Integrating Predictive Analytics Into High-Value Carethe Dawn Of

Opinion

INNOVATIONS IN HEALTH CARE DELIVERY VIEWPOINT Integrating Predictive Analytics Into High-Value Care The Dawn of Precision Delivery

Ravi B. Parikh, MD, United States health care costs are twice as high as els around clinical issues, such as acute intensive care unit 1 MPP spending in most industrialized countries. One key op- decompensation and hospital readmissions. Department of portunity for health systems to improve value is by lim- As organizations like Amazon and American Air- Medicine, Brigham iting overuse of costly resources, in part by focusing lines have shown, however, development of these mod- and Women’s 1 Hospital, Boston, these resources toward high-risk patient groups. Some els is only a first step. Few health systems currently use Massachusetts; and healthsystemshavebeenusingretrospectiveclaimsdata predictive analytics at scale to influence health care de- Harvard Medical or other approaches, like the Framingham risk model, to livery.Health systems must identify strategies to imple- School, Boston, Massachusetts. identify high-risk individuals. However, most systems to- ment predictive risk algorithms into clinical practice. day are doing little in the way of risk stratification, and Meetali Kakad, MD, physicians often find it difficult to apply these charac- Using Predictive Analytics to Focus Intensity MPH terizations of risk to the care of an individual patient. of Services Across the Care Continuum Department of Electronichealthrecords(EHRs)haveheldtheprom- Acute Care Medicine, Brigham and Women’s ise of allowing clinicians and health systems to deter- Antibioticoverusepredisposespatientstoadverseevents Hospital, Boston, mine an individual’s real-time risk of a clinical event and resistant infections, the treatment of which results in Massachusetts; and through predictive analytics. The use of EHRs is becom- significant health care costs in the United States. Kaiser Harvard Medical ing ubiquitous in the United States. This sea change can Permanente of Northern California (KPNC), an integrated School, Boston, Massachusetts. be linked with advances in big data techniques and com- health service organization, has used predictive analytics puterized decision support to transform health care de- to reduce antibiotic overuse in neonates. KPNC used ma- David W. Bates, MD, livery. Just as “precision medicine” is generally linked to ternal health data from more than 600 000 live births to MSc the concept of using genetic and genomic data to per- determine the probability of early-onset neonatal sepsis Department of sonalizetreatments,“precisiondelivery”involvesusingan Medicine, Brigham innonprematureinfantspriortobirth.Thesedatawerein- and Women’s individual’s electronic health data to predict risk and per- tegrated with objective clinical data from the newborn at Hospital, Boston, sonalize care to substantially improve value. In this View- birth to assess the probability of sepsis by categorizing Massachusetts; and point, we make the case for precision delivery by describ- newborns as at low, medium, or high risk of sepsis. KPNC Harvard Medical School, Boston, inghowsomehealthsystemsarebeginningtosuccessfully obstetricians and neonatologists then used this score to Massachusetts. implement analytics into practice and discussing future determinewhethertoadministerantibiotics.3 Afterimple- directions for using predictive analytics to improve value. mentation of this algorithm, use of systemic antibiotics in theneonatalperiodamongnewbornsof34weeksormore Lessons From Other Fields gestationwasestimatedtodecreaseby33%to60%,and Other industries have successfully used predictive ana- up to an estimated 250 000 newborns nationally could lytics to tailor service delivery in real time. Familiar ex- potentially be spared antibiotics at birth annually.3 amples include Amazon’s product recommendation system for online shopping based on an individual’s prior Postdischarge Care purchases, and American Airlines’ ticket pricing system Hospital readmissions represent an important driver of basedonpriorcustomerpurchasingtrends.Sportsteams spending, with all-cause 30-day readmissions costing like the Oakland Athletics have relied heavily on analyt- the US health system more than $41 billion annually,and ics to select player rosters, outperforming expecta- thus are a major quality indicator for health systems.4 tions despite having a much smaller payroll than other Parkland Health and Hospital System used an algo- teams. These organizations use large amounts of data rithm based on 29 clinical, social, behavioral, and utili- and sophisticated machine learning algorithms to meet zation factors available within 24 hours of admission consumer and organizational needs. to predict risk of readmission for patients with heart In health care, predictive analytics offers an auto- failure.5 In a prospective study, 228 patients with heart mated means to forecast future health outcomes for in- failure deemed at high risk of 30-day readmission re- Corresponding dividualsorpopulationsbasedonalgorithmsderivedfrom ceived targeted evidence-based interventions includ- Author: David W. historical patient data. Some smartphone apps have suc- ing (1) detailed patient education by a multidisciplinary Bates, MD, MSc, cessfully applied predictive analytics to influence health team including a pharmacist, nutritionist, and case Brigham and manager; (2) follow-up telephone calls within 48 hours Women’s Hospital, care: Ginger.io, for example, uses analytics based on cell 1620 Tremont St, phone data to identify patients at risk for depression cri- to ensure medication adherence; (3) outpatient heart Boston, MA 02120 ses, cueing physicians and caregivers to intervene.2 As failure specialist appointments within 7 days; and (dbates@partners more electronic health data become available, some (4) a primary care appointment scheduled according .org). health systems have begun to develop predictive mod- to the urgency of noncardiac issues. Compared with

jama.com (Reprinted) JAMA February 16, 2016 Volume 315, Number 7 651

Opinion Viewpoint

834 patients enrolled in the study prior to intervention, there was Organizationsthataimforprecisiondeliveryofcarewillneedsev- a 26% relative reduction in risk-adjusted odds of readmission among eralkeypiecesofinfrastructure.First,theywillneedanintegratedEHR 913 patients with heart failure enrolled in the postintervention pe- infrastructure and access to long-term data, like the VHA CDW, on riod (26% vs 21% 30-day readmission rates).5 which to base predictive algorithms. Second, they will need robust, responsivetoolstoaddresssuggestionsandimproveclinicians’work- Serious Illness flowwithinclinicalsystems.Third,outputsofthealgorithmswillneed The Veterans Health Administration (VHA) applied analytics to im- to be actionable and prompt prespecified, evidence-based activi- prove quality of care for serious illness by creating its Corporate Data ties, similar to the KPNC antibiotic guidelines. Fourth, predictive al- Warehouse (CDW), a repository for patient-level data aggregated gorithms will need to be flexible enough to quickly adjust for real- from across the VHA, in 2006.6 The CDW was used to calculate risk time patient data, as the Parkland readmission model does. Such scores predicting hospitalization and death for VHA’s primary care iteration allows for flexible “dosing” of services across the continuum population, based on variables including demographics, vital signs, of care, with intensity geared up and down as the need requires. laboratory results, and prior utilization. Accessed 3000 to 4000 Some health professionals have raised concerns about the ap- times monthly by more than 1200 clinicians, these scores are widely plication of predictive analytics, not the least of which is the per- used in practice. Nurse care managers used these scores to guide ceived diminution of the role of the physician in managing clinical services, including end-of-life and palliative care, delivered by mul- uncertainty.8 Other concerns include protection of patient privacy, tidisciplinary patient-aligned care teams (PACTs) to high-risk indi- diminishment of patient preferences, and inadequate medical viduals. Compared with 87 practices with the lowest implementa- training.9 Health professionals had similar hesitations more than a tion of PACTs, the 77 practices with highest PACT implementation decade ago when considering implementing EHRs. However, algo- demonstrated a 17% reduction in hospitalizations (4.42 vs 3.68 quar- rithms routinely outperform practitioners’ clinical intuition without terly admissions per 1000 veterans) for ambulatory care–sensitive decision support. Algorithms also may enhance the quality of inter- conditions and a 27% reduction in emergency department visits actionbetweenphysiciansandpatients—forexample,machinelearn- (188 vs 245 visits per 1000 patients) over a 7-month period.7 ing algorithms based on retrospective data can provide survival pro- jections that may help inform discussions regarding end-of-life care Future Directions for patients with advanced cancer. However, physicians will still need These organizations are examples of health systems that apply pre- to exercise clinical judgment, and with appropriate training can com- dictive analytics to improve value for high-risk patient groups. Un- bine new insights learned from predictive analytics alongside pa- der accountable care, successful organizations will use a broad ar- tient preferences to make higher-value treatment decisions. ray of tools to predict important outcomes, including to identify The time for precision delivery is now. With the advent of ac- patients likely to require expensive care, be readmitted, or experi- countable care, the health care organizations that succeed will be ence a specific type of adverse event.1 However, just as important those that deliver high value. Perhaps the most important step to as prediction is how the predictions are integrated with clinical sys- improvingvaluewillbeimplementingclinicalanalyticsinroutinecare. tems to help physicians and other health care professionals make Organizations that adapt by integrating these tools may do better decisions and track real-time quality. both clinically and financially going forward.

ARTICLE INFORMATION Funding/Support: Dr Kakad is funded by The newborns Ն 34 weeks’ gestation. Pediatrics. 2014; Conflict of Interest Disclosures: All authors have Commonwealth Fund and the Research Council of 133(1):30-36. completed and submitted the ICMJE Form for Norway under the Norwegian Harkness Fellowship 4. Hines AL, Barrett ML, Jiang J, Steiner CA. Disclosure of Potential Conflicts of Interest. in Health Care Policy and Practice. Conditions With the Largest Number of Adult Dr Bates reported receiving equity from Intensix, Role of the Funder/Sponsor: The funders had no Hospital Admissions by Payer, 2011. https://www which makes software to support clinical role in neither design and conduct of the study; .hcup-us.ahrq.gov/reports/statbriefs/sb172 decision-making in intensive care; being named as collection, management, analysis, and -Conditions-Readmissions-Payer.jsp. April 2014. coinventor on patent No. 6029138 held by Brigham interpretation of the data; preparation, review, or Accessed December 22, 2015. and Women’s Hospital on the use of decision approval of the manuscript; nor decision to submit 5. Amarasingham R, Patel PC, Toto K, et al. support software for medical management, the manuscript for publication. Allocating scarce resources in real-time to reduce licensed to the Medicalis Corporation, and holding a Disclaimer: The authors have no commercial or heart failure readmissions: a prospective, controlled minority equity position in Medicalis, which financial ties to any of the organizations or study. BMJ Qual Saf. 2013;22(12):998-1005. develops web-based decision support for radiology companies mentioned in this manuscript. test ordering; serving on the clinical advisory board 6. Fihn SD, Francis J, Clancy C, et al. Insights from advanced analytics at the Veterans Health for Zynx Inc, which develops evidence-based REFERENCES algorithms; consulting for EarlySense, which makes Administration. Health Aff (Millwood). 2014;33(7): patient safety monitoring systems; receiving equity 1. Bates DW, Saria S, Ohno-Machado L, Shah A, 1203-1211. and cash compensation from QPID Inc, a company Escobar G. Big data in health care. Health Aff 7. Nelson KM, Helfrich C, Sun H, et al. focused on intelligence systems for electronic (Millwood). 2014;33(7):1123-1131. Implementation of the patient-centered medical health records; receiving cash compensation from 2. Madan A, Bebrian M, Lazer D, Pentland A. home in the Veterans Health Administration. JAMA CDI (Negev) Ltd, which is a not-for-profit incubator Social sensing to model epidemiological behavior Intern Med. 2014;174(8):1350-1358. for health IT startups; receiving equity from Enelgy, change. In: Proceedings of UBICOMP 2010 12th ACM 8. Sniderman AD, D’Agostino RB Sr, Pencina MJ. which makes software to support evidence-based Conference on Ubiquitous Computing. New York, The role of physicians in the era of predictive clinical decisions, from Ethosmart, which makes NY: Association for Computing Machinery; analytics. JAMA. 2015;314(1):25-26. software to help patients with chronic diseases, and 2010:291-300. from MDClone, which takes clinical data and 9. Amarasingham R, Patzer RE, Huesch M, et al. 3. Escobar GJ, Puopolo KM, Wi S, et al. Implementing electronic health care predictive produces deidentified versions of it. No other Stratification of risk of early-onset sepsis in authors reported disclosures. analytics. Health Aff (Millwood). 2014;33(7):1148-1154.

652 JAMA February 16, 2016 Volume 315, Number 7 (Reprinted) jama.com

Downloaded From: by a Kaiser Permanente User on 10/06/2017