- 61539, 61541–61546,

61550, 61552, 61556–61559,

61563- 61564, 61570–61571,

61575–61576, 61580–61586,

61590–61592, 61596–61598,

61600–61601, 61605–61613,

and 61615–61616.

Chemotherapy CPT (Intravenous) 90782 – 90788, 96408 – administration services 96419, 96520, 96530, 96545,

96549

82 CPT (Not intravenous) 61517, 96400 – 96406, 96420-

96425, 96440-96446, 96450,

96520-96522, 96542, 99601-

99602

HCPCS Level II Q0083-Q0085, J9000-J9999

ICD-9 CM (Diagnostic) V58.1, V58.11, V58.12

Table 1-2: Claims codes used to identify patient characteristics and treatment. With the exception of radiotherapy, codes refer to cranial procedures. See “Cancer

Therapy Look-up Tables” from the National Cancer Institute for chemotherapy drug

NDC codes.

83 Table 2-1 - Medicare claims brain metastasis classification accuracy

For Medicare claims data algorithms identifying synchronous and lifetime brain metastases based on the presence of a diagnosis code for secondary cancer of the

CNS and a procedure code indicating intracranial imaging within sixty days of the other, this table shows classification accuracy compared to a SEER SBM gold standard. The rows under “Algorithm: Synchronous” represent such an algorithm with the additional criteria that codes must have occurred within 60 days of primary cancer diagnosis, whereas “Algorithm: Lifetime” indicates that codes may have occurred at anytime in Medicare claims. “Sens” refers to sensitivity; “spec”, specificity; PPV, positive predictive value. An asterisk (“*”) represents values that are suppressed to avoid reporting data from 1 through 11 cases.

84 Primary Predicted True Sens PPV Spec Kappa

Algorithm:

Synchronous

Lung Overall 5100 6789 0.58 0.77 0.98 0.63 (0.62 - 0.64)

Adenocarcinoma 2103 2806 0.59 0.79 0.98 0.63 (0.62 - 0.65)

Carcinoma 403 571 0.57 0.8 0.98 0.62 (0.59 - 0.66)

NSCLC 506 745 0.56 0.82 0.98 0.62 (0.58 - 0.65)

Other 572 784 0.58 0.8 0.99 0.65 (0.62 - 0.68)

SCLC 872 1144 0.59 0.78 0.97 0.62 (0.60 - 0.65)

Squamous 644 739 0.57 0.66 0.99 0.60 (0.56 - 0.63)

Breast Overall 196 203 0.46 0.47 1 0.46 (0.40 - 0.53)

Her2-/Hr+ 103 86 0.41 0.34 1 0.37 (0.28 - 0.46)

Her2+/ Hr(+/-) 27 31 0.48 0.56 1 0.51 (0.36 - 0.67)

Other 38 50 0.44 0.58 1 0.50 (0.37 - 0.63)

Triple (-) 28 36 0.58 0.75 1 0.65 (0.52 - 0.79)

Skin Overall 152 230 0.55 0.83 1 0.66 (0.60 - 0.71)

MMJN * * 0.43 0.6 1 0.50 (0.15 - 0.85)

85 Nevi & 134 206 0.54 0.83 1 0.65 (0.59 - 0.71)

Melanomas

Other 13 * 0.71 0.92 1 0.79 (0.64 - 0.95)

Algorithm:

Lifetime

Lung Overall 9138 6789 0.65 0.48 0.92 0.49 (0.48 - 0.50)

Adenocarcinoma 3630 2806 0.65 0.5 0.91 0.50 (0.48 - 0.51)

Carcinoma 575 571 0.63 0.62 0.95 0.57 (0.53 - 0.61)

NSCLC 804 745 0.61 0.57 0.92 0.52 (0.49 - 0.55)

Other 1028 784 0.65 0.5 0.95 0.52 (0.49 - 0.55)

SCLC 1863 1144 0.67 0.41 0.84 0.40 (0.38 - 0.43)

Squamous 1238 739 0.64 0.38 0.95 0.45 (0.42 - 0.48)

Breast Overall 967 203 0.57 0.12 0.99 0.19 (0.16 - 0.22)

Her2-/Hr+ 430 86 0.53 0.11 0.99 0.18 (0.13 - 0.22)

Her2+/ Hr(+/-) 181 31 0.68 0.12 0.97 0.19 (0.12 - 0.26)

Other 144 50 0.54 0.19 0.99 0.27 (0.19 - 0.36)

Triple (-) 212 36 0.58 0.1 0.96 0.16 (0.10 - 0.22)

86 Skin Overall 658 230 0.66 0.23 0.98 0.33 (0.29 - 0.37)

MMJN 114 * 0.71 0.04 0.99 0.08 (0.01 - 0.15)

Nevi & 498 206 0.65 0.27 0.97 0.36 (0.32 - 0.41)

Melanomas

Other 46 * 0.82 0.3 0.95 0.42 (0.27 - 0.57)

87 Table 2-2 - Diagnosis code-only algorithms performance

For Medicare claims data algorithms identifying synchronous and lifetime brain metastases based only on the presence of a diagnosis code for secondary cancer of the CNS, this table shows classification accuracy compared to a SEER SBM gold standard. The rows under “Algorithm: Synchronous” represent such an algorithm with the additional criteria that codes must have occurred within 60 days of primary cancer diagnosis, whereas “Algorithm: Lifetime” indicates that codes may have occurred at any time in Medicare claims. “Sens” refers to sensitivity; “spec”, specificity; PPV, positive predictive value. An asterisk (“*”) represents values that are suppressed to avoid reporting data from 1 through 11 cases. NSCLC stands for non-small-cell lung cancer; SCLC, small-cell lung cancer; MMJN, malignant melanoma in junctional nevi.

88 Primary Predicted True Sens PPV Spec Kappa

Algorithm:

Synchronous

Lung Overall 6445 6789 0.74 0.78 0.98 0.73 (0.72 - 0.74)

Adenocarcinoma 2643 2806 0.75 0.79 0.97 0.74 (0.72 - 0.75)

Carcinoma 539 571 0.75 0.79 0.97 0.74 (0.70 - 0.77)

NSCLC 664 745 0.73 0.82 0.97 0.74 (0.71 - 0.77)

Other 723 784 0.74 0.8 0.99 0.75 (0.73 - 0.78)

SCLC 1079 1144 0.74 0.78 0.97 0.72 (0.69 - 0.74)

Squamous 797 739 0.72 0.67 0.98 0.68 (0.65 - 0.71)

Breast Overall 294 203 0.6 0.41 1 0.49 (0.43 - 0.54)

Her2-/Hr+ 167 86 0.57 0.29 1 0.39 (0.31 - 0.46)

Her2+/ Hr(+/-) 34 31 0.55 0.5 1 0.52 (0.37 - 0.67)

Other 55 50 0.6 0.55 1 0.57 (0.46 - 0.68)

Triple (-) 38 36 0.69 0.66 1 0.67 (0.55 - 0.80)

Skin Overall 197 230 0.72 0.84 1 0.78 (0.73 - 0.82)

MMJN * * 0.43 0.6 1 0.50 (0.15 - 0.85)

89 Nevi & 177 206 0.72 0.84 1 0.77 (0.73 - 0.82)

Melanomas

Other 15 17 0.82 0.93 1 0.87 (0.75 - 1.00)

Algorithm:

Lifetime

Lung Overall 12126 6789 0.85 0.48 0.89 0.55 (0.54 - 0.56)

Adenocarcinoma 4807 2806 0.86 0.5 0.88 0.56 (0.55 - 0.58)

Carcinoma 780 571 0.83 0.61 0.92 0.65 (0.62 - 0.68)

NSCLC 1102 745 0.84 0.57 0.89 0.61 (0.58 - 0.64)

Other 1359 784 0.84 0.49 0.93 0.58 (0.55 - 0.60)

SCLC 2411 1144 0.86 0.41 0.79 0.45 (0.42 - 0.47)

Squamous 1667 739 0.85 0.38 0.93 0.49 (0.46 - 0.51)

Breast Overall 1467 203 0.76 0.11 0.98 0.18 (0.16 - 0.21)

Her2-/Hr+ 701 86 0.74 0.09 0.98 0.16 (0.12 - 0.19)

Her2+/ Hr(+/-) 255 31 0.81 0.1 0.96 0.17 (0.11 - 0.22)

Other 217 50 0.72 0.17 0.99 0.27 (0.20 - 0.34)

Triple (-) 294 36 0.83 0.1 0.95 0.17 (0.12 - 0.23)

Skin Overall 892 230 0.87 0.22 0.97 0.35 (0.31 - 0.38)

90 MMJN 150 * 0.86 0.04 0.98 0.07 (0.02 - 0.13)

Nevi & 689 206 0.87 0.26 0.96 0.39 (0.35 - 0.43)

Melanomas

Other 53 * 0.88 0.28 0.93 0.40 (0.26 - 0.55)

91 Table 2-3 - Incidence of primary cancer associated with brain metastases

The first set of AAIR, total at-risk, and present/absent/missing values in this table reflects Surveillance, Epidemiology, and End-Results program synchronous brain metastases data, while the second set reflects Medicare lifetime brain metastases data. The “Present” columns reflect the incidence proportion of brain metastases associated with the cancer described in that row; “Absent”, absence of brain metastases; and N/A, missing. * indicates use of crude values rather than annual or average annual age-adjusted values, and values contained in parentheses reflect estimated 95% confidence intervals. “Her2” stands for Human Epidermal Growth

Factor Receptor 2, “Hr” stands for hormone receptors, reflecting either progesterone receptor or estrogen receptor expression status. SCLC stands for small-cell lung cancer; NSCLC, non-small-cell lung cancer.

92 Site Histology SEER At risk Pres. Abs. NA Medicare At risk Pres. Abs.

SBM LBM

AAIR % % % AAIR % %

Lung Overall 9422 70974 9.6 84.5 5.9 13255 120405 13.5 86.5

Adenocarcinoma 11449 23801 11.8 83.3 4.9 15059 38102 15.5 84.5

Carcinoma 12028 5127 11.1 77.8 11 12829 9339 11.9 88.1

NSCLC 13336 5547 13.4 80.5 6.1 15011 11990 15.3 84.7

Other 6704 11819 6.6 84.7 8.7 9610 19911 9.4 90.6

SCLC 12607 8467 13.5 81 5.5 21539 14476 23.1 76.9

Squamous 4427 16213 4.6 91.7 3.8 7920 26587 8.1 91.9

Breast Overall 309 67362 0.3 98 1.7 1790 110983 1.8 98.2

Her2-/Hr+ * 20.9 41135 0.2 98.7 1.1 1065 41135 1.1 98.9

Her2+/ Hr(+/-) * 52.5 5900 0.5 97.8 1.7 3017 5900 3.1 96.9

Other * 36.7 15250 0.4 96.1 3.6 1996 20406 1.7 98.3

Triple Negative * 70.9 5077 0.7 98 1.3 4103 5077 4.2 95.8

Skin Overall 1049 21860 1.1 97.2 1.8 3583 35268 3.6 96.4

Mal. Mel. In ** 7382 ** ** ** 1758 12053 1.8 98.2

Junct. Nevus

93 Nevi & 1473 13862 1.5 96.3 2.2 4380 22041 4.3 95.7

Melanomas

Other 2903 616 2.8 94.6 2.6 7530 1174 7.7 92.3

94 Table 2-4 - Lung cancer patient demographics and clinical characteristics

This table shows summary statistics for lung cancer cases diagnosed in 2010 through 2012. The column labeled “SEER Synchronous” reflects patients diagnosed with brain metastases during staging workup, and the column “Medicare Lifetime” reflects patients with evidence of brain metastases at any time throughout available

Medicare claims data. None of the columns in this table are mutually exclusive.

Continuous data are presented as “mean (standard deviation)”; categorical, “count

(%)”.

95 Variable Overall SEER Medicare Lifetime

Synchronous n 70974 6789 9440

Age - continuous 75.27 (6.96) 73.40 (6.33) 73.33 (6.30)

Race

American Indian 232 ( 0.3) 24 ( 0.4) 37 ( 0.4)

Asian/Pacific Islander 4043 ( 5.7) 436 ( 6.4) 536 ( 5.7)

Black 6484 ( 9.1) 704 ( 10.4) 944 ( 10.0)

White Hispanic 3708 ( 5.2) 338 ( 5.0) 451 ( 4.8)

White Non-Hispanic 56410 ( 79.6) 5284 ( 77.9) 7462 ( 79.1)

Histology

Adenocarcinoma, Nos 23801 ( 33.5) 2806 ( 41.3) 3748 ( 39.7)

Carcinoma, Nos 5127 ( 7.2) 571 ( 8.4) 599 ( 6.3)

Non-Small Cell Carcinoma, 5547 ( 7.8) 745 ( 11.0) 831 ( 8.8)

Nos

Other 11819 ( 16.7) 784 ( 11.5) 1074 ( 11.4)

Small Cell Carcinoma, Nos 8467 ( 11.9) 1144 ( 16.9) 1915 ( 20.3)

96 Squamous Cell Carcinoma, 16213 ( 22.8) 739 ( 10.9) 1273 ( 13.5)

Nos

Size in mm 43.10 (32.62) 48.01 (29.01) 46.26 (27.77)

SEER Stage

Distant 39100 ( 55.1) 6789 (100.0) 7416 ( 78.6)

In situ 180 ( 0.3) 0 ( 0.0) NA

Localized 13102 ( 18.5) 0 ( 0.0) 552 ( 5.8)

Regional, direct 5634 ( 7.9) 0 ( 0.0) 306 ( 3.2)

Regional, extension and 4987 ( 7.0) 0 ( 0.0) 473 ( 5.0)

nodes

Regional, lymph nodes only 5769 ( 8.1) 0 ( 0.0) 555 ( 5.9)

Unknown 2202 ( 3.1) 0 ( 0.0) 128 ( 1.4)

Bone metastases: Yes 11288 ( 16.9) 1962 ( 29.8) 2383 ( 26.3)

Liver metastases: Yes 7828 ( 11.7) 1338 ( 20.3) 1495 ( 16.6)

97 Table 2-5 - Breast cancer patient demographics and clinical characteristics

This table shows summary statistics for breast cancer cases diagnosed in 2010 through 2012. The column labeled “SEER Synchronous” reflects patients diagnosed with brain metastases during staging workup, and the column “Medicare Lifetime” reflects patients with evidence of brain metastases at any time throughout available

Medicare claims data. None of the columns in this table are mutually exclusive.

Continuous data are presented as “mean (standard deviation)”; categorical, “count

(%)”.

98 Variable Overall SEER Medicare

Synchronous Lifetime n 67096 203 1009

Age - continuous 74.21 (7.12) 74.09 (6.79) 73.43 (6.81)

Race

Black 6190 ( 9.3) 29 ( 14.3) 134 (13.3)

Other 4273 ( 6.4) NA 31 ( 3.1)

White Hispanic 4709 ( 7.0) 23 ( 11.3) 62 ( 6.2)

White Non-Hispanic 51667 (77.3) 146 ( 71.9) 780 (77.5)

Histology

Adenoid Carcinoma 1169 ( 1.7) NA 0 ( 0.0)

Duct Carcinoma 44967 (67.0) 124 ( 61.1) 705 (69.9)

Lobular And Other Ductal Ca. 15187 (22.6) 21 ( 10.3) 159 (15.8)

Mucinous Adenocarcinoma 1551 ( 2.3) NA NA

Other 4222 ( 6.3) 54 ( 26.6) 137 (13.6)

Size in mm 21.13 (21.60) 48.80 (38.83) 39.96 (30.98)

SEER Stage

99 Distant 3334 ( 5.0) 203 (100.0) 432 (42.8)

In situ 10929 (16.3) 0 ( 0.0) 31 ( 3.1)

Localized 38096 (56.8) 0 ( 0.0) 232 (23.0)

Regional, direct 1481 ( 2.2) 0 ( 0.0) 25 ( 2.5)

Regional, extension and 2069 ( 3.1) 0 ( 0.0) 75 ( 7.4)

nodes

Regional, lymph nodes only 10468 (15.6) 0 ( 0.0) 191 (18.9)

Unknown 719 ( 1.1) 0 ( 0.0) 23 ( 2.3)

Bone metastases: Yes 2044 ( 3.1) 117 ( 60.0) 275 (28.4)

Liver metastases: Yes 656 ( 1.0) 41 ( 22.0) 86 ( 8.9)

Lung metastases: Yes 1081 ( 1.6) 71 ( 37.8) 149 (15.5)

PR Status

PR Negative 16025 (23.9) 85 ( 41.9) 469 (46.5)

PR Other and unknown 4884 ( 7.3) 36 ( 17.7) 85 ( 8.4)

PR Positive 46187 (68.8) 82 ( 40.4) 455 (45.1)

ER Status

ER Negative 8939 (13.3) 56 ( 27.6) 331 (32.8)

100 ER Other and unknown 4084 ( 6.1) 31 ( 15.3) 75 ( 7.4)

ER Positive 54073 (80.6) 116 ( 57.1) 603 (59.8)

SEER Subtype

Her2+/HR(+/-) 5900 ( 8.8) 31 ( 15.3) 185 (18.3)

Her2-/HR+ 41135 (61.3) 86 ( 42.4) 444 (44.0)

Triple Negative 5077 ( 7.6) 36 ( 17.7) 216 (21.4)

Unknown 14984 (22.3) 50 ( 24.6) 164 (16.3)

101 Table 2-6 - Melanoma demographics and clinical characteristics

This table shows summary statistics for melanoma cases diagnosed in 2010 through 2012. The column labeled “SEER Synchronous” reflects patients diagnosed with brain metastases during staging workup, and the column “Medicare Lifetime” reflects patients with evidence of brain metastases at any time throughout available

Medicare claims data. None of the columns in this table are mutually exclusive.

Continuous data are presented as “mean (standard deviation)”; categorical, “count

(%)”.

102 Variable Overall SEER Medicare

Synchronous Lifetime n 21860 230 685

Age - continuous 75.21 (7.37) 74.11 (7.03) 75.47 (7.09)

Race

Other 234 ( 1.1) NA 18 ( 2.6)

White Hispanic 483 ( 2.3) NA 21 ( 3.1)

White Non-Hispanic 20338 ( 96.6) 216 ( 93.9) 645 ( 94.3)

Histology

Mal. Mel. In Junct. Nevus 7382 ( 33.8) NA 115 ( 16.8)

Nevi & Melanomas 13862 ( 63.4) 206 ( 89.6) 523 ( 76.4)

Other 616 ( 2.8) 17 ( 7.4) 47 ( 6.9)

Size in mm 17.67 (37.45) 6.82 (20.71) 19.50 (44.20)

SEER Stage

Distant 797 ( 3.6) 230 (100.0) 231 ( 33.7)

In situ 8450 ( 38.7) 0 ( 0.0) 27 ( 3.9)

Localized 10503 ( 48.0) 0 ( 0.0) 238 ( 34.7)

Regional, direct 583 ( 2.7) 0 ( 0.0) 38 ( 5.5)

103 Regional, extension and 133 ( 0.6) 0 ( 0.0) 16 ( 2.3)

nodes

Regional, lymph nodes only 761 ( 3.5) 0 ( 0.0) 81 ( 11.8)

Unknown 633 ( 2.9) 0 ( 0.0) 54 ( 7.9)

Bone metastases: Yes 131 ( 0.6) 35 ( 16.2) 38 ( 5.9)

Liver metastases: Yes 166 ( 0.8) 50 ( 23.1) 42 ( 6.4)

104 Table 2-7 - Stage-specific incidence proportions of Brain Metastases

Using the lifetime brain metastases identification algorithm that relies on the presence of a diagnosis code for secondary cancer of the CNS and a diagnostic imaging procedure code, each within sixty days of the other, this table shows stage- specific incidence proportions (%) of brain metastases for each primary cancer. The population under consideration includes patients diagnosed during any of the five years (2008-2012) for which data was available; stage is reported according to

SEER Summary Stage 2000, which reflects stage at primary cancer diagnosis.

Values are presented as incidence proportions followed by a numerator and denominator for that calculation. Regional, LN refers to primary cancers that were found to have spread only to regional lymph nodes; Regional, dir, to those that have spread by “direct extension” only; and Regional, ext, to those that had “both direct extension and lymph node involvement”.

Stage Lung Cancers Breast Cancers Melanoma

Distant 18.8 (12,440 / 66,100) 13.6 (748 / 5,501) 30.4 (381 / 1,253)

Regional, LN 10.5 (1,067 / 10,149) 2.4 (424 / 17,363) 13.2 (159 / 1,202)

Regional, ext 9.9 (813 / 8,228) 4 (140 / 3,484) 15.2 (35 / 230)

Regional, dir 5.9 (544 / 9,158) 2 (54 / 2,697) 7.8 (76 / 969)

Unknown 6.7 (267 / 3,963) 4.2 (54 / 1,287) 9.5 (101 / 1,068)

In situ 7.3 (24 / 330) 0.4 (68 / 18,200) 0.4 (59 / 13,412)

Localized 4.8 (1,068 / 22,455) 0.9 (533 / 62,451) 2.6 (448 / 17,134)

105 Table 3-1: NSCLC patient characteristics by SBM and bevacizumab treatment status

NSCLC patient characteristics by SBM and bevacizumab treatment status, SEER-

Medicare data 2010-2012. An asterisk (*) indicates censoring where fewer than 11 observations were available.

106 Bevacizumab No Bevacizumab

SBM No SBM SBM No SBM

N 81 666 4343 35197

Age (mean (sd)) 69.4 (3.5) 71.1 (4.3) 71.5 (4.4) 72.3 (4.5)

Nonwhite (%) 15 ( 18.5) 153 ( 23.0) 1,060 ( 24.4) 7,755 ( 22.0)

Histology

Adenocarcinoma 61 ( 75.3) 471 ( 70.7) 2,195 ( 50.5) 13,462 ( 38.2)

Squamous Cell * 30 ( 4.5) 600 ( 13.8) 10,845 ( 30.8)

Other * 165 ( 24.8) 1,548 ( 35.6) 10,890 ( 30.9)

Male (%) 35 ( 43.2) 336 ( 50.5) 2,293 ( 52.8) 18,773 ( 53.3)

Stage (%)

Stage IA 0 ( 0.0) 25 ( 3.8) 0 ( 0.0) 5,423 ( 15.4)

Stage IB 0 ( 0.0) 35 ( 5.3) 0 ( 0.0) 4,392 ( 12.5)

Stage IIA 0 ( 0.0) * 0 ( 0.0) 444 ( 1.3)

Stage IIB 0 ( 0.0) 20 ( 3.0) 0 ( 0.0) 1,567 ( 4.5)

Stage IIIA 0 ( 0.0) 51 ( 7.7) 0 ( 0.0) 3,740 ( 10.6)

Stage IIIB 0 ( 0.0) 122 ( 18.3) 0 ( 0) 5,801 ( 16.5)

Stage IV 81 (100.0) 396 ( 59.5) 4,343 ( 99.3) 12,759 ( 36.3)

107 Stage Occult 0 ( 0.0) * 0 ( 0.0) 449 ( 1.3)

Stage Unknown 0 ( 0.0) * 0 ( 0.0) 622 ( 1.8)

Not applicable 0 ( 0.0) 25 ( 3.8) 0 ( 0.0) 5,423 ( 15.4)

Bone Metastases (%)

None 56 ( 69.1) * 2,882 ( 66.4) 29,918 ( 85.0)

Not applicable 0 ( 0.0) * 0 ( 0.0) *

Unknown 0 ( 0.0) * 129 ( 3.0) *

Yes 25 ( 30.9) 198 ( 29.7) 1,332 ( 30.7) 5,116 ( 14.5)

Liver Metastases (%)

None * * 3,436 ( 79.1) 32,307 ( 91.8)

Not applicable * * 0 ( 0.0) 0 ( 0.0)

Unknown * * 145 ( 3.3) 178 ( 0.5)

Yes * 62 ( 9.3) 762 ( 17.5) 2,712 ( 7.7)

Cranial SRS (%) * 529 ( 79.4) 2,631 ( 60.6) 20,465 ( 58.1)

Any Radiotherapy (%) 81 (100.0) 610 ( 91.6) 2,927 ( 67.4) 22,537 ( 64.0)

Carboplatin (%) 65 ( 80.2) 566 ( 85.0) 860 ( 19.8) 7,459 ( 21.2)

Cisplatin (%) * 90 ( 13.5) 106 ( 2.4) 1,783 ( 5.1)

108 Dexamethasone (%) 69 ( 85.2) 581 ( 87.2) 1,229 ( 28.3) 10,908 ( 31.0)

Paclitaxel (%) 39 ( 48.1) 338 ( 50.8) 421 ( 9.7) 4,682 ( 13.3)

Pemetrexed (%) 51 ( 63.0) 501 ( 75.2) 517 ( 11.9) 3,152 ( 9.0)

109 Table 3-2: Hazard ratios of mortality associated with the use of Bevacizumab

Hazard ratios of mortality associated with the use of bevacizumab for NSCLC patients overall and those with synchronous brain metastases. Hazard ratios are presented for matched and unmatched populations, as well as for univariable and multivariable models. An asterisk (‘*’) indicates the use of weighted estimation of average hazard ratios, which places greater weight on events occurring sooner after the index date, rather than Cox proportional hazards modeling due to non- proportional hazards.

Population Matching Univariable Multivariable

NSCLC, Overall Unmatched 0.73 (0.66-0.8 p: <0.001)* 0.88 (0.81-0.96 p:0.0032)

Matched 0.72 (0.64-0.82 p: <0.001)

NSCLC, Unmatched 0.47 (0.37-0.6 p: <0.001) 0.75 (0.59-0.96 p:0.0198)

Synchronous

Brain

Metastases

Matched 0.69 (0.5-0.96 p:0.0293)

110 Table 3-3: Odds of bevacizumab prescription among non small-cell carcinoma patients with brain metastases

Odds of bevacizumab prescription among NSCLC synchronous brain metastases patients diagnosed in the years 2010 through 2012; Reference levels include 65 to

70 years of age, non-adenocarcinoma histology, no medication administered, female sex, and no surgical management. For categories with more than two possible values, asterisks are used to denote reference levels.

Univariable Multivariable

Age: 65 to 70 * *

Age: 70 to 74 0.66 (0.39-1.09, p: 0.1124) 0.63 (0.36-1.05, p: 0.0860)

Age: 75+ 0.25 (0.09-0.53, p: 0.0012) 0.31 (0.12-0.70, p: 0.0097)

Cisplatin 5.17 (2.35-10.13, p < 0.0001) 1.46 (0.63-3.03, p: 0.3417)

Dexamethasone 13.74 (7.69-26.80, p < 0.0001) 2.92 (1.42-6.36, p: 0.0048)

Adenocarcinoma 2.79 (1.70-4.76, p: 0.0001) 1.63 (0.94-2.90, p: 0.0869)

Male 0.66 (0.42-1.03, p: 0.0723) 0.74 (0.46-1.20, p: 0.2235)

Paclitaxel 7.92 (4.99-12.49, p < 0.0001) 4.47 (2.65-7.56, p < 0.001)

Pemetrexed 13.94 (8.74-22.73, p < 0.0001) 6.06 (3.45-10.90, p < 0.001)

111 Table 3-4: Odds of bevacizumab prescription among all non small-cell carcinoma patients

Odds of bevacizumab prescription among all NSCLC patients diagnosed in the years 2010 through 2012. Reference levels include 65 to 70 years of age, stage I disease, non-adenocarcinoma histology, no medication administered, white race, female sex, and no surgical management. For categories with more than two possible values, asterisks are used to denote reference levels.

112 Univariable Multivariable

Age: 65 to 70 * *

Age: 70 to 74 0.72 (0.61-0.85, p: 0.0001) 0.77 (0.64-0.92, p: 0.0038)

Age: 75+ 0.51 (0.42-0.62, p: 0.0000) 0.76 (0.62-0.94, p: 0.0106)

Stage I * *

Stage II 1.95 (1.19-3.10, p: 0.0059) 0.81 (0.49-1.32, p: 0.4092)

Stage III 2.97 (2.22-4.02, p: 0.0000) 0.91 (0.65-1.28, p: 0.5742)

Stage IV 4.61 (3.55-6.10, p: 0.0000) 1.49 (1.09-2.08, p: 0.0152)

Stage Occult 2.55 (1.06-5.24, p: 0.0200) 1.49 (0.58-3.29, p: 0.3645)

Stage Unknown 1.16 (0.35-2.83, p: 0.7722) 0.54 (0.16-1.38, p: 0.2473)

Adenocarcinoma 3.74 (3.19-4.40, p: 0.0000) 1.68 (1.41-2.01, p: 0.0000)

Carboplatin 20.27 (16.66-24.89, p: 0.0000) 2.56 (1.96-3.37, p: 0.0000)

Dexamethasone 15.03 (12.18-18.75, p: 0.0000) 2.00 (1.53-2.64, p: 0.0000)

Paclitaxel 6.81 (5.88-7.89, p: 0.0000) 2.47 (2.07-2.95, p: 0.0000)

Pemetrexed 28.00 (23.71-33.20, p: 0.0000) 8.93 (7.27-11.00, p: 0.0000)

Non-White 1.00 (0.84-1.19, p: 0.9587) 1.18 (0.98-1.43, p: 0.0778)

Male 0.86 (0.74-1.00, p: 0.0449) 0.88 (0.75-1.03, p: 0.1083)

113 Surgery 0.46 (0.37-0.56, p: 0.0000) 0.94 (0.72-1.21, p: 0.6352)

114 Table 3-5: Hazard ratios of mortality among the NSCLC synchronous brain metastases patients, SEER-Medicare dataset 2010-2012.

When considering only patients diagnosed with brain metastases during staging workup for primary cancer, the above shows ratios for hazard of mortality. After adjusting for clinical and demographic characteristics, patients with brain metastases who were treated with bevacizumab were found to have 0.75 times the hazard of mortality at any given time, compared to the corresponding untreated population. For categories with more than two possible values, asterisks are used to denote reference levels.

115 Univariable Multivariable

Bevacizumab 0.47 (0.37-0.60, p: 0.0000) 0.75 (0.59-0.96, p:0.0194)

Age: 65 to 70 * *

Age: 70 to 74 1.10 (1.02-1.19, p: 0.0105) 1.14 (1.06-1.23, p:0.0009)

Age: 75+ 1.32 (1.21-1.43, p: 0.0000) 1.26 (1.16-1.38, p:0.0000)

Metropolitan 0.93 (0.85-1.01, p: 0.0883) 0.89 (0.81-0.97, p:0.0064)

Non-White 0.88 (0.82-0.95, p: 0.0010) 0.82 (0.76-0.89, p:0.0000)

Adenocarcinoma 0.75 (0.70-0.80, p: 0.0000) 0.80 (0.74-0.85, p:0.0000)

Male 1.14 (1.07-1.21, p: 0.0001) 1.13 (1.06-1.20, p:0.0002)

Carboplatin 0.57 (0.53-0.61, p: 0.0000) 0.77 (0.67-0.88, p:0.0001)

Cisplatin 0.57 (0.47-0.69, p: 0.0000) 0.67 (0.55-0.82, p:0.0001)

Dexamethasone 0.63 (0.59-0.68, p: 0.0000) 0.96 (0.87-1.06, p:0.4845)

Paclitaxel 0.62 (0.56-0.69, p: 0.0000) 0.78 (0.68-0.89, p:0.0003)

Pemetrexed 0.51 (0.46-0.56, p: 0.0000) 0.69 (0.61-0.78, p:0.0000)

116 Table 3-6: Hazard ratios of mortality in the overall NSCLC population, SEER-

Medicare dataset 2010-2012.

Reference levels include 65 to 70 years of age, stage I disease, non- adenocarcinoma histology, no medication administered, white race, female sex, non-metropolitan diagnosing or treating facility, and no surgical management.

117 Univariable Multivariable

Bevacizumab 0.73 (0.66-0.80, p: 0.0000) 0.88 (0.81-0.96, p:0.0032)

Age: 65 to 70 * *

Age: 70 to 74 1.10 (1.07-1.13, p: 0.0000) 1.12 (1.08-1.15, p:0.0000)

Age: 75+ 1.21 (1.18-1.25, p: 0.0000) 1.25 (1.22-1.29, p:0.0000)

Stage I * *

Stage II 1.84 (1.72-1.96, p: 0.0000) 2.00 (1.87-2.13, p:0.0000)

Stage III 3.10 (2.98-3.23, p: 0.0000) 3.49 (3.35-3.64, p:0.0000)

Stage IV 5.96 (5.74-6.18, p: 0.0000) 6.88 (6.62-7.15, p:0.0000)

Stage Occult 2.97 (2.64-3.35, p: 0.0000) 3.07 (2.72-3.46, p:0.0000)

Stage Unknown 3.75 (3.39-4.14, p: 0.0000) 3.96 (3.58-4.37, p:0.0000)

Metropolitan 0.86 (0.84-0.89, p: 0.0000) 0.86 (0.83-0.89, p:0.0000)

Non-White 1.04 (1.01-1.07, p: 0.0191) 0.93 (0.90-0.96, p:0.0000)

Adenocarcinoma 0.96 (0.94-0.99, p: 0.0020) 0.88 (0.86-0.90, p:0.0000)

Male 1.28 (1.25-1.31, p: 0.0000) 1.25 (1.22-1.28, p:0.0000)

Carboplatin 1.08 (1.05-1.11, p: 0.0000) 0.89 (0.85-0.94, p:0.0000)

Cisplatin 0.74 (0.70-0.79, p: 0.0000) 0.72 (0.68-0.76, p:0.0000)

118 Dexamethasone 0.95 (0.93-0.97, p: 0.0001) 0.93 (0.89-0.96, p:0.0001)

Paclitaxel 1.05 (1.02-1.09, p: 0.0026) 0.92 (0.88-0.97, p:0.0012)

Pemetrexed 1.00 (0.97-1.04, p: 0.8968) 0.81 (0.77-0.85, p:0.0000)

119 Table 3-7: Missing covariate values in the overall NSCLC population, SEER-

Medicare dataset 2010-2012.

Data element Proportion missing (%)

Age 0.00%

Bone Metastases 4.73%

Liver Metastases 4.84%

Brain Metastases 5.00%

Derived AJCC Stage 0.00%

Urban/Rural classification 0.02%

Race 0.00%

Histology 0.00%

Sex 0.00%

120 Table 4-1 Clinical and demographic characteristics, SEER-Medicare lung cancer population 2010-2012

No SBM:No SBM:No SRS No SBM:SRS SBM:SRS

SRS n 53,892 5,933 11,058 3,259

Age at Diagnosis 75.1 (6.9) 73.4 (6.3) 75.4 (6.8) 72.0 (5.4)

(mean (sd))

Race (%)

American Indian 167 ( 0.3) 22 ( 0.4) 49 ( 0.4) 11 ( 0.3)

Asian/Pacific Islander 3,186 ( 5.9) 400 ( 6.7) 672 ( 6.1) 202 ( 6.2)

Black 5,029 ( 9.3) 644 ( 10.9) 722 ( 6.5) 250 ( 7.7)

White Hispanic 2,815 ( 5.2) 301 ( 5.1) 423 ( 3.8) 219 ( 6.7)

White Non-Hispanic 42,619 ( 79.2) 4,563 ( 76.9) 9,178 ( 83.1) 2,577 ( 79.1)

Histology (%)

Adenocarcinoma 17,846 ( 33.1) 2,374 ( 40.0) 4,515 ( 40.8) 1,953 ( 59.9)

Carcinoma 3,613 ( 6.7) 516 ( 8.7) 519 ( 4.7) 128 ( 3.9)

Non-Small Cell 3,899 ( 7.2) 646 ( 10.9) 998 ( 9.0) 358 ( 11.0)

Carcinoma

Small Cell Carcinoma 6,299 ( 11.7) 1,049 ( 17.7) 463 ( 4.2) 219 ( 6.7)

121 Squamous Cell 13,328 ( 24.7) 655 ( 11.0) 3,035 ( 27.4) 349 ( 10.7)

Carcinoma

Other 8,907 ( 16.5) 693 ( 11.7) 1,528 ( 13.8) 252 ( 7.7)

Male (%) 28,071 ( 52.1) 3,098 ( 52.2) 4,916 ( 44.5) 1,592 ( 48.8)

Geo. Classification (%)

Metro 45,484 ( 84.4) 4,996 ( 84.2) 9,547 ( 86.4) 2,790 ( 85.6)

Urban 2,653 ( 4.9) 312 ( 5.3) 491 ( 4.4) 177 ( 5.4)

Not Metro 5,751 ( 10.7) 623 ( 10.5) 1,011 ( 9.2) 292 ( 9.0)

Stage (%)

I 12,412 ( 23.0) 0 ( 0.0) 5,606 ( 50.7) 0 ( 0.0)

II 3,181 ( 5.9) 0 ( 0.0) 625 ( 5.7) 0 ( 0.0)

III 14,765 ( 27.4) 39 ( 0.7) 2,251 ( 20.4) 21 ( 0.6)

IV 21,316 ( 39.6) 5,894 ( 99.3) 2,156 ( 19.5) 3,238 ( 99.4)

Other 1,912 ( 3.5) 0 ( 0.0) 301 ( 2.7) 0 ( 0.0)

Missing 306 ( 0.6) 0 ( 0.0) 119 ( 1.1) 0 ( 0.0)

Bone Metastases (%)

None 45,424 ( 84.3) 4,019 ( 67.7) 10,426 ( 94.3) 2,388 ( 73.3)

Not applicable * * 0 ( 0.0) 0 ( 0.0)

122 Unknown * * 17 ( 0.2) 126 ( 3.9)

Yes 8,159 ( 15.1) 1,742 ( 29.4) 615 ( 5.6) 745 ( 22.9)

Liver Metastases (%)

None 47,897 ( 88.9) 4,519 ( 76.2) 10,862 ( 98.2) 2,829 ( 86.8)

Not applicable * * * 0 ( 0.0)

Unknown * * * 96 ( 2.9)

Yes 5,665 ( 10.5) 1,219 ( 20.5) 187 ( 1.7) 334 ( 10.2)

Neurosurgery (%) 33 ( 0.1) 29 ( 0.5) 121 ( 1.1) 82 ( 2.5)

Table 4-1 Clinical and demographic characteristics by SRS treatment and presence of BM

Table 4-1 provides summary statistics of clinical and demographic characteristics, stratified by the use of SRS and presence of SBM. … In the detailed description of stage, Stage IC, IIC and Stages I, II, III, and IV of the NOS classification were removed because all summary values were either zero or censored.

123 Table 4-2 Odds of stereotactic radiosurgery among lung cancer patients with brain metastases, SEER-Medicare population 2010-2012

Univariable Multivariable

Age: 70 to 74 1.03 (0.93-1.14 p: 0.603) 1.04 (0.93-1.15 p: 0.517)

Age: 75+ 0.63 (0.57-0.70 p: 0.000) 0.63 (0.57-0.70 p: 0.000)

Histology: Other 0.55 (0.47-0.64 p: 0.000) 0.50 (0.42-0.58 p: 0.000)

Histology: SCLC 0.31 (0.27-0.37 p: 0.000) 0.31 (0.27-0.36 p: 0.000)

Location: Urban 1.02 (0.84-1.23 p: 0.876) 0.97 (0.79-1.18 p: 0.737)

Location: Not Metro 0.84 (0.72-0.97 p: 0.019) 0.83 (0.72-0.97 p: 0.020)

Race: A/PI 0.89 (0.75-1.07 p: 0.213) 0.92 (0.77-1.11 p: 0.392)

Race: Black 0.69 (0.59-0.80 p: 0.000) 0.66 (0.57-0.78 p: 0.000)

Race: Other 1.26 (1.06-1.50 p: 0.010) 1.23 (1.03-1.48 p: 0.025)

Sex: Male 0.87 (0.80-0.95 p: 0.002) 0.85 (0.77-0.92 p: 0.000)

Surgery 3.75 (2.98-4.71 p: 0.000) 3.86 (3.04-4.90 p: 0.000)

Table 4-2 describes the odds of stereotactic radiosurgery among age-eligible

Medicare patients diagnosed with lung cancer and brain metastases between the years 2010-2012. Reference levels for age were 65-70 years; histology, NSCLC; and for location, Big Metro.

124 Table 4-3 Odds of stereotactic radiosurgery in the overall lung cancer patient population, SEER-Medicare population 2010-2012

Univariable Multivariable

Age: 70 to 74 0.97 (0.92-1.02 p: 0.197) 0.89 (0.85-0.93 p: 0.000)

Age: 75+ 0.88 (0.84-0.92 p: 0.000) 0.68 (0.65-0.71 p: 0.000)

Histology: Other 0.67 (0.64-0.71 p: 0.000) 0.70 (0.66-0.74 p: 0.000)

Histology: SCLC 0.33 (0.30-0.36 p: 0.000) 0.35 (0.32-0.38 p: 0.000)

Location: Urban 0.93 (0.85-1.01 p: 0.076) 0.88 (0.81-0.96 p: 0.006)

Location: Not Metro 0.82 (0.77-0.88 p: 0.000) 0.78 (0.73-0.84 p: 0.000)

Race: A/PI 0.95 (0.88-1.03 p: 0.199) 1.01 (0.93-1.09 p: 0.859)

Race: Black 0.70 (0.65-0.75 p: 0.000) 0.64 (0.60-0.69 p: 0.000)

Race: Other 0.86 (0.79-0.93 p: 0.000) 0.87 (0.80-0.95 p: 0.001)

Sex: Male 0.77 (0.74-0.80 p: 0.000) 0.76 (0.74-0.79 p: 0.000)

Stage:II 0.44 (0.41-0.49 p: 0.000) 0.45 (0.41-0.50 p: 0.000)

Stage:III 0.33 (0.31-0.35 p: 0.000) 0.22 (0.21-0.24 p: 0.000)

Stage:IV 0.43 (0.41-0.44 p: 0.000) 0.27 (0.25-0.28 p: 0.000)

Stage:Occult 0.27 (0.24-0.30 p: 0.000) 0.17 (0.16-0.19 p: 0.000)

125 Surgery 0.64 (0.61-0.67 p: 0.000) 0.27 (0.26-0.29 p: 0.000)

Table 4-3 describes the odds of stereotactic radiosurgery among all age-eligible

Medicare patients diagnosed with lung cancer between the years 2010-2012.

Reference levels for age were 65-70 years; histology, NSCLC; location, Big Metro; and for stage, Stage I.

126 Table 4-4 Hazards of mortality for lung cancer patients with synchronous brain metastases, SEER-Medicare population 2010-2012

Univariable Multivariable

Age: 70 to 74 1.01 (0.96-1.06, p: 0.6870) 1.10 (1.04-1.15 p: 0.0006)

Age: 75+ 1.44 (1.37-1.51, p: 0.0000) 1.45 (1.38-1.52 p: 0.0000)

Histology: NOS 1.19 (1.11-1.28, p: 0.0000) 1.10 (1.03-1.18 p: 0.0062)

Histology: SCLC NOS 1.19 (1.12-1.27, p: 0.0000) 0.93 (0.87-0.99 p: 0.0197)

Location: Urban 0.93 (0.85-1.02, p: 0.1255) 0.93 (0.85-1.03 p: 0.1565)

Location: Not Metro 1.14 (1.06-1.22, p: 0.0004) 1.10 (1.03-1.18 p: 0.0066)

Race: A/PI 0.68 (0.62-0.74, p: 0.0000) 0.61 (0.55-0.66 p: 0.0000)

Race: Black 1.18 (1.10-1.27, p: 0.0000) 1.04 (0.97-1.12 p: 0.2338)

Race: Other 0.86 (0.78-0.94, p: 0.0009) 0.88 (0.81-0.97 p: 0.0082)

Sex: Male 1.11 (1.06-1.15, p: 0.0000) 1.13 (1.09-1.18 p: 0.0000)

SRS 0.36 (0.34-0.38, p: 0.0000) 0.36 (0.34-0.37 p: 0.0000)

Table 4-4 describes the hazards of mortality associated with various patient characteristics, adjusting for the use of stereotactic radiosurgery. The population includes age-eligible Medicare patients diagnosed with lung cancer and synchronous brain metastases between the years 2010 and 2012. Reference levels for age were 65-70 years; histology, NSCLC; and location, Big Metro. 127 Table 4-5 Hazards of mortality for the overall lung cancer patient population,

SEER-Medicare population 2010-2012

Univariable Multivariable

Age: 70 to 74 1.09 (1.07-1.12, p: 0.0000) 1.15 (1.12-1.17 p: 0.0000)

Age: 75+ 1.35 (1.32-1.37, p: 0.0000) 1.56 (1.53-1.59 p: 0.0000)

Histology: Other 0.83 (0.81-0.85, p: 0.0000) 0.94 (0.92-0.96 p: 0.0000)

Histology: SCLC NOS 1.69 (1.65-1.73, p: 0.0000) 1.25 (1.22-1.28 p: 0.0000)

Location: Urban 1.10 (1.06-1.14, p: 0.0000) 1.07 (1.03-1.11 p: 0.0002)

Location: Not_Metro 1.15 (1.12-1.18, p: 0.0000) 1.12 (1.09-1.15 p: 0.0000)

Race: A/PI 0.85 (0.82-0.88, p: 0.0000) 0.76 (0.73-0.78 p: 0.0000)

Race: Black 1.10 (1.07-1.13, p: 0.0000) 1.04 (1.01-1.07 p: 0.0099)

Race: Other 1.07 (1.03-1.11, p: 0.0001) 0.99 (0.96-1.02 p: 0.5365)

Sex: Male 1.23 (1.22-1.25, p: 0.0000) 1.21 (1.19-1.23 p: 0.0000)

SRS 0.52 (0.51-0.53, p: 0.0000) 0.54 (0.52-0.55 p: 0.0000)

Stage:II 1.53 (1.46-1.59, p: 0.0000) 1.44 (1.38-1.51 p: 0.0000)

Stage:III 2.69 (2.61-2.76, p: 0.0000) 2.55 (2.48-2.62 p: 0.0000)

Stage:IV 4.65 (4.54-4.76, p: 0.0000) 4.70 (4.58-4.82 p: 0.0000)

128 Stage:Occult 3.27 (3.15-3.40, p: 0.0000) 3.03 (2.91-3.14 p: 0.0000)

Table 4-5 describes the hazards of mortality associated with various patient characteristics, adjusting for the use of SRS. The population includes all age-eligible

Medicare patients diagnosed with lung cancer between the years 2010 and 2012.

Reference levels for age were 65-70 years; histology, NSCLC; location, Big Metro; and stage, Stage I.

129 Figures

Figure 2-1: Cancer registry and Medicare claims data selection, SEER-

Medicare population 2008-2012

Figure 2.1 shows the years available for each of synchronous and lifetime brain metastasis data sources. Lifetime brain metastasis includes both synchronous brain metastasis diagnosis and cases diagnosed thereafter, whereas synchronous brain metastasis are those diagnosed during staging workup. Data source restrictions led to the use of cases diagnosed in 2010-2012 for estimates of synchronous brain metastasis, but 2008 - 2012 for lifetime estimates.

130 Figure 2-2: Lung and breast cancer incidence proportions of brain metastases by race, SEER-Medicare population 2008 to 2012

This figure illustrates incidence proportions of synchronous brain metastasis as found in SEER data (SEER SBM) and lifetime brain metastases as found in

Medicare data (Medicare LBM). In addition to stratification by race, lung cancer incidence proportions are stratified by histology and breast cancers stratified by molecular subtype. “AI” stands for American Indian/Native American; “API”,

Asian/Pacific Islander; “WH”, White Hispanic; and “WNH”, White Non-Hispanic.

Incidence proportions are not shown where fewer than 11 subjects were available for that proportion. “Her2” stands for Human Epidermal Growth Factor Receptor 2,

“Hr” stands for hormone receptors, reflecting either progesterone receptor or estrogen receptor expression status. Due to data availability, SEER SBM and all breast cancer cases reflect patients diagnosed between 2010 and 2012, and

Medicare LBM cases reflect 2008-2012 data.

131

132 Figure 2-3: Lung and melanoma incidence proportions by sex, SEER-Medicare population 2008-2012

This figure illustrates incidence proportions of synchronous brain metastasis as found in SEER data (SEER SBM) and lifetime brain metastases as found in

Medicare data (Medicare LBM). Each of lung cancers and melanomas are presented, stratified by histology and sex. Incidence proportions are not shown where fewer than 11 subjects were available for that proportion. “MMJN” stands for malignant melanoma in junctional nevi; “N & M”, nevi and melanomas. Due to data availability, SEER SBM cases reflect patients diagnosed between 2010 and 2012 and Medicare LBM cases reflect 2008-2012 data.

133 Figure 2-4: Population selection diagram, SEER-Medicare population 2008-

2012

This figure shows the numbers of primary cancer cases available (labeled

“initial”), the exclusion criteria applied to these cases, and the final selected number of cases (labeled “selected”). Exclusion criteria are organized by restrictions imposed due to the use of Medicare data as opposed to more general exclusion criteria. Note that these numbers of subjects meeting exclusion criteria are not mutually exclusive, and these values reflect cases diagnosed in 2008-2012.

134 Figure 3-1: Kaplan-Meier estimates of survival among non small-cell carcinoma lung cancer patients with brain metastases, SEER-Medicare population 2010-2012

Estimated survival among the NSCLC synchronous brain metastasis population, SEER-Medicare dataset 2010-2012; For NSCLC patients diagnosed with brain metastases, the top panel in this figure shows a Kaplan-Meier estimate of survival for each group of patients treated with and without bevacizumab, and the middle panel shows a conditional survival curve in which each observation is weighted with respect to odds of bevacizumab treatment. At the bottom, a table describes the Kaplan-Meier estimated proportion of patients at risk is provided, with asterisks (‘*’) indicating censored values.

135

136 Figure 3-2: Kaplan-Meier Survival estimates, overall SEER-Medicare population 2010-2012

Estimated survival among NSCLC patients, SEER-Medicare dataset 2010-

2012; For all NSCLC patients, the top panel in this figure shows a Kaplan-Meier estimate of survival for each group of patients treated with and without bevacizumab, and the middle panel shows a conditional survival curve in which each observation is weighted with respect to odds of bevacizumab treatment. At the bottom, a table of the Kaplan-Meier estimated proportion of patients at risk is provided. Due to non-proportional hazards, weighted estimation of hazard ratios was performed to yield an average hazard ratio of mortality associated with the use of bevacizumab.

137

138 Figure 3-3: Population selection diagram, SEER-Medicare population 2010-

2012

This figure describes the population selection procedure required to identify Non-

Small Cell Carcinoma (NSCLC) patients with brain metastases (BM). Small-Cell

Lung Carcinoma patients were excluded, as were patients who did not meet

Medicare criteria and those patients who were diagnosed on autopsy or had unverified death.

139 Figure 3-4: Propensity score distributions among patients with brain metastases, SEER-Medicare population 2010-2012

This figure shows the distributions of predicted propensity to receive bevacizumab treatment for each of the unmatched (top) and matched (bottom) populations of patients with synchronous brain metastases (SBM). The greater overlap between propensity to receive bevacizumab treatment indicates a successful match between patients treated with bevacizumab (BEV+) and those not treated with bevacizumab

(BEV-).

140 Figure 3-5: Propensity score distributions among all patients, SEER-Medicare population 2010-2012

These plots describe the distributions of predicted propensity to receive bevacizumab treatment for each of the unmatched (top) and matched (bottom) populations of all patients who met inclusion criteria. The greater overlap between propensity to receive bevacizumab treatment indicates a successful match between patients treated with bevacizumab (BEV+) and those not treated with bevacizumab

(BEV-).

141 Figure 3-6: Standardized mean differences before and after matching for both the overall and brain metastases populations, SEER-Medicare population

2010-2012

This is a depiction of the standardized mean differences of potential covariates across matched versus unmatched populations for each of the Synchronous Brain

Metastases (SBM) and overall populations. Points shown in blue reflect the matched populations, while those in gold reflect the unmatched populations. Patient matching was successful, as evidenced by the decreased standardized mean differences in both the overall and SBM populations after matching.

142 Figure 4-1: Survival following lung cancer with synchronous brain metastases

(SBM) diagnosis between 2010-2012 among SEER-Medicare patients by SRS status and race/ethnicity

Figure 4-1 displays Kaplan-Meier estimates of mortality for all lung cancer patients diagnosed with synchronous brain metastases (SBM) (left) and those patients with

SBM who were treated with stereotactic radiosurgery (right), by race/ethnicity.

143 Figure 4-2: Time to stereotactic surgery across race among patients diagnosed with lung cancer and synchronous brain metastases between 2010 and 2012 in the SEER-Medicare population

Figure 4.2 displays Kaplan-Meier estimates of cumulative incidence of stereotactic radiosurgery among patients diagnosed with brain metastases at the same time as primary lung cancer. Here, mortality is considered a right-censoring event and therefore there may be survivorship bias present among those patients treated beyond the median survival time among lung cancer patients reported by Cagney et al (Table 1-1).

144 Bibliography

Chapter 1

Abbott, N. J., Rönnbäck, L. & Hansson, E. Astrocyte-endothelial interactions at the blood-brain barrier. Nat. Rev. Neurosci. 7, 41–53 (2006).

Adams, H., Adams, H.H., Jackson, C., Rincon-Torroella, J., Jallo, G.I. and

Quiñones-Hinojosa, A., 2016. Evaluating extent of resection in pediatric glioblastoma: a multiple propensity score-adjusted population-based analysis.

Child's Nervous System, 32(3), pp.493-503.

Agarwala SS, Kirkwood JM. Temozolomide, a novel alkylating agent with activity in the central nervous system, may improve the treatment of advanced metastatic melanoma. Oncologist. 2000;5(2):144-51.

Alliance for Clinical Trials in Oncology. Corticosteroids + bevacizumab vs. corticosteroids + placebo (best) for radionecrosis after radiosurgery for brain metastases. ClinicalTrials.gov. Bethesda, MD: National Library of Medicine (US)

2015-[cited 2018 Mar 4]. Available at http://clinicaltrials.gov/show/NCT02490878

NLM Identifier: NCT02490878.

Arvold ND, Lee EQ, Mehta MP, et al. Updates in the management of brain metastases. Neuro-oncology. 2016;18(8):1043-65. PMID: 27382120

Awad, R., Fogarty, G., Hong, A., Kelly, P., Ng, D., Santos, D. and Haydu, L., 2013.

145 Hippocampal avoidance with volumetric modulated arc therapy in melanoma brain metastases–the first Australian experience. Radiation oncology, 8(1), p.62.

Bafaloukos, D. and Gogas, H., 2004. The treatment of brain metastases in melanoma patients. Cancer treatment reviews, 30(6), pp.515-520.

Canadian Task Force on Preventive Health Care, 2016. Recommendations on screening for lung cancer. CMAJ, 188(6), pp.425-432.

ClinicalTrials.gov [Internet]. Bethesda (MD): National Library of Medicine (US). 9

Dec 2008 - April 2010. Identifier NCT00805636, Evaluation of the Efficacy and

Safety of [18F]-ML-10, as a PET Imaging Radiotracer, in Early Detection of

Response of Brain Metastases of Solid Tumors to Radiation Therapy. Available from: https://clinicaltrials.gov/ct2/show/NCT00805636

Chao ST, Ahluwalia MS, Barnett GH et al. Challenges with the diagnosis and treatment of cerebral radiation necrosis. Int J Radiat Oncol Biol Phys.

2013;87(3):449–457.

Collins, L.G., Haines, C., Perkel, R. and Enck, R.E., 2007. Lung cancer: diagnosis and management. Am Fam Physician, 75(1), pp.56-63. PMID: 17225705

Dasgupta T, Brasfield R. Metastatic melanoma. A clinicopatho- logical study. Cancer. 1964;17:1323–39.

146

Davies, M.A., Liu, P., McIntyre, S., Kim, K.B., Papadopoulos, N., Hwu, W.J., Hwu,

P. and Bedikian, A., 2011. Prognostic factors for survival in melanoma patients with brain metastases. Cancer, 117(8), pp.1687-1696.

Davis, F. G., Dolecek, T. A., McCarthy, B. J., & Villano, J. L. (2012). Toward determining the lifetime occurrence of metastatic brain tumors estimated from 2007

United States cancer incidence data. Neuro-oncology, 14(9), 1171-7.

de la Monte SM, Moore GW, Hutchins GM. Patterned distribution of metastases from malignant melanoma in humans. Cancer Res. 1983;43:3427–33.

Disibio, G. and French, S.W., 2008. Metastatic patterns of cancers: results from a large autopsy study. Archives of pathology & laboratory medicine, 132(6), pp.931-

939.

Du Four S, Wilgenhof S, Duerinck J et al. Radiation necrosis of the brain in melanoma patients successfully treated with ipilimumab, three case studies. Eur J

Cancer. 2012;48(16):3045–3051.

Eichler, A.F., Chung, E., Kodack, D.P., Loeffler, J.S., Fukumura, D. and Jain, R.K.,

2011. The biology of brain metastases—translation to new therapies. Nature reviews Clinical oncology, 8(6), p.344.

147 Ewend MG, Morris DE, Carey LA, Ladha AM, Brem S. Guidelines for the initial management of metastatic brain tumors: role of surgery, radiosurgery, and radiation therapy. J Natl Compr Cancer Netw 2008; 6:505-513.

Flieger, M., Ganswindt, U., Schwarz, S.B., Kreth, F.W., Tonn, J.C., La Fougère, C.,

Ertl, L., Linn, J., Herrlinger, U., Belka, C. and Niyazi, M., 2014. Re-irradiation and bevacizumab in recurrent high-grade glioma: an effective treatment option. Journal of neuro-oncology, 117(2), pp.337-345.

Frisk, G., Svensson, T., Bäcklund, L.M., Lidbrink, E., Blomqvist, P. and Smedby,

K.E., 2012. Incidence and time trends of brain metastases admissions among breast cancer patients in Sweden. British journal of cancer, 106(11), p.1850.

Gondi, V., Hermann, B.P., Mehta, M.P. and Tomé, W.A., 2013. Hippocampal dosimetry predicts neurocognitive function impairment after fractionated stereotactic radiotherapy for benign or low-grade adult brain tumors. International Journal of

Radiation Oncology* Biology* Physics, 85(2), pp.348-354.

Hirsch, F.R., Hansen, H.H., Paulson, O.B. and Vraa-Jensen, J., 1979. Development of brain metastases in small-cell anaplastic carcinoma of the lung. In CNS complications of malignant disease (pp. 175-184). Palgrave Macmillan, London.

Hirsch, F.R., Paulson, O.B., Hansen, H.H. and Larsen, S.O., 1983. Intracranial metastases in small cell carcinoma of the lung prognostic aspects. Cancer, 51(3),

148 pp.529-533.

Hong, D.S., Kurzrock, R., Falchook, G.S., Andresen, C., Kwak, J., Ren, M., Xu, L.,

George, G.C., Kim, K.B., Nguyen, L.M. and O'Brien, J.P., 2015. Phase 1b study of lenvatinib (E7080) in combination with temozolomide for treatment of advanced melanoma. Oncotarget, 6(40), p.43127.

Hudachek, S.F. and Gustafson, D.L., 2013. Physiologically based pharmacokinetic model of lapatinib developed in mice and scaled to humans. Journal of pharmacokinetics and pharmacodynamics, 40(2), pp.157-176.

Japuntich, S.J., Krieger, N.H., Salvas, A.L. and Carey, M.P., 2018. Racial disparities in lung cancer screening: an exploratory investigation. Journal of the National

Medical Association, 110(5), pp.424-427.

Johnson, C.B., Postow, M.A., Chapman, P., Wolchok, J.D., Brennan, C.W., Tabar,

V.S., Chan, T.A., Yamada, Y., Yang, T.J. and Beal, K., 2018. Safety and Clinical

Outcomes of Ipilimumab and Nivolumab Plus Concurrent Stereotactic Radiosurgery for Brain Metastases. International Journal of Radiation Oncology• Biology• Physics,

102(3), p.e228.

Kienast, Y., Von Baumgarten, L., Fuhrmann, M., Klinkert, W.E., Goldbrunner, R.,

Herms, J. and Winkler, F., 2010. Real-time imaging reveals the single steps of brain metastasis formation. Nature medicine, 16(1), p.116.

149

Kebir, S., Rauschenbach, L., Galldiks, N., Schlaak, M., Hattingen, E., Landsberg, J.,

Bundschuh, R.A., Langen, K.J., Scheffler, B., Herrlinger, U. and Glas, M., 2016.

Dynamic O-(2-[18F] fluoroethyl)-L-tyrosine PET imaging for the detection of checkpoint inhibitor-related pseudoprogression in melanoma brain metastases.

Neuro-oncology, 18(10), pp.1462-1464.

Kocher M, Soffietti R, Abacioglu U, et al. Adjuvant whole-brain radiotherapy versus observation after radiosurgery or surgical resection of one to three cerebral metastases: results of the EORTC 22952-26001 study. J Clin Oncol 2011; 29:134.

Kromer C, Xu J, Ostrom QT, et al. Estimating the annual frequency of synchronous brain metastasis in the united states 2010–2013: A population-based study. Journal of Neuro-Oncology. 2017:1-10.

Lassman AB, Deangelis LM. Brain metastases. Neurol Clin. 2003;21(1):1-23, vii.

PMID: 12690643

Lesser GJ. Chemotherapy of cerebral metastases from solid tumors. Neurosurg Clin

N Am. 1996;7(3):527-36. PMID: 8823780

LeCun, Y., Bengio, Y. and Hinton, G., 2015. Deep learning. Nature, 521(7553), p.436.

150 Lockman PR, Mittapalli RK, Taskar KS et al. Heterogeneous blood-tumor barrier permeability determines drug efficacy in experimental brain metastases of breast cancer. Clin Cancer Res. 2010;16(23):5664–5678.

Loeffler, J.S., Patchell, R.A. and Sawaya, R., 1997. Treatment of metastatic cancer.

Cancer: Principles and Practice of Oncology. 5th ed. Philadelphia: Lippincott–Raven

Publishers, 2523.

Long, G.V., Atkinson, V., Menzies, A.M., Lo, S., Guminski, A.D., Brown, M.P.,

Gonzalez, M.M., Diamante, K., Sandhu, S.K., Scolyer, R.A. and Emmett, L., 2017. A randomized phase II study of nivolumab or nivolumab combined with ipilimumab in patients (pts) with melanoma brain metastases (mets): The Anti-PD1 Brain

Collaboration (ABC).

Mahajan A, Ahmed S, McAleer MF, et al. Post-operative stereotactic radiosurgery versus observation for completely resected brain metastases: a single-centre, randomised, controlled, phase 3 trial. Lancet Oncol 2017; 18:1040.

Mansfield P, Coxon R, Glover P. Echo-planar imaging of the brain at 3.0 T: first normal volunteer results. J Comput Assist Tomogr. 1994;18(3):339-43. PMID:

8188896

151 Martinage, G., Hong, A.M., Fay, M., Thachil, T., Roos, D., Williams, N., Lo, S. and

Fogarty, G., 2018. Quality assurance analysis of hippocampal avoidance in a melanoma whole brain radiotherapy randomized trial shows good compliance.

Radiation Oncology, 13(1), p.132.

Margolin K, Ernstoff MS, Hamid O, et al. Ipilimumab in patients with melanoma and brain metastases: an open-label, phase 2 trial. Lancet Oncol. 2012;13(5):459-65.

PMID: 22456429

McQuade, J.L., Posada, L.P., Lecagoonporn, S., Cain, S., Bassett Jr, R.L., Patel,

S.P., Hwu, W.J., Hwu, P., Davies, M.A., Bedikian, A.Y. and Amaria, R.N., 2016. A phase I study of TPI 287 in combination with temozolomide for patients with metastatic melanoma. Melanoma research, 26(6), p.604.

Morikawa A, Peereboom DM, Thorsheim HR et al. Capecitabine and lapatinib uptake in surgically resected brain metastases from metastatic breast cancer patients: a prospective study. Neuro Oncol. 2015;17(2):289–295.

Moraes, F.Y., Taunk, N.K., Marta, G.N., Suh, J.H. and Yamada, Y., 2016. The rationale for targeted therapies and stereotactic radiosurgery in the treatment of brain metastases. The oncologist, 21(2), pp.244-251.

Nayak, L., Lee, E.Q. and Wen, P.Y., 2012. Epidemiology of brain metastases.

Current oncology reports, 14(1), pp.48-54.

152

Neuwelt, E. A. Mechanisms of disease: the blood-brain barrier. Neurosurgery 54,

131–140 (2004).

Noordijk EM, Vecht CJ, Haaxma-Reiche H, et al. The choice of treatment of single brain metastasis should be based on extracranial tumor activity and age. Int J

Radiat Oncol Biol Phys 1994; 29:711.

Palmieri, D. ed., 2012. Central nervous system metastasis, the biological basis and clinical considerations (Vol. 18). Springer Science & Business Media.

Park, H.S., Wang, E.H., Rutter, C.E., Corso, C.D., Chiang, V.L. and James, B.Y.,

2016. Changing practice patterns of Gamma Knife versus linear accelerator-based stereotactic radiosurgery for brain metastases in the US. Journal of neurosurgery,

124(4), pp.1018-1024.

Parker GM, Dunn IF, Ramkissoon SH et al. Recurrent radiation necrosis in the brain following stereotactic radiosurgery. Pract Radiat Oncol. 2015;5(3):e151–e154.

Patchell RA, Tibbs PA, Regine WF, et al. Postoperative radiotherapy in the treatment of single metastases to the brain: a randomized trial. JAMA 1998; 280:1485.

Patel JK, Didolkar MS, Pickren JW, Moore RH. Metastatic pattern of malignant

153 melanoma. A study of 216 autopsy cases. Am J Surg. 1978;135:807–10.

Patil CG, Pricola K, Sarmiento JM, Garg SK, Bryant A, Black KL. Whole brain radiation therapy (WBRT) alone versus WBRT and radiosurgery for the treatment of brain metastases. Cochrane Database of Systematic Reviews 2017, Issue 9. Art.

No.: CD006121. DOI: 10.1002/14651858.CD006121.pub4.

Peretti-viton P, Taieb D, Viton JM, et al. Contrast-enhanced magnetisation transfer

MRI in metastatic lesions of the brain. Neuroradiology. 1998;40(12):783-7.

Postmus PE, Smit EF. Chemotherapy for brain metastases of lung cancer: a review.

Ann Oncol. 1999;10(7):753-9. PMID: 10470420

Robitaille PM, Abduljalil AM, Kangarlu A, et al. Human magnetic resonance imaging at 8 T. NMR Biomed. 1998;11(6):263-5. PMID: 9802467

Sampson JH, Carter Jr JH, Friedman AH, Seigler HF. Demographics, prognosis, and therapy in 702 patients with brain metastases from malignant melanoma. J

Neurosurg. 1998;88:11–20.

Samala, R., Thorsheim, H.R., Goda, S., Taskar, K., Gril, B., Steeg, P.S. and Smith,

Q.R., 2016. Vinorelbine delivery and efficacy in the MDA-MB-231BR preclinical model of brain metastases of breast cancer. Pharmaceutical research, 33(12), pp.2904-2919.

154

Shaw AT, Yeap BY, Solomon BJ, et al. Effect of crizotinib on overall survival in patients with advanced non-small-cell lung cancer harbouring ALK gene rearrangement: a retrospective analysis. Lancet Oncol. 2011;12(11):1004-12.

Senan S, Brade A, Wang LH, et al. PROCLAIM: randomized phase III trial of pemetrexed-cisplatin or etoposide-cisplatin plus thoracic radiation therapy followed by consolidation chemotherapy in locally advanced nonsquamous non-small-cell lung cancer. J Clin Oncol 2016;34:953-962. PMID: 26811519.

Smith, R.A., Andrews, K.S., Brooks, D., Fedewa, S.A., Manassaram‐Baptiste, D.,

Saslow, D., Brawley, O.W. and Wender, R.C., 2018. Cancer screening in the United

States, 2018: a review of current American Cancer Society guidelines and current issues in cancer screening. CA: a cancer journal for clinicians, 68(4), pp.297-316.

Smith, R.A., Andrews, K.S., Brooks, D., Fedewa, S.A., Manassaram‐Baptiste, D.,

Saslow, D., Brawley, O.W. and Wender, R.C., 2017. Cancer screening in the United

States, 2017: a review of current American Cancer Society guidelines and current issues in cancer screening. CA: a cancer journal for clinicians, 67(2), pp.100-121.

Soffietti, R., Abacioglu, U., Baumert, B., Combs, S.E., Kinhult, S., Kros, J.M.,

Marosi, C., Metellus, P., Radbruch, A., Villa Freixa, S.S. and Brada, M., 2017.

Diagnosis and treatment of brain metastases from solid tumors: guidelines from the

European Association of Neuro-Oncology (EANO). Neuro-oncology, 19(2), pp.162-

155 174.

Stemmler HJ, Schmitt M, Willems A, Bernhard H, Harbeck N, Heinemann V. Ratio of trastuzumab levels in serum and cerebrospinal fluid is altered in HER2-positive breast cancer patients with brain metastases and impairment of blood-brain barrier.

Anticancer Drugs. 2007;18(1):23-8. PMID: 17159499

Stewart DJ. A critique of the role of the blood-brain barrier in the chemotherapy of human brain tumors. J Neurooncol. 1994;20(2):121-39. PMID: 7807190

Tsimberidou AM, Letourneau K, Wen S et al. Phase I clinical trial outcomes in 93 patients with brain metastases: the MD Anderson Cancer Center experience. Clin

Cancer Res. 2011;17(12):4110–4118.

Vecht CJ, Haaxma-Reiche H, Noordijk EM, et al. Treatment of single brain metastasis: radiotherapy alone or combined with neurosurgery? Ann Neurol 1993;

33:583.

Voog, J., Brastianos, P., Loeffler, J.S., Shih, H.A. and Oh, K.S., 2017. Clinically

Targetable Mutation Status Associated with Local Control Following Radiation

Treatment for Brain Metastases. International Journal of Radiation Oncology•

Biology• Physics, 99(2), pp.E114-E115.

Wender, R., Fontham, E.T., Barrera Jr, E., Colditz, G.A., Church, T.R., Ettinger,

156 D.S., Etzioni, R., Flowers, C.R., Scott Gazelle, G., Kelsey, D.K. and LaMonte, S.J.,

2013. American Cancer Society lung cancer screening guidelines. CA: a cancer journal for clinicians, 63(2), pp.106-117.

Yin W, Jiang Y, Shen Z, Shao Z, Lu J. Trastuzumab in the adjuvant treatment of

HER2-positive early breast cancer patients: a meta-analysis of published randomized controlled trials. PLoS ONE. 2011;6(6):e21030.

Zatloukal P, Petruzelka L, Zemanova M, et al. Concurrent versus sequential chemoradiotherapy with cisplatin and vinorelbine in locally advanced non-small cell lung cancer: a randomized study. Lung Cancer 2004;46:87-98. PMID: 15364136.

Zhang RD, Price JE, Fujimaki T, Bucana CD, Fidler IJ. Differential permeability of the blood-brain barrier in experimental brain metastases produced by human neoplasms implanted into nude mice. Am J Pathol. 1992;141(5):1115-24. PMCID:

PMC1886664

Chapter 2

Ascha MS. (2018, August 5). mustafaascha/brain-mets-seer: Reproducibility repository for a study of brain metastases in SEER-Medicare (Version v1.0.2).

Zenodo.

Barnholtz-Sloan JS, Sloan AE, Davis FG, Vigneau FD, Lai P, Sawaya RE. Incidence

157 proportions of brain metastases in patients diagnosed (1973 to 2001) in the metropolitan detroit cancer surveillance system. Journal of clinical oncology.

2004;22(14):2865-2872.

Bedikian AY, Wei C, Detry M, et al. Predictive factors for the development of brain metastasis in advanced unresectable metastatic melanoma. American journal of clinical oncology. 2011;34(6):603-610.

Cagney DN, Martin AM, Catalano PJ, et al. Incidence and prognosis of patients with brain metastases at diagnosis of systemic malignancy: A population-based study.

Neuro-Oncology. 2017:nox077.

Cooper GS, Yuan Z, Stange KC, Amini SB, Dennis LK, Rimm AA. The utility of medicare claims data for measuring cancer stage. Medical care. 1999;37(7):706-

711.

Collaborative Stage Data Collection System User Documentation and Coding

Instructions Version 02.03.02. Chicago, Illinois, United States of America:

Collaborative Stage Work Group of the American Joint Committee on Cancer;

American Joint Committee on Cancer; 2017.

Dawood S, Broglio K, Esteva F, et al. Survival among women with triple receptor- negative breast cancer and brain metastases. Annals of Oncology. 2009;20(4):621-

627.

158

Eichler AF, Lamont EB. Utility of administrative claims data for the study of brain metastases: A validation study. Journal of neuro-oncology. 2009;95(3):427-431.

Fox BD, Cheung VJ, Patel AJ, Suki D, Rao G. Epidemiology of metastatic brain tumors. Neurosurgery Clinics of North America. 2011;22(1):1-6.

Garcia, M., Derocq, D., FREiss, G.I.L.L.E.S. and Rochefort, H., 1992. Activation of estrogen receptor transfected into a receptor-negative breast cancer cell line decreases the metastatic and invasive potential of the cells. Proceedings of the

National Academy of Sciences, 89(23), pp.11538-11542.

Gore, E.M., Bae, K., Wong, S.J., Sun, A., Bonner, J.A., Schild, S.E., Gaspar, L.E.,

Bogart, J.A., Werner-Wasik, M. and Choy, H., 2011. Phase III comparison of prophylactic cranial irradiation versus observation in patients with locally advanced non–small-cell lung cancer: Primary analysis of Radiation Therapy Oncology Group

Study RTOG 0214. Journal of Clinical Oncology, 29(3), p.272.

Gregor, A., Cull, A., Stephens, R.J., Kirkpatrick, J.A., Yarnold, J.R., Girling, D.J.,

Macbeth, F.R., Stout, R., Machin, D. and for Cancer, U.K.C.C., 1997. Prophylactic cranial irradiation is indicated following complete response to induction therapy in small cell lung cancer: results of a multicentre randomised trial. European Journal of

Cancer, 33(11), pp.1752-1758.

159 Halpern MT, Ward EM, Pavluck AL, Schrag NM, Bian J, Chen AY. Association of insurance status and ethnicity with cancer stage at diagnosis for 12 cancer sites: A retrospective analysis. The lancet oncology. 2008;9(3):222-231.

Haiman CA, Stram DO, Wilkens LR, et al. Ethnic and racial differences in the smoking-related risk of lung cancer. New England Journal of Medicine.

2006;354(4):333-342.

Hicks, D.G., Short, S.M., Prescott, N.L., Tarr, S.M., Coleman, K.A., Yoder, B.J.,

Crowe, J.P., Choueiri, T.K., Dawson, A.E., Budd, G.T. and Tubbs, R.R., 2006.

Breast cancers with brain metastases are more likely to be estrogen receptor negative, express the basal cytokeratin CK5/6, and overexpress HER2 or EGFR.

The American journal of surgical pathology, 30(9), pp.1097-1104.

Hochstenbag, M.M.H., Twijnstra, A., Wilmink, J.T., Wouters, E.F.M. and Ten Velde,

G.P.M., 2000. Asymptomatic brain metastases (BM) in small cell lung cancer

(SCLC): MR-imaging is useful at initial diagnosis. Journal of neuro-oncology, 48(3), pp.243-248.

Jacot W, Quantin X, Boher J, et al. Brain metastases at the time of presentation of non-small cell lung cancer: A multi-centric aerio* analysis of prognostic factors.

British journal of cancer. 2001;84(7):903.

Keyfitz N. Sampling variance of standardized mortality rates. Human Biology.

160 1966;38(3):309-317.

Kromer C, Xu J, Ostrom QT, et al. Estimating the annual frequency of synchronous brain metastasis in the united states 2010–2013: A population-based study. Journal of Neuro-Oncology. 2017:1-10.

Little AG, Gay EG, Gaspar LE, Stewart AK. National survey of non-small cell lung cancer in the united states: Epidemiology, pathology and patterns of care. Lung cancer. 2007;57(3):253-260.

Loeffler JS, Wen PY. Epidemiology, Clinical Manifestations, and Diagnosis of Brain

Metastases. (DeAngelis LM, Eichler AF, eds.). Wolters Kluwer Health, UpToDate;

2018.

Martin AM, Cagney DN, Catalano PJ, et al. Brain metastases in newly diagnosed breast cancer: A population-based study. JAMA oncology. 2017.

National Cancer Institute. ICD-O-3 SEER Site/Histology validation list. 2018.

National Cancer Institute. SEER-Medicare: Requirements of investigators following receipt of data. 2018.

National Comprehensive Cancer Network (NCCN). NCCN Framework for Resource

Stratification of NCCN Guidelines. Version 3.2018: Non-Small Cell Lung Cancer,

161 Core. 2018. MS-49.

Nussbaum, E.S., Djalilian, H.R., Cho, K.H. and Hall, W.A., 1996. Brain metastases: histology, multiplicity, surgery, and survival. Cancer: Interdisciplinary International

Journal of the American Cancer Society, 78(8), pp.1781-1788.

Pàez-Ribes M, Allen E, Hudock J, et al. Antiangiogenic therapy elicits malignant progression of tumors to increased local invasion and distant metastasis. Cancer cell. 2009;15(3):220-231.

R Core Team. R: A Language and Environment for Statistical Computing. Vienna,

Austria: R Foundation for Statistical Computing; 2017.

Shao Y-Y, Lu L-C, Cheng A-L, Hsu C-H. Increasing incidence of brain metastasis in patients with advanced hepatocellular carcinoma in the era of antiangiogenic targeted therapy. The oncologist. 2011;16(1):82-86.

Sherman, R.E., Anderson, S.A., Dal Pan, G.J., Gray, G.W., Gross, T., Hunter, N.L.,

LaVange, L., Marinac-Dabic, D., Marks, P.W., Robb, M.A. and Shuren, J., 2016.

Real-world evidence—what is it and what can it tell us. N Engl J Med, 375(23), pp.2293-2297.

Slotman, B., Faivre-Finn, C., Kramer, G., Rankin, E., Snee, M., Hatton, M.,

Postmus, P., Collette, L., Musat, E. and Senan, S., 2007. Prophylactic cranial

162 irradiation in extensive small-cell lung cancer. New England Journal of Medicine,

357(7), pp.664-672.

Stallman RM, McGrath R, Smith PD. GNU Make: A Program for Directing

Recompilation, for Version 3.81. Free Software Foundation; 2004.

Stodden V, Borwein J, Bailey DH. Setting the default to reproducible. computational science research SIAM News. 2013;46(5):4-6.

Surveillance E-RP Epidemiology. SEER Research Data Record Description: Cases

Diagnosed in 1973-2014. Chicago, Illinois, United States of America: National

Cancer Institute, National Institutes of Health; 2017.

Lin NU, Claus E, Sohl J, Razzak AR, Arnaout A, Winer EP. Sites of distant recurrence and clinical outcomes in patients with metastatic triple-negative breast cancer. Cancer. 2008;113(10):2638-2645.

National Comprehensive Cancer Network (NCCN). NCCN Clinical Practice

Guidelines Version 2.2018: Melanoma. 2018.

Nordstrom BL, Whyte JL, Stolar M, Mercaldi C, Kallich JD. Identification of metastatic cancer in claims data. Pharmacoepidemiology and drug safety.

2012;21(S2):21-28.

163 Ostrom QT, Wright CH, Barnholtz-Sloan JS. Brain metastases: Epidemiology.

Handbook of clinical neurology. 2018;149:27-42.

Takahashi, T., Yamanaka, T., Seto, T., Harada, H., Nokihara, H., Saka, H., Nishio,

M., Kaneda, H., Takayama, K., Ishimoto, O. and Takeda, K., 2017. Prophylactic cranial irradiation versus observation in patients with extensive-disease small-cell lung cancer: a multicentre, randomised, open-label, phase 3 trial. The Lancet

Oncology, 18(5), pp.663-671.

Videtic GM, Reddy CA, Chao ST, et al. Gender, race, and survival: A study in non–

Small-cell lung cancer brain metastases patients utilizing the radiation therapy oncology group recursive partitioning analysis classification. International Journal of

Radiation Oncology• Biology• Physics. 2009;75(4):1141-1147.

Warren JL, Klabunde CN, Schrag D, Bach PB, Riley GF. Overview of the seer- medicare data: Content, research applications, and generalizability to the united states elderly population. Medical care. 2002;40(8):IV-3.

Werner CA. The older population: 2010. 2010 census briefs. document c2010br-09.

2011.

Whyte JL, Engel-Nitz NM, Teitelbaum A, Rey GG, Kallich JD. An evaluation of algorithms for identifying metastatic breast, lung, or colorectal cancer in administrative claims data. Medical care. 2015;53(7):e49-e57.

164

Chapter 3

Alboukadel Kassambara and Marcin Kosinski (2018). survminer: Drawing Survival

Curves using 'ggplot2'. R package version 0.4.3.999. http://www.sthda.com/english/rpkgs/survminer/

Amaravadi R, Schuchter LM, McDermott DF, et al. Updated results of a randomized phase II study comparing two schedules of temozolamide in combination with sorafenib in patients with advanced melanoma [abstract] J Clin Oncol.

2007;25(suppl 18):8527.

Ascha, MS. 0.0.9000 mustafaascha/seerm-bevacizumab: Repository init. January

19, 2019. DOI: 10.5281/zenodo.2544421

Besse B, Lasserre SF, Compton P, Huang J, Augustus S, Rohr UP. Bevacizumab safety in patients with central nervous system metastases. Clin Cancer Res.

2010;16(1):269-78. PMID: 20028762

Carden CP, Larkin JM, Rosenthal MA. What is the risk of intracranial bleeding during anti-VEGF therapy?. Neuro-oncology. 2008;10(4):624-30. PMID: 18539884

Cedrych I, Kruczała MA, Walasek T, Jakubowicz J, Blecharz P, Reinfuss M.

Systemic treatment of non-small cell lung cancer brain metastases. Contemp Oncol

165 (Pozn). 2016;20(5):352-357. PMCID: PMC5371701/

ClinicalTrials.gov [Internet]. Bethesda (MD): National Library of Medicine (US) July

12, 2013 - . Identifier NCT01898130, Bevacizumab in Pats w/Recurrent ST Brain

Metas Who Have Failed Whole Brain Radiation Therapy; June 14, 2018. Accessed

January 2019.

Cloughesy TF, Prados MD, Wen PY, Mikkelsen T, Abrey LE, Schiff D, Yung WK,

Maoxia Z, Dimery I, Friedman HS (2008) A phase II, randomized, non-comparative clinical trial of the effect of bevacizumab (BV) alone or in combination with irinotecan

(CPT) on 6-month progression free survival (PFS6) in recurrent, treatment-refractory glioblastoma (GBM). J Clin Oncol 26(Suppl):Abstract 2010b, oral presentation update. DOI: 10.1200/jco.2008.26.15_suppl.2010b

Caffo M, Barresi V, Caruso G, et al. Innovative therapeutic strategies in the treatment of brain metastases. Int J Mol Sci. 2013;14(1):2135-74. PMCID:

PMC3565370/

Diaz RJ, Ali S, Qadir MG, De la fuente MI, Ivan ME, Komotar RJ. The role of bevacizumab in the treatment of glioblastoma. J Neurooncol. 2017;133(3):455-467.

Dietrich J, Rao K, Pastorino S, Kesari S. Corticosteroids in brain cancer patients: benefits and pitfalls. Expert Rev Clin Pharmacol. 2011;4(2):233-42. PMCID:

PMC3109638/

166

De Braganca KC, Janjigian YY, Azzoli CG, et al. Efficacy and safety of bevacizumab in active brain metastases from non-small cell lung cancer. J Neurooncol.

2010;100(3):443-7. PMID: 20440540

Dunkler D, Ploner M, Schemper M and Heinze G (2018). “Weighted Cox Regression

Using the R Package coxphw.” _Journal of Statistical Software_, *84*(2), pp. 1-26. doi: 10.18637/jss.v084.i02 (URL: http://doi.org/10.18637/jss.v084.i02).

Ellis, L.M., 2006, October. Mechanisms of action of bevacizumab as a component of therapy for metastatic colorectal cancer. Seminars in oncology (Vol. 33, pp. S1-S7).

WB Saunders.

FDA News Release. FDA approves first biosimilar for the treatment of cancer.

September 14, 2017. https://www.fda.gov/NewsEvents/Newsroom/PressAnnouncements/ucm576112.htm

Ferrara N, Hillan KJ, Gerber HP, Novotny W. Discovery and development of bevacizumab, an anti-VEGF antibody for treating cancer. Nat Rev Drug Discov.

2004;3(5):391-400.

Friedman J, Hastie T, Tibshirani R (2010). Regularization Paths for Generalized

Linear Models via Coordinate Descent. Journal of Statistical Software, 33(1), 1-22.

URL http://www.jstatsoft.org/v33/i01/.

167

Geynisman DM, Chien CR, Smieliauskas F, Shen C, Shih YC. Economic evaluation of therapeutic cancer vaccines and immunotherapy: a systematic review. Hum

Vaccin Immunother. 2014;10(11):3415-24. PMCID: PMC4514043/

Gilbert MR, Dignam JJ, Armstrong TS, Wefel JS, Blumenthal DT, Vogelbaum MA,

Colman H, Chakravarti A, Pugh S, Won M, Jeraj R, Brown PD, Jaeckle KA, Schiff D,

Stieber VW, Brachman DG, Werner-Wasik M, Tremont-Lukats IW, Sulman EP,

Aldape KD, Curran WJ Jr, Mehta MP. A randomized trial of bevacizumab for newly diagnosed glioblastoma. N Engl J Med. 2014 Feb 20;370(8):699-708. doi:10.1056/NEJMoa1308573. PubMed PMID: 24552317; PubMed Central PMCID:

PMC4201043.

Gil-gil MJ, Mesia C, Rey M, Bruna J. Bevacizumab for the treatment of glioblastoma.

Clin Med Insights Oncol. 2013;7:123-35. PMCID: PMC3682734

Giovanni Bernardo, Quinto Cuzzoni, Maria Rosa Strada, Antonio Bernardo,

Giuseppe Brunetti, Ivona Jedrychowska, Umberto Pozzi & Raffaella Palumbo (2002)

First-Line Chemotherapy with Vinorelbine, Gemcitabine, and Carboplatin in the

Treatment of Brain Metastases from Non-Small-Cell Lung Cancer: A Phase II Study,

Cancer Investigation, 20:3, 293-302, DOI: 10.1081/CNV-120001173

Heiss, J.D., Papavassiliou, E., Merrill, M.J., Nieman, L., Knightly, J.J., Walbridge, S.,

Edwards, N.A. and Oldfield, E.H., 1996. Mechanism of dexamethasone suppression

168 of brain tumor-associated vascular permeability in rats. Involvement of the glucocorticoid receptor and vascular permeability factor. The Journal of clinical investigation, 98(6), pp.1400-1408.

Hollingsworth HC, Kohn EC, Steinberg SM, Rothenberg ML, Merino MJ. Tumor angiogenesis in advanced stage ovarian carcinoma. Am J Pathol. 1995;147(1):33-

41.

Heymach JV, Johnson BE, Prager D, et al. A phase II trial of ZD6474 plus docetaxel in patients with previously treated NSCLC: follow up results [abstract] J Clin Oncol.

2006;24(suppl 18):7016.

Henderson CA, Bukowski RM, Stadler WM, et al. The Advanced Renal Cell

Carcinoma Sorafenib (ARCCS) expanded access trial: subset analysis of patients with brain metastases [abstract] J Clin Oncol. 2007;25(suppl 18):15506.

Kim, Michelle M et al. “Gemcitabine Plus Radiation Therapy for High-Grade Glioma:

Long-Term Results of a Phase 1 Dose-Escalation Study” International journal of radiation oncology, biology, physics vol. 94,2 (2015): 305-11.

Kural C, Atac GK, Tehli O, et al. The evaluation of the effects of steroid treatment on the tumor and peritumoral edema by DWI and MR spectroscopy in brain tumors.

Neurol Neurochir Pol. 2018;52(4):495-504. PMID: 29588064

169 Kurzrock R, Stewart DJ. Exploring the Benefit/Risk Associated with Antiangiogenic

Agents for the Treatment of Non-Small Cell Lung Cancer Patients. Clin Cancer Res.

2017;23(5):1137-1148.

Li, Y., Appius, A., Pattipaka, T., Feyereislova, A., Cassidy, A. and Ganti, A.K., 2019.

Real-world management of patients with epidermal growth factor receptor (EGFR) mutation-positive non–small-cell lung cancer in the USA. PloS one, 14(1), p.e0209709.

Lois A. Lampson (2011) Monoclonal antibodies in neuro-oncology, mAbs, 3:2, 153-

160. DOI: 10.4161/mabs.3.2.14239

Matsuoka, H., Yoshiuchi, K., Koyama, A., Otsuka, M., & Nakagawa, K. (2015).

Chemotherapeutic drugs that penetrate the blood–brain barrier affect the development of hyperactive delirium in cancer patients. Palliative and Supportive

Care, 13(4), 859-864. doi:10.1017/S1478951514000765

Melosky B, Reardon DA, Nixon AB, Subramanian J, Bair AH, Jacobs I.

Bevacizumab biosimilars: scientific justification for extrapolation of indications.

Future Oncol. 2018;14(24):2507-2520.

National Cancer Institute Cancer Research Network. Cancer Therapy Lookup

Tables: Chemotherapy, hormone therapy, and immunotherapy. Cancer Research

Network (U24 CA171524). Accessed November 2018.

170 https://crn.cancer.gov/resources/codes.html

National Cancer Institute Surveillance, Epidemiology, and End-Results Program.

SEER Data and Software, Documentation for Data, Site Recode ICD-O-3/WHO

2008 Definition. https://seer.cancer.gov/siterecode/icdo3_dwhoheme/index.html

National Cancer Institute Surveillance, Epidemiology, and End-Results Program.

Lung Solid Tumor Rules, National Lung Equivalent Terms and Definitions, C340-

C343, C348, C349. Updated 10/12/2018. https://seer.cancer.gov/tools/solidtumor/lung_STM.pdf

Seoane J, De Mattos-Arruda L. Brain metastasis: new opportunities to tackle therapeutic resistance. Mol Onc. 2014;8;1120-1131. https://febs.onlinelibrary.wiley.com/doi/full/10.1016/j.molonc.2014.05.009/

Simon N, Friedman J, Hastie T, Tibshirani R (2011). Regularization Paths for Cox's

Proportional Hazards Model via Coordinate Descent. Journal of Statistical Software,

39(5), 1-13. URL http://www.jstatsoft.org/v39/i05/.

Socinski MA, Langer CJ, Huang JE, et al. Safety of bevacizumab in patients with non-small-cell lung cancer and brain metastases. J Clin Oncol. 2009;27(31):5255-

61. PMID: 19738122

Tamaskar I, Shaheen P, Wood L, et al. Antitumour effects of sorafenib and sunitinib

171 in patients with metastatic renal cell carcinoma (mRCC) who had prior therapy antiangiogenic agents [abtract] J Clin Oncol. 2006;24(suppl 18):4597.

Therneau T, Crowson C, Atkinson E. 'Adjusted Survival Curves'. (2015)

‘https://cran.r-project.org/web/packages/survival/vignettes/adjcurve.pdf’

Yang L, Chen CJ, Guo XL, Wu XC, Lv BJ, Wang HL, Guo Z, Zhao, XY.

Bevacizumab and risk of intracranial hemorrhage in patients with brain metastases: a meta-analysis. J Neurooncol. 2018; 137(1): 49-56. PMCID: PMC5846997/

Chapter 4

Agbai, O. N., Buster, K., Sanchez, M., Hernandez, C., Kundu, R. V., Chiu, M., ... &

Lim, H. W. (2014). Skin cancer and photoprotection in people of color: a review and recommendations for physicians and the public. Journal of the American Academy of Dermatology, 70(4), 748-762.

Agbai, O. N., Buster, K., Sanchez, M., Hernandez, C., Kundu, R. V., Chiu, M., ... &

Lim, H. W. (2014). Skin cancer and photoprotection in people of color: a review and recommendations for physicians and the public. Journal of the American Academy of Dermatology, 70(4), 748-762.

Akinyemiju, T., Sakhuja, S., Waterbor, J., Pisu, M., & Altekruse, S. F. (2018).

Racial/ethnic disparities in de novo metastases sites and survival outcomes for

172 patients with primary breast, colorectal, and prostate cancer. Cancer Medicine, 7(4),

1183-1193.

Alam, M. S., & Mukherjee, B. (2017). Squamous cell carcinoma. In Emergencies of the Orbit and Adnexa (pp. 415-422). Springer, New Delhi.

American Cancer Society. Cancer Facts & Figures 2018. Atlanta: American Cancer

Society; 2018.

American Cancer Society. Cancer Facts & Figures for African Americans 2016-

2018.

American Cancer Society. Cancer Facts & Figures for Hispanics/Latinos 2015-2017.

Atlanta: American Cancer Society; 2015.

American Cancer Society. Special Section: Cancer in Asian Americans, Native

Hawaiians, and Pacific Islanders. Atlanta: American Cancer Society, 2016.

Blake, K. D., Moss, J. L., Gaysynsky, A., Srinivasan, S., & Croyle, R. T. (2017).

Making the case for investment in rural cancer control: an analysis of rural cancer incidence, mortality, and funding trends. Cancer Epidemiology and Prevention

Biomarkers.

Blake, K. D., Moss, J. L., Gaysynsky, A., Srinivasan, S., & Croyle, R. T. (2017).

173 Making the case for investment in rural cancer control: An analysis of rural cancer incidence, mortality, and funding trends. Cancer Epidemiology and Prevention

Biomarkers, 26(7), 992-997.

Bodekær, M., Petersen, B., Philipsen, P. A., Heydenreich, J., Thieden, E., & Wulf,

H. C. (2015). Sun exposure patterns of urban, suburban, and rural children: a dosimetry and diary study of 150 children. Photochemical & Photobiological

Sciences, 14(7), 1282-1289.

Christ, G., Messner, C., & Behar, L. (Eds.). (2015). Handbook of oncology social work: Psychosocial care for people with cancer. Oxford University Press.

Conic, R. Z., Cabrera, C. I., Khorana, A. A., & Gastman, B. R. (2018). Determination of the impact of melanoma surgical timing on survival using the National Cancer

Database. Journal of the American Academy of Dermatology, 78(1), 40-46.

Cronin, K. A., Lake, A. J., Scott, S., Sherman, R. L., Noone, A. M., Howlader, N., ...

& Kohler, B. A. (2018). Annual Report to the Nation on the Status of Cancer, part I:

National cancer statistics. Cancer, 124(13), 2785-2800.

DeSantis, C. E., Siegel, R. L., Sauer, A. G., Miller, K. D., Fedewa, S. A., Alcaraz, K.

I., & Jemal, A. (2016). Cancer statistics for African Americans, 2016: progress and opportunities in reducing racial disparities. CA: a cancer journal for clinicians, 66(4),

290-308.

174

Dwyer-Lindgren, L., Mokdad, A. H., Srebotnjak, T., Flaxman, A. D., Hansen, G. M.,

& Murray, C. J. (2014). Cigarette smoking prevalence in US counties: 1996-2012.

Population health metrics, 12(1), 5.

Goldenberg, A., Vujic, I., Sanlorenzo, M., & Ortiz-Urda, S. (2015). Melanoma risk perception and prevention behavior among African-Americans: the minority melanoma paradox. Clinical, cosmetic and investigational dermatology, 8, 423.

Guadagnolo, B. A., Petereit, D. G., & Coleman, C. N. (2017, April). Cancer care access and outcomes for American Indian populations in the United States: challenges and models for progress. In Seminars in radiation oncology (Vol. 27, No.

2, pp. 143-149). Elsevier.

Harvey, V. M., Enos, C. W., Chen, J. T., Galadima, H., & Eschbach, K. (2017). The

Role of Neighborhood Characteristics in Late Stage Melanoma Diagnosis among

Hispanic Men in California, Texas, and Florida, 1996–2012. Journal of cancer epidemiology, 2017.

Lawler, S., Spathonis, K., Masters, J., Adams, J., & Eakin, E. (2011). Follow-up care after breast cancer treatment: experiences and perceptions of service provision and provider interactions in rural Australian women. Supportive Care in Cancer, 19(12),

1975-1982.

175 McClelland III, S., Page, B. R., Jaboin, J. J., Chapman, C. H., Deville Jr, C., &

Thomas Jr, C. R. (2017). The pervasive crisis of diminishing radiation therapy access for vulnerable populations in the United States, part 1: African-American patients. Advances in Radiation Oncology, 2(4), 523-531.

Moore, J., Zelen, D., Hafeez, I., Ganti, A. K., Beal, J., & Potti, A. (2003). Risk- awareness of cutaneous malignancies among rural populations. Medical Oncology,

20(4), 369-373.

Morris, A. M., Rhoads, K. F., Stain, S. C., & Birkmeyer, J. D. (2010). Understanding racial disparities in cancer treatment and outcomes. Journal of the American

College of Surgeons, 211(1), 105-113

Muller, C. J., Robinson, R. F., Smith, J. J., Jernigan, M. A., Hiratsuka, V., Dillard, D.

A., & Buchwald, D. (2017). Text message reminders increased colorectal cancer screening in a randomized trial with Alaska Native and American Indian people.

Cancer, 123(8), 1382-1389.

Plescia, M., Henley, S. J., Pate, A., Underwood, J. M., & Rhodes, K. (2014). Lung cancer deaths among American Indians and Alaska Natives, 1990–2009. American journal of public health, 104(S3), S388-S395.

Skin Cancer Facts & Statistics. (3 May 2018) In Skin Cancer Foundation. Retrieved

24 June 2018, from https://www.skincancer.org/skin-cancer-information/skin-cancer-

176 facts

Slominski, R. M., Zmijewski, M. A., & Slominski, A. T. (2015). The role of melanin pigment in melanoma. Experimental dermatology, 24(4), 258-259.

Stubblefield, J., & Kelly, B. (2014). Melanoma in non-caucasian populations.

Surgical Clinics, 94(5), 1115-1126.

Thompson, C. A., Gomez, S. L., Hastings, K. G., Kapphahn, K., Yu, P., Shariff-

Marco, S., ... & Palaniappan, L. P. (2016). The burden of cancer in Asian

Americans: a report of national mortality trends by Asian ethnicity. Cancer

Epidemiology and Prevention Biomarkers.

Torre, L. A., Sauer, A. M. G., Chen, M. S., Kagawa‐Singer, M., Jemal, A., & Siegel,

R. L. (2016). Cancer statistics for Asian Americans, Native Hawaiians, and Pacific

Islanders, 2016: Converging incidence in males and females. CA: a cancer journal for clinicians, 66(3), 182-202.

Trifiletti, D. M., Sheehan, J. P., Grover, S., Dutta, S. W., Rusthoven, C. G.,

Kavanagh, B. D., ... & Showalter, T. N. (2017). National trends in radiotherapy for brain metastases at time of diagnosis of non-small cell lung cancer. Journal of

Clinical Neuroscience, 45, 48-53.

Varlotto, J. M., Voland, R., McKie, K., Flickinger, J. C., DeCamp, M. M., Maddox,

177 D., ... & Oliveira, P. (2018). Population‐based differences in the outcome and presentation of lung cancer patients based upon racial, histologic, and economic factors in all lung patients and those with metastatic disease. Cancer medicine, 7(4),

1211-1220.

Wang, Y., Zhao, Y., & Ma, S. (2016). Racial differences in six major subtypes of melanoma: descriptive epidemiology. BMC cancer, 16(1), 691.

White, M. C., Espey, D. K., Swan, J., Wiggins, C. L., Eheman, C., & Kaur, J. S.

(2014). Disparities in cancer mortality and incidence among American Indians and

Alaska Natives in the United States. American journal of public health, 104(S3),

S377-S387.

Zahnd, W. E., Fogleman, A. J., & Jenkins, W. D. (2018). Rural–Urban Disparities in

Stage of Diagnosis Among Cancers With Preventive Opportunities. American journal of preventive medicine, 54(5), 688-698.

Zhang, H., Dziegielewski, P. T., Nguyen, T. J., Jeffery, C. C., O’Connell, D. A.,

Harris, J. R., & Seikaly, H. (2015). The effects of geography on survival in patients with oral cavity squamous cell carcinoma. Oral oncology, 51(6), 578-585.

Chapter 5

Bachelot, T., Romieu, G., Campone, M., Diéras, V., Cropet, C., Dalenc, F., Jimenez,

178 M., Le Rhun, E., Pierga, J.Y., Gonçalves, A. and Leheurteur, M., 2013. Lapatinib plus capecitabine in patients with previously untreated brain metastases from

HER2-positive metastatic breast cancer (LANDSCAPE): a single-group phase 2 study. The lancet oncology, 14(1), pp.64-71.

Banda JM, Halpern Y, Sontag D, Shah NH. Electronic phenotyping with

APHRODITE and the Observational Health Sciences and Informatics (OHDSI) data network. AMIA Jt Summits Transl Sci Proc. 2017;2017:48-57. PMCID:

PMC5543379

Breast cancer drug approved for new indication. Womens Health (Lond).

2010;6(2):173. PMID: 20187722

Dismuke CE. Underreporting of computed tomography and magnetic resonance imaging procedures in inpatient claims data. Med Care. 2005;43(7):713-7. PMID:

15970787

Etzioni DA, Lessow C, Bordeianou LG, et al. Concordance Between Registry and

Administrative Data in the Determination of Comorbidity: A Multi-institutional Study.

Ann Surg. 2019; PMID: 30817356Haut ER, Pronovost PJ. Surveillance bias in outcomes reporting. JAMA. 2011;305(23):2462-3. PMID: 21673300

Etzioni DA, Lessow C, Bordeianou LG, et al. Concordance Between Registry and

Administrative Data in the Determination of Comorbidity: A Multi-institutional Study.

179 Ann Surg. 2019; PMID: 30817356

Haut ER, Pronovost PJ, Schneider EB. Limitations of Administrative Databases.

JAMA. 2012;307(24):2589–2590. doi:10.1001/jama.2012.6626

Higa GM, Abraham J. Lapatinib in the treatment of breast cancer. Expert Rev

Anticancer Ther. 2007;7(9):1183-92. PMID: 17892419 cancer with brain metastases: A systematic review and pooled analysis. European Journal of Cancer,

84, pp.141-148. PMID: 17892419

McCoy Jr, T.H., Yu, S., Hart, K.L., Castro, V.M., Brown, H.E., Rosenquist, J.N.,

Doyle, A.E., Vuijk, P.J., Cai, T. and Perlis, R.H., 2018. High throughput phenotyping for dimensional psychopathology in electronic health records. Biological psychiatry,

83(12), pp.997-1004.

Nielsen, S.S., Warden, M.N., Camacho-Soto, A., Willis, A.W., Wright, B.A. and

Racette, B.A., 2017. A predictive model to identify Parkinson disease from administrative claims data. Neurology, 89(14), pp.1448-1456.

Petrelli, F., Ghidini, M., Lonati, V., Tomasello, G., Borgonovo, K., Ghilardi, M.,

Cabiddu, M. and Barni, S., 2017. The efficacy of lapatinib and capecitabine in HER-

2 positive breast

Romano PS, Mark DH. Bias in the coding of hospital discharge data and its

180 implications for quality assessment. Med Care. 1994;32(1):81-90. PMID: 8277803

Sarrazin MS, Rosenthal GE. Finding pure and simple truths with administrative data.

JAMA. 2012;307(13):1433-5. PMID: 22474208

Schenck AP, Klabunde CN, Warren JL, et al. Data sources for measuring colorectal endoscopy use among Medicare enrollees. Cancer Epidemiol Biomarkers Prev.

2007;16(10):2118-27. PMID: 17932360

US Department of Health and Human Services, National Institutes of Health,

National Cancer Institute, Surveillance, Epidemiology and End-Results Program.

“Measures that are limited or not available in the data”. Last updated 14 Sep 2018, accessed 4 March 2019. https://healthcaredelivery.cancer.gov/seermedicare/considerations/measures.html

William J. Culpepper, Ruth Ann Marrie, Annette Langer-Gould, Mitchell T. Wallin,

Jonathan D. Campbell, Lorene M. Nelson, Wendy E. Kaye, Laurie Wagner, Helen

Tremlett, Lie H. Chen, Stella Leung, Charity Evans, Shenzhen Yao, Nicholas G.

LaRocca, on behalf of the United States Multiple Sclerosis Prevalence Workgroup

(MSPWG) Validation of an algorithm for identifying MS cases in administrative health claims datasets. Neurology Feb 2019, 10.1212/WNL.0000000000007043;

DOI: 10.1212/WNL.0000000000007043

Zhang, J., Zhang, L., Yan, Y., Li, S., Xie, L., Zhong, W., Lv, J., Zhang, X., Bai, Y.

181 and Cheng, Z., 2015. Are capecitabine and the active metabolite 5-FU CNS penetrable to treat breast cancer brain metastasis?. Drug Metabolism and

Disposition, 43(3), pp.411-417.

182