LEGEND: Clinical Insights for Hypertension

bumetanide bumetanide Trials: 40 Comparisons: 10, 278 N = 102 [1148] 33K N = 3502 [212K] 1.9M e168 Whelton et al. JACC VOL. 71, NO. 19, 2018 2017 High Blood Pressure Clinical Practice Guideline MAY 15, 2018:e127– 248

2. The treatment of patients with hypertension without decreases CVD morbidity and mortality. The clinical elevated risk has been systematically understudied trial evidence is strongest for a target BP of 140/90 because lower-risk groupswouldrequireprolonged mm Hg in this population. However, observational follow-up to have a sufﬁcient number of clinical events studies suggest that these individuals often have a high to provide useful information. Although there is clin- lifetime risk and would beneﬁtfromBPcontrolearlier ical trial evidence that both drug and nondrug therapy in life (S8.1.5-19,S8.1.5-20). willFirst-line interrupt the progressive agents course of hypertension (S8.1.5-6),thereisnotrialevidencethatthistreatment 8.1.6. Choice of Initial Medication

Recommendation for Choice of Initial Medication 2017References ACC/AHA that support theGuidelines recommendation are summarized in Online Data Supplement 27 and Systematic Review Report.

COR LOE RECOMMENDATION

1. For initiation of antihypertensive drug therapy, ﬁrst-line agents include thiazide diuretics, CCBs, and ACE SR IA inhibitors or ARBs. (S8.1.6-1,S8.1.6-2)

SR indicates systematic review. “In particular, there is inadequate evidence to support the initial use of beta blockersSynopsis for hypertension in the absence of specificeffective for cardiovascular prevention of CVDcomorbidities.” than other first-step The overwhelming majority of persons with BP suffi- agents, such as thiazide diuretics (S8.1.6-3,S8.1.6-10). ciently elevated to warrant pharmacological therapy may Recommendation-SpecificSupportiveText be best treated initially with 2 agents (see Section 8.1.6.1). When initiation of pharmacological therapy with a single 1. The overall goal of treatment should be reduction in BP, 2018medication ESC/ESH is appropriate, Guidelines: primary consideration Class should I, Levelin A the context of underlying CVD risk. Five drug classes be given to comorbid conditions (e.g., HF, CKD) for which have been shown, in high-quality RCTs, to prevent CVD “Amongspecifi allcclassesofBP-loweringmedicationareindicated antihypertensive drugs, ACE inhibitors,as compared ARBs, with beta-blockers, placebo (diuretics, ACE CCBs, inhibitors, (see Section 9) (S8.1.6-1).Inthelargesthead-to-head ARBs, CCBs, and beta blockers) (S8.1.6-11,S8.1.6-12).In and [thiazide]comparison of diureticsfirst-step drug . . . therapy have for demonstrated hypertension effectivehead-to-head reduction comparisons of of BPfirst-step andCV therapy, events(S8.1.6-3) in RCTs,,thethiazide-typediureticchlorthalidonewas and thus are indicated as thedifferent basis drug of classesantihypertensive have been reported treat- to provide superior to the CCB amlodipine and the ACE inhibitor somewhat divergent capacity to prevent specificCVD mentlisinopril strategies.” in preventing “. . . beta-blockers HF,aBP-relatedoutcomeof are usuallyevents. equivalent Interpretation in preventing ofmeta-analysescomparing major CV increasing importance in the growing population of older agents from different drug classes is challenging events,persons except with hypertension for less(S8.1.6-4 effective—S8.1.6-7) prevention.Addition- of strokebecause the . . relevant . ” RCTs were conducted in different ally, ACE inhibitors were less effective than thiazide di- time periods, during which concurrent antihyperten- uretics and CCBs in lowering BP and in prevention of sive therapy was less or more common, and the efficacy stroke. For black patients, ACE inhibitors were also of agents from certain drug classes may have changed. notably less effective than CCBs in preventing HF (S8.1.6- In recognition of this, some (S8.1.6-2) but not all (S8.1.6- 8) and in the prevention of stroke (S8.1.6-9) (see Section 11,S8.1.6-12) meta-analyses, as well as the largest indi- 10.1). ARBs may be better tolerated than ACE inhibitors vidual RCT that compared first-step agents (S8.1.6-3), in black patients, with less cough and angioedema, but have suggested that diuretics, especially the long-acting according to the limited available experience they offer no thiazide-type agent chlorthalidone, may provide an proven advantage over ACE inhibitors in preventing optimal choice for first-step drug therapy of hyperten- stroke or CVD in this population, making thiazide diuretics sion. In contrast, some meta-analyses have suggested (especially chlorthalidone) or CCBs the best initial choice that beta blockers may be less effective, especially for for single-drug therapy. For stroke, in the general popu- stroke prevention in older adults, but interpretation is lation, beta blockers were less effective than CCBs (36% hampered by inclusion of RCTs that used beta blockers lower risk) and thiazide diuretics (30% lower risk). CCBs that are now considered to be inferior for prevention of have been shown to be as effective as diuretics for CVD (S8.1.6-13,S8.1.6-14).Inasystematicreview reducing all CVD events other than HF, and CCBs are a and network meta-analysis conducted for the good alternative choice for initial therapy when thiazide present guideline, beta blockers were significantly diuretics are not tolerated. Alpha blockers are not used as less effective than diuretics for prevention of stroke first-line therapy for hypertension because they are less and cardiovascular events (S8.1.6-1).Diureticswerealso First-line agents: comparisons from RCTs

RCT − AMI Efﬁcacy outcomes: myocardial infarction, heart failure, stroke

Figure 3.3 Network of clinical trials of antihypertensive drug classes in which myocardial infarction was reported (N=29). *

ACE Target ACEIs CI overlap

full ●

ARBs ● ALPINE THZ partial ARB 0.41 (0.01-5.6)

● ●

cBBs ● ●

● ● ●

dCCBs ● ● ●

● ● ● ● ASCOT-BPLA, ELSA CCB BB TZDs ● ● ● ● 0.89 (0.55-1.5)

*The trials included in each pair-wise comparison are labeled above the arrow. Summary of relative risk (RR) and 95% CIs for the direct comparisons are shown below the arrow. For each pair-wise comparison, the line starts at the reference group and points towards the comparison group. Thus, the arrowhead points to a class of antihypertensive drugs for which a higher RR would signify an increased risk of myocardial infarction.

Comparator cBBs TZDs ARBs

Figure 3.4. Network of clinical trials of antihypertensive drug classes in which heart failure was reported (N=21). ACEIs dCCBs

RCT − AMI Acute myocardial infarction Meta−analysis Efﬁcacy outcome: myocardial infarction, heart failure, stroke RCTs LEGEND

ACEIs ACEIs

ARBs ARBs

cBBs cBBs

dCCBs dCCBs

TZDs TZDs cBBs cBBs TZDs TZDs ARBs ARBs ACEIs ACEIs dCCBs dCCBs

Data source: meta-analysis, 1 2M total patients per study ⇠ Beta blockers underperform alternatives Unexpected: TZDs > ACEIs. Reliable? TZDs vs. ACEIs: hazard ratio

Efﬁcacy outcome: myocardial infarction

Patients Events Patients Events ALLHAT CTD 15,255 1,362 lisinopril 9,054 796 ● ANBP2 HCTZ 3,039 82 enalapril 3,044 58 ● PHYLLIS HCTZ 127 - fosinopril 127 -

Review 18,421 1,447 12,225 854 ●

LEGEND TZDs 631,154 1,122 ACEIs 1,666,241 5,073 ●

0.67 1 1.25 2

TZDs vs. ACEIs calibrated HR: 0.83, 95% CI 0.73 - 0.95, p = 0.01 Is concordant with systematic review, but shows effect difference Suspect previous studies lack power Statistical analysis. We conduct our cohort study using the open-source OHDSI CohortMethod R package (Schuemie et al., Exposed 2018c) , with large-scale analytics achieved through the Cy- TZDs: n = 9801694 clops R package (Suchard et al., 2013) . We use propensity ACEIs: n = 10173421 scores (PSs) – estimates of treatment exposure probability Y conditional on pre-treatment baseline features in the one year TZDs: n = 9204499 prior to treatment initiation – to control for potential mea- Mono−therapy new users ACEIs: n = 8883713 sured confoudning and improve balance between the target N (TZDs) and comparator (ACEIs) cohorts (Rosenbaum and Y Rubin, 1983) . We use an expansive PS model that includes all available patient demographics, drug, condition and proce- TZDs: n = 291454 Having indication ACEIs: n = 510667 dure covariates generated through the FeatureExtraction R N package (Schuemie et al., 2018d) instead of a prespecified set of investigator-selected confounders. We perform PS stratifica- Y tion or variable-ratio matching and then estimate comparative TZDs: n = 0 TZDs-vs-ACEIs hazard ratios (HRs) using a Cox proportional Restricted to common period ACEIs: n = 0 hazards model. Detailed covariate and methods informations N are provided in the Supporting Information. We present PS Y and covariate balance metrics to assess successful confounding control, and provide HR estimates and Kaplan-Meier survival Restricting duplicate TZDs: n = 0 subjects to first cohort ACEIs: n = 0 plots for the outcome of acute myocardial infarction. We addi- N tionally estimate HRs for pre-specified subgroups to evaluate Y interactions with the treatment eect. For eciency reasons, we fit subgroup Cox models using PS stratification only. TZDs: n = 335 No prior outcome ACEIs: n = 2529 Residual study bias from unmeasured and systematic N sources can exist in observational studies after controlling for measured confounding (Schuemie et al., 2014, 2016).Toesti- Y mate such residual bias, we conduct negative control outcome TZDs: n = 1424 experiments with 265 negative control outcomes identified Have at least 1 days at risk ACEIs: n = 3226 through a data-rich algorithm (Voss et al., 2017). We fit the N negative control estimates to an empirical null distribution TZDsY vs. ACEIs: study diagnostics that characterizes the study residual bias and is an important Study population: artifact from which to assess the study design (Schuemie et al., TZDs: n = 303982 2018a) . Using the empirical null distribution and synthetic ACEIs: n = 773286 positive controls (Schuemie et al., 2018b) , we additionally calibrate all HR estimates, their 95% confidence intervals (CIs) and the p-value to reject the null hypothesis of no dierential Fig. 1. Attrition diagram for selecting new-users of TZDs and ACEIs from the CCAE database. eect (HR = 1). Empirical calibration serves as an importantLarge-scale propensity score model controls for observed confounding diagnostic tool to evaluate if residual systematic error is sufficient to cast doubt on the accuracy of the unknown eect estimate. TZDs ACEIs Cohort stratification / balance: 2.5 Results Population characteristics. Figure 1 diagrams the inclusion 2.0 Achieved across all 10,868 of study subjects from the CCAE database under the on- treatment with stratification design. We augment these counts 1.5 baseline characteristics with cohort sizes we identify for the remaining designs in Table 1. This table also reports total patient follow-up time, numbers Density 1.0 (CCAE) of acute myocardial infarction events these patients experience and unadjusted incidence rates. Table 2 compares base-line characteristics between patient cohorts. 0.5 Blood pressure (pop. means in

Patient characteristics balance. Figure 2 plots the preference 0.0 mmHg) (Panther) score distributions, re-scalings of PS estimates to adjust for 0.00 0.25 0.50 0.75 1.00 dierential treatment prevalences, for patients treated with Preference score TZDs and ACEIs. We assess characteristics balance achieved TZDs ACEIs Fig. 2. Preference score distribution for TZDs and ACEIs new-users. The pref- through PS adjustment by comparing all characteristics’ stan- | | erence score is a transformation of the propensity score that adjusts for prevalence before 145/89 145/87 0.13 dardized dierence (StdDi) between treatment group means differences between populations. A higher overlap indicates that subjects in the two before and after PS trimming and stratification (Table 2). populations are more similar in terms of their predicted probability of receiving one after 145/88 145/87 0.02 Figure ?? plots StdDi for all 8863 base-line patient features treatment over the other. that serve as input for the PS model. Before stratification, 59 2 | No BP measurementsSchuemie et al. used in PS model, but still balanced after stratification TZDs vs. ACEIs: study outcomes Table 5. Subgroup analyses. We report HR estimates, their 95% CIs and uncalibrated and calibrated (cal) p-values to reject the null hypothesis of no difference in five pre-specified patient subgroups.

Subjects On-treatment Intent-to-treat Calibration returnsSubgroup near T nominal C HR (95%HR CI) p estimatecal-p HR (95% coverage CI) p cal-p

True hazard ratio = 1 True hazard ratio = 1.5 True hazard ratio = 2 True hazard ratio = 4 1.00

74 estimates 59 estimates 59 estimates 59 estimates Uncalibrated 0.75 73.0% of CIs include 1 47.5% of CIs include 1.5 42.4% of CIs include 2 35.6% of CIs include 4 0.50

0.25

0.00 1.00 74 estimates 59 estimates 59 estimates 59 estimates Calibrated

Standard Error 0.75 94.6% of CIs include 1 93.2% of CIs include 1.5 91.5% of CIs include 2 91.5% of CIs include 4 0.50

0.25

0.00 0.1 0.25 0.5 1 2 4 6 810 0.1 0.25 0.5 1 2 4 6 810 0.1 0.25 0.5 1 2 4 6 810 0.1 0.25 0.5 1 2 4 6 810 Hazard ratio

Fig. 5. Evaluation of effect estimation between TZs and ACEIs new-users. The top plots HRs and their corresponding standard errors before calibration for each negative and synthetic positive control. The bottom plots the same estimates after calibration. Good diagnostics comparable cohorts (observed and • No mention of the drug-condition pair! on a US product – Ingredient labelunobserved); in the “Adverse Drug Reactions” calibration or “Postmarketing”controls for– ATC residual class systematic bias section (Duke et al., 2013), ! • Risk Scores (Charlson comorbidity index) • NoTZDs US spontaneous are reports more suggesting effective that the than pair is in ACEIs in preventing MI an adverse event relationship (Evans et al., 2001; Banda We exclude all covariates that occur in fewer than 0.1% of et al., 2016), patients within the target and comparator cohorts prior to • OMOP vocabulary does not suggest that the drug is model fitting for computational eciency. indicated for the condition, • Vocabulary conditional concepts are usable (i.e., not too References broad, not suggestive of an adverse event relationship, no pregnancy related), and ALLHAT Officers and Coordinators for the ALLHAT Collaborative Research • Exact condition concept itself is used in patient level data. Group (2002). “Major outcomes in high-risk hypertensive patients ran- domized to angiotensin-converting enzyme inhibitor or calcium channel We optimize remaining condition concepts, such that parent blocker vs diuretic: The antihypertensive and lipid-lowering treatment to concepts remove children as defined by the OMOP vocabulary prevent heart attack trial (ALLHAT).” Journal of the American Medical and perform manual review to exclude any pairs that may still Association, 288, 2981–2997. be in a causal relationship or too similar to the study outcome. Austin PC (2009). “Balance diagnostics for comparing the distribution of baseline covariates between treatment groups in propensity-score For hypertension, this process led to a candidate list of 76 matched samples.” Statistics in medicine, 28(25), 3083–3107. negative controls for which table can be found in study protocol Banda JM, Evans L, Vanguri RS, Tatonetti NP, Ryan PB, Shah NH (2016). (https://github.com/OHDSI/Legend/tree/master/Documents). In “A curated and standardized adverse drug event resource to accelerate the comparison of TZs and ACEIs in the Optum database, 74 drug safety research.” Scientific data, 3, 160026. negative controls had sucient outcomes to return estimable Duke J, Friedlin J, Li X (2013). “Consistency in the safety labeling of HRs. We list these conditions in Table 6 bioequivalent medications.” Pharmacoepidemiology and Drug Safety, 22(3), 294–301. Covariate sets. Evans S, Waller PC, Davis S (2001). “Use of proportional reporting ratios (PRRs) for signal generation from spontaneous adverse drug reaction • Demographics (age in 5-year bands, gender, index year, reports.” Pharmacoepidemiology and Drug Safety, 10(6), 483–486. index month) Forouzanfar MH, Liu P, Roth GA, Ng M, Biryukov S, Marczak L, Alexander • Conditions (condition occurrence in lookback window) L, Estep K, Abate KH, Akinyemiju TF, et al. (2017). “Global burden of – in 365 days prior to index date hypertension and systolic blood pressure of at least 110 to 115 mm Hg, 1990-2015.” Journal of the American Medical Association, (2), – in 30 days prior to index date 317 165–182. • Condition aggregation Hripcsak G, Duke J, Shah N, Reich C, Huser V, Schuemie M, Suchard – SMOMED M, Park R, Wong I, Rijnbeek P, et al. (2015). “Observational Health Data Sciences and Informatics (OHDSI): Opportunities for Observational • Drugs (drug occurrence in lookback window) Researchers.” Studies in health technology and informatics, 216, 574– – in 365 days prior to index date 578. Overhage J, Ryan P, Reich C, Hartzema A, Stang P (2012). “Validation of – in 30 days prior to index date a common data model for active safety surveillance research.” Journal • Drug aggregation of the American Medical Informatics Association, 19, 54–60.

Schuemie et al. OHDSI version 1.0 | October 8, 2018 | 5 Cardiovascular efﬁcacy by drug Cardiovascular disease Meta−analysis

Chlorthalidone HCTZ Indapamide Benazepril Captopril Enalapril Fosinopril Lisinopril Quinapril Ramipril Prescriptions are not written Azilsartan Candesartan Irbesartan Losartan at the class-level; must Olmesartan Telmisartan Valsartan choose an individual drug Amlodipine Felodipine Nicardipine Nifedipine for the patient Diltiazem Verapamil Furosemide Spironolactone Atenolol Bisoprolol st nd Metoprolol Nebivolol 1 -line > 2 -line Nadolol Propranolol Carvedilol Labetalol Hydralazine Doxazosin Some within-class Terazosin Clonidine Methyldopa differences failed

HCTZ diagnostics, Nadolol Atenolol Ramipril Enalapril Losartan Lisinopril Captopril Quinapril Labetalol Nebivolol Diltiazem Clonidine Valsartan Terazosin Fosinopril Bisoprolol Verapamil Nifedipine Azilsartan Carvedilol Irbesartan Felodipine Doxazosin Metoprolol Benazepril Amlodipine Nicardipine Propranolol Methyldopa Telmisartan Olmesartan Indapamide Hydralazine Furosemide Candesartan Chlorthalidone Spironolactone e.g. captopril Composite (MI, HF, stroke) outcome in meta-analysis JACC VOL. 71, NO. 19, 2018 Whelton et al. e169 MAY 15, 2018:e127– 248 2017 High Blood Pressure Clinical Practice Guideline

signiﬁcantly better than CCBs for prevention of HF. hypertension (without HF or CKD), initial antihyper- There were some other nonsigniﬁcant differences tensive treatment should include a thiazide diuretic between diuretics, ACE inhibitors, ARBs, and CCBs, but or CCB. the general pattern was for similarity in effect. As indicatedInitiating in Section 8.1.6.1 mono-,mostadultswith vs. combination-therapy hypertension require >1drugtocontroltheirBP. 8.1.6.1. Choice of Initial Monotherapy Versus Initial As recommended in Section 10.1,forblackadultswith Combination Drug Therapy

2017Recommendations ACC/AHA for Choice Guidelines of Initial Monotherapy Versus Initial Combination Drug Therapy*

COR LOE RECOMMENDATIONS

1. Initiation of antihypertensive drug therapy with 2 ﬁrst-line agents of different classes, either as separate I C-EO agents or in a ﬁxed-dose combination, is recommended in adults with stage 2 hypertension and an average BP more than 20/10 mm Hg above their BP target.

2. Initiation of antihypertensive drug therapy with a single antihypertensive drug is reasonable in adults with IIa C-EO stage 1 hypertension and BP goal <130/80 mm Hg with dosage titration and sequential addition of other agents to achieve the BP target.

*Fixed-dose combination antihypertensive medications are listed in Online Data Supplement D.

Synopsis CVD risk, or those who have a history of hypotension or 2018Systematic ESC/ESH review of the Guidelines: evidence comparingClass the initi- I, Leveldrug-associated A side effects. However, caution is ation of antihypertensive treatment with monotherapy advised in initiating antihypertensive pharmacotherapy “Combinationand sequential (stepped-care) treatment titration is recommended of additional with for 2 most drugs in hypertensive older patients because patients hypotension as or initial agents versus initiation of treatment with combination orthostatic hypotension maydevelopinsomepatients; therapy.”therapy (including “. . . nofixed-dose RCT combinations)has compared did not majorBP shouldCV outcomes be carefully monitored. between initial combina- tionidentify therapy any RCTs and meeting monotherapy the systematic review . . .” ques- 2. The stepped-care approach defined by the initiation of tions posed in the PICOTS format (P population, antihypertensive drug therapy with a single agent ¼ I intervention, C comparator, O outcome, T timing, followed by the sequential titration of the dose and ¼ ¼ ¼ ¼ S setting). However, in both ACCORD and SPRINT, 2-drug addition of other agents has been the recommended ¼ therapy was recommended for most participants in the treatment strategy since the first report of the National intensive- but not standard-therapy groups. High Blood Pressure Education Program (S8.1.6.1-7). This approach is also reasonable in older adults or those Recommendation-SpecificSupportiveText at risk or who have a history of hypotension or drug- associated side effects. This strategy has been used 1. Because most patients with hypertension require mul- successfully in nearly all hypertension treatment trials tiple agents for control of their BP and those with higher but has not been formally tested against other antihy- BPs are at greater risk, more rapid titration of antihypertensive drug treatment strategies for effectiveness in pertensive medications began to be recommended in achieving BP control or in preventing adverse outcomes. patients with BP >20/10 mm Hg above their target, beginning with the JNC 7 report (S8.1.6.1-1).Inthese patients, initiation of antihypertensive therapy with 2 8.2. Achieving BP Control in Individual Patients agents is recommended. Evidence favoring this Recommendations for lifestyle modifications and drug approach comes mostly from studies using fixed-dose selection are specified in Sections 6.2, 8.1.4,and8.1.6. combination products showing greater BP lowering Initial drug selections should be based on trial evidence of with fixed-dose combination agents than with single treatment efficacy, combined with recognition of agents, as well as better adherence to therapy (S8.1.6.1- compelling indications for use of an agent from a specific 2,S8.1.6.1-3).Thesafetyandefficacy of this strategy drug class, as well as the individual patient’slifestyle have been demonstrated in adults to reduce BPs preferences and traits. For a subset of patients (25% to to <140/90 mm Hg though not compared with other 50%) (S8.2-1),theinitialdrugtherapywillbewell strategies (S8.1.6.1-4—S8.1.6.1-6).Ingeneral,this tolerated and effective in achieving the desired level of approach is reasonable in older adults, those at high BP, with only the need for subsequent monitoring Mono- vs. combo-therapy

Efﬁcacy outcome:Acutemyocardial myocardial infarction infarction Meta−analysis, heart failure, stroke

Similar efﬁcacy ACEIs for HF and stroke ARBs

cBBs Recommended: no ACEIs + dCCBs ARBs combos TZDs ACEIs + TZDs combos are most cBBs & TZDs ARBs & cBBs ARBs & TZDs ACEIs & cBBs ACEIs ACEIs & TZDs ACEIs

ACEIs & ARBs ACEIs prevalent cBBs & dCCBs dCCBs & TZDs ARBs & dCCBs ACEIs & dCCBs ACEIs

ACEIs vs. ACEIs + TZDs TZDs vs. ACEIs + TZDs Mono- vs. combo-therapy

Safety outcome: kidneyKidney diseasedisease Meta−analysis

ACEIs

ARBs Antihypertensives

cBBs tend to affect

dCCBs kidney function

TZDs Initiating combo increases risk cBBs & TZDs ARBs & cBBs ARBs & TZDs ACEIs & cBBs ACEIs ACEIs & TZDs ACEIs ACEIs & ARBs ACEIs cBBs & dCCBs dCCBs & TZDs ARBs & dCCBs ACEIs & dCCBs ACEIs

ACEIs vs. ACEIs + TZDs – ACEIs are more prevalent TZDs vs. ACEIs + TZDs Statistical analysis. We conduct our cohort study using the open-source OHDSI CohortMethod R package (Schuemie et al.,

2018c) , with large-scale analytics achieved through the Cy- Exposed: clops R package (Suchard et al., 2013) . We use propensity ACE inhibitors: n = 779041 scores (PSs) – estimates of treatment exposure probability ACE inhibitors and thiazide or thiazide−like diuretics: n = 233420 conditional on pre-treatment baseline features in the one year prior to treatment initiation – to control for potential measured Y confoudning and improve balance between the target (ACEIs) and comparator (ACEIs and TZDs) cohorts (Rosenbaum and ACE inhibitors: n = 2 Restricted to common period ACE inhibitors and thiazide or thiazide−like diuretics: n = 0 Rubin, 1983) . We use an expansive PS model that includes N all available patient demographics, drug, condition and procedure covariates generated through the FeatureExtraction R Y package (Schuemie et al., 2018d) instead of a prespecified set of investigator-selected confounders. We perform PS stratification or variable-ratio matching and then estimate comparative Restricting duplicate ACE inhibitors: n = 0 subjects to first cohort ACE inhibitors and thiazide or thiazide−like diuretics: n = 0 ACEIs-vs-ACEIs and TZDs hazard ratios (HRs) using a Cox N proportional hazards model. Detailed covariate and methods informations are provided in the Supporting Information.We Y present PS and covariate balance metrics to assess successful confounding control, and provide HR estimates and Kaplan- ACE inhibitors: n = 8268 No prior outcome ACE inhibitors and thiazide or thiazide−like diuretics: n = 1054 Meier survival plots for the outcome of chronic kidney disease. N We additionally estimate HRs for pre-specified subgroups to Kidney disease in CCAE evaluate interactions with the treatment eect. For eciency Y reasons, we fit subgroup Cox models using PS stratification only. ACE inhibitors: n = 3200 Residual study bias from unmeasured and systematic Have at least 1 days at risk ACE inhibitors and thiazide or thiazide−like diuretics: n = 1054 sources can exist in observational studies after controlling for N Table 1. Patient cohorts. Target (T) cohort is ACEIs new-users. Comparative (C) cohort is ACEIs and TZDs new-users. We report total number measured confounding (Schuemie et al., 2014, 2016).Toesti- of patients, follow-up time (in years), number of chronic kidney disease events, and event incidence rate (IR) per 1,000 patient years (PY) mate such residual bias, we conduct negative control outcome Y experiments with 292 negative control outcomes identifiedin patient cohorts, as well as the their minimum detectable relative risk (MDRR). Note that the IR does not account for any stratification or through a data-rich algorithm (Voss et al., 2017). We fit thematching.Study population: ACE inhibitors: n = 767571 negative control estimates to an empirical null distribution ACE inhibitors and thiazide or thiazide−like diuretics: n = 231312 Patients PYs Events IR that characterizes the study residual bias and is an important artifact from which to assess the study design (Schuemie et al., Design T C T C T C T C MDRR 2018a) . Using the empirical null distribution and synthetic Fig. 1. AttritionPS diagramstratification, for selecting on-treatment new-users of ACE767,571 inhibitors and231,312 ACE 579,641 153,465 3,473 980 5.99 6.39 1.10 positive controls (Schuemie et al., 2018b) , we additionally inhibitors and thiazide or thiazide-like diuretics from the CCAE database. calibrate all HR estimates, their 95% confidence intervals (CIs) PS stratification, intent-to-treat 769,659 232,022 2,089,649 635,031 11,872 3,673 5.68 5.78 1.05 and the p-value to reject the null hypothesis of no dierential eect (HR = 1). Empirical calibration serves as an important diagnostic tool to evaluate if residual systematic error is suf- ACEIs ACEIs and TZDs ficient to cast doubt on the accuracy of the unknown eect estimate. 0.25

Table 2. Patient3 demographics. We report the standardized differ- Results ence of population means (StdDiff) before and after stratification for 0.20 Population characteristics. Figure 1 diagrams the inclusionselected base-line patient characteristics. of study subjects from the CCAE database under the on- 2 Cohort stratification / balance: Before stratification After stratification treatment with stratification design. We augment these counts 0.15 Characteristic T (%) C (%) StdDiff T (%) C (%) StdDiff with cohort sizes we identify for the remaining designs in Table Density Age group 1. This table also reports total patient follow-up time, numbers Achieved across all 8,541 15-19 1 0.7 0.2 0.07 0.6 0.5 0.02 of chronic kidney disease events these patients experience20-24 1.4 0.9 0.04 1.3 1.2 0.00 0.10 25-29 2.6 2.5 0.00 2.6 2.4 0.01 and unadjusted incidence rates. Table 2 compares base-line baselineAfter stratification characteristics 30-34 5.0 5.2 -0.01 5.1 4.9 0.01 characteristics between patient cohorts. 35-39 8.1 8.7 -0.02 8.2 8.2 0.00 50-54 0 18.7 19.2 -0.01 18.8 18.9 0.00 0.05 Patient characteristics balance. Figure 2 plots the preference55-59 0.00 0.25 0.50 18.3 0.7517.8 0.011.00 18.2 18.4 0.00 Remarkable overlap in cohorts 60-64 Preference score15.5 14.0 0.04 15.1 15.6 -0.01 score distributions, re-scalings of PS estimates to adjust for65-69 1.3 1.1 0.02 1.3 1.3 0.00 dierential treatment prevalences, for patients treated withGender:Fig. female 2. Preference score distribution for ACEIs38.4 and ACEIs46.4 and-0.16 TZDs new-users.40.3 40.6 -0.01 0.00 ACEIs and ACEIs and TZDs. We assess characteristics balanceMedicalThe history: preference General score is a transformation of the propensity score that adjusts for Acute respiratory disease 24.5 23.2 0.03 24.3 24.4 0.00 prevalence differences between populations. A higher overlap indicates that subjects 0.00 0.05 0.10 0.15 0.20 0.25 achieved through PS adjustment by comparing all character-Attention deficit hyperactivity disorder 1.2 0.8 0.04 1.1 1.1 0.00 in the two populations are more similar in terms of their predicted probability of Before stratification istics’ standardized dierence (StdDi) between treatmentChronic liver disease 1.5 1.1 0.03 1.4 1.4 0.00 receiving one treatment over the other. group means before and after PS trimming and stratificationChronic obstructive lung disease 1.8 1.5 0.02 1.7 1.8 0.00 Crohn’s disease 0.3 0.2 0.01 0.3 0.2 0.00 Fig. 3. Patient characteristics balance before and after stratification. As a rule- (Table 2). Figure ?? plots StdDi for all 8541 base-line patientDepressive disorder 7.4 6.3 0.04 7.1 7.4 -0.01 Diabetes mellitus 18.3 9.8 0.24 16.4 15.8 0.02 of-thum, all values < 0.1 is generall considered well-balance (Austin, 2009). Gastroesophageal reflux disease 7.8 6.4 0.05 7.5 7.5 0.00 2 | Human immunodeficiency virus infection 0.2 0.2 0.01Schuemie0.2 et al.0.2 -0.01 Lesion of liver 0.2 0.2 0.02 0.2 0.2 0.00 Obesity 8.5 9.8 -0.04 8.8 8.7 0.00 features that serve as input for the PS model. Before stratifi- Osteoarthritis 11.3 10.1 0.04 11.0 11.3 -0.01 Pneumonia 1.5 1.3 0.02 1.4 1.4 0.00 cation, 70 features have a StdDi > 0.1. After stratification, Psoriasis 1.0 0.8 0.02 1.0 1.0 0.00 Renal impairment 1.1 0.6 0.05 0.9 0.9 0.01 the count is 0. Rheumatoid arthritis 0.8 0.7 0.01 0.8 0.7 0.00 Ulcerative colitis 0.3 0.2 0.02 0.3 0.2 0.00 Urinary tract infectious disease 5.1 4.7 0.02 5.0 5.1 0.00 Outcome assessment. Table 3 details the time to first chronic Viral hepatitis C 0.4 0.4 0.01 0.4 0.4 0.00 Visual system disorder 15.5 12.9 0.08 14.9 15.0 0.00 kidney disease or censoring distributions for patients in the Medical history: Cardiovascular disease ACEIs and ACEIs and TZDs cohorts. We report in Table 4 Atrial fibrillation 0.4 0.2 0.04 0.4 0.4 0.00 Cerebrovascular disease 1.7 1.0 0.06 1.6 1.5 0.01 estimated HRs comparing ACEIs to ACEIs and TZDs for the Coronary arteriosclerosis 2.2 1.1 0.09 2.0 1.8 0.02 Heart disease 9.0 6.2 0.11 8.4 8.3 0.00 on-treatment and intent-to-treat designs with stratification Heart failure 0.5 0.3 0.02 0.4 0.5 0.00 or matching. Figure 4 plots Kaplan-Meier survival curves Ischemic heart disease 1.7 0.9 0.07 1.5 1.4 0.01 Peripheral vascular disease 4.1 2.9 0.07 3.9 3.8 0.00 for patients under the intent-to-treat design. To examine Pulmonary embolism 0.2 0.1 0.02 0.2 0.2 0.00 Medical history: Neoplasms possible subgroup dierences in treatment-eect, we include Hematologic neoplasm 0.5 0.4 0.02 0.5 0.5 0.00 Table{tab:subgroups} that reports HR estimates separately Malignant lymphoma 0.2 0.1 0.01 0.2 0.2 0.00 Malignant neoplasm of anorectum 0.1 0.1 0.01 0.1 0.1 0.00 for children (age < 18), the elderly (age 65), female pa- Malignant neoplastic disease 4.2 3.2 0.05 4.0 4.0 0.00 Malignant tumor of breast 0.7 0.6 0.01 0.7 0.7 0.00 tients, pregnant women, patients with hepatic impairment and Malignant tumor of colon 0.2 0.2 0.02 0.2 0.2 0.01 patients with renal impairment, using PS stratification. Malignant tumor of lung 0.1 0.1 0.01 0.1 0.1 0.00 Malignant tumor of urinary bladder 0.1 0.1 0.01 0.1 0.1 0.00 Primary malignant neoplasm of prostate 0.5 0.4 0.02 0.5 0.5 0.00 Medication use Residual systematic error. In the absense of bias, we expect Antibacterials for systemic use 48.8 45.1 0.07 48.0 48.0 0.00 95% of negative and positive control estimate 95% confidence Antidepressants 17.7 15.4 0.06 17.2 17.5 -0.01 Antiepileptics 6.2 5.0 0.05 6.0 5.9 0.00 intervals to include their presumed HR. In the case of negative Antiinflammatory and antirheumatic products 24.0 23.4 0.01 23.9 24.0 0.00 Antineoplastic agents 1.4 1.1 0.03 1.4 1.3 0.00 controls, the presumed HR = 1. Figure 5 describes the negative Antipsoriatics 0.4 0.3 0.02 0.4 0.4 0.00 and positive control estimates under the on-treatment with PS Antithrombotic agents 3.3 2.1 0.08 3.1 2.8 0.01 Beta blocking agents 0.5 0.4 0.01 0.4 0.5 -0.01 stratification design. Before calibration, negative and positive Drugs for acid related disorders 14.1 11.6 0.07 13.6 13.6 0.00 Drugs for obstructive airway diseases 18.1 16.2 0.05 17.7 17.6 0.00 controls demonstrate poor coverage. After calibration, controls Drugs used in diabetes 15.6 7.5 0.25 13.8 13.1 0.02 demonstrate acceptable coverage. Immunosuppressants 1.5 1.2 0.03 1.4 1.4 0.00 Lipid modifying agents 24.6 17.2 0.18 22.9 22.8 0.00 Opioids 15.2 14.2 0.03 15.2 14.8 0.01 Psycholeptics 17.6 14.9 0.07 17.0 17.0 0.00 Conclusions Psychostimulants, agents used for adhd and nootropics 2.9 2.3 0.03 2.7 2.8 0.00 We find that ACEIs has a lower risk of chronic kidney disease as compared to ACEIs and TZDs within the population that the CCAE represents.

Schuemie et al. OHDSI version 1.0 | October 9, 2018 | 3 Table 3. Time-at-risk distributions as percentiles in the target and comparator cohorts after stratiﬁcation.

min 10% 25% 50% 75% 90% max ACEIs 1 29 37 116 528 718 6,087 ACEIs and TZDs 1 29 29 98 451 627 5,989

Table 4. HazardTable ratio (HR) 3. Time-at-risk estimates and distributions their conﬁdence as intervals percentiles (CIs) and inp the-value target to reject and the comparator null hypothesis cohorts of no difference after stratiﬁcation. (HR = 1) under various designs.

minUncalibrated 10% 25% 50% Calibrated 75% 90% max Kidney diseaseDesignACEIs in HR1 (95% CCAE CI)29 p37 HR116 (95% CI)528 p718 6,087 PS stratiﬁcation,ACEIs on-treatment and TZDs 0.781 (0.72 - 0.84)29 0.0029 0.7898 (0.71 - 0.87)451 0.00627 5,989 PS stratiﬁcation, intent-to-treat 0.79 (0.76 - 0.82) 0.00 0.80 (0.73 - 0.89) 0.00

Table 4. Hazard ratio (HR) estimates and their confidence intervals (CIs) and p-value to reject the null hypothesis of no difference (HR = 1) under various designs. Residual systematic error. In the absense of bias, we expect 95% of negative and positive control estimate 95% confidence intervals to include their presumed HR. In the case of negative Uncalibrated Calibrated controls, the presumed HR = 1. Figure 5 describes the negative Design HRand (95% positive CI) controlp estimatesHR under (95% the CI) on-treatmentp with PS stratification design. Before calibration, negative and positive PS stratification, on-treatment 0.78controls (0.72 - demonstrate 0.84) 0.00 poor0.78 coverage. (0.71 -After 0.87) calibration,0.00 controls PS stratification, intent-to-treat 0.79demonstrate (0.76 - 0.82) acceptable0.00 coverage.0.80 (0.73 - 0.89) 0.00

Conclusions ACEIs ACEIs and TZDs We find thatResidual ACEIs has systematic a lower risk error. of chronicIn thekidney absense disease of bias, we expect 1.0000 as compared to ACEIs and TZDs within the population that the CCAEMeta-analysis95% represents. of negative and positive safety control estimateestimates: 95% confidence intervals to include their presumed HR. In the case of negative 0.9975 Supportingcontrols, Information the presumed HR = 1. FigureHR5 describes the negative and positive control estimates under the on-treatment with PS 0.9950 Here we enumerate the guiding principles of LEGEND and provide linkingstratification details on design.study cohorts Before and calibration, design. negative and positive controls demonstrate poor coverage. After calibration, controls Survival probability 0.9925 ACEIs 0.84 (0.77 - 0.91) LEGEND principles.demonstrate acceptable coverage. TZDs 0.80 (0.67 - 0.96) 0.9900 1. Evidence will be generated at large-scale. 2. DisseminationConclusions of the evidence will not depend on the 0 250 500 estimated eects. ACEIsTime in days ACEIs and TZDs 3. EvidenceWe find will bethat generated ACEIs by has consistently a lower risk applying of chronic a kidney disease Number at risk systematic approach across all research questions. 1.0000ACEIs 769,659 595,554 453,889 as compared to ACEIs and TZDs within the population that ACEIs and TZDs 232,022 180,936 139,197 4. EvidenceMono will be generated vs. combo-therapy: using a pre-specified analysis 0 250 500 the CCAE represents. Time in days design. 5. Evidence will be generated using open source software 0.9975 Fig. 4. Kaplan Meier plot of chronic kidney disease-free survival. This plot is that is freely availableComparable to all. efficacy adjusted for the propensity score stratification; the ACEIs curve shows the actual 6. EvidenceSupporting generation process Information will be empirically evaluated observed survival. The ACEIs and TZDs curve applies reweighting to approximate by including control research questions where the true the counterfactual0.9950 of what ACEIs survival would look like had the ACEIs cohort been Here we enumerate the guiding principles of LEGEND and exposed to ACEIs and TZDs instead. The shaded area denotes the 95% CI. eect size is known.Greater safety risk 7. Evidenceprovide will be linking generated details using onbest-practices. study cohorts and design. 8. LEGEND will not be used to evaluate methods.

Survival probability 0.9925 9. EvidenceLEGEND will be principles. updated on a regular basis. 10. No patient-level data will be shared between sites in the network, only aggregated data. 0.9900 1. Evidence will be generated at large-scale. Study cohorts.2. DisseminationPlease see the of LEGEND the evidence hypertension will not depend on the 0 250 500 Study protocolestimated (https://github.com/OHDSI/Legend/tree/master/ eects. Time in days Documents)3. forEvidence complete specification will be generated of the ACEIs, by ACEIs consistently applying a Number at risk and TZDs andsystematic chronic kidney approach disease cohorts across using all ATLAS research questions. ACEIs 769,659 595,554 453,889 (http://www.ohdsi.org/web/atlas). ACEIs and TZDs 232,022 180,936 139,197 4. Evidence will be generated using a pre-specified analysis 0 250 500 4 | Time in days design. Schuemie et al. 5. Evidence will be generated using open source software Fig. 4. Kaplan Meier plot of chronic kidney disease-free survival. This plot is that is freely available to all. adjusted for the propensity score stratification; the ACEIs curve shows the actual 6. Evidence generation process will be empirically evaluated observed survival. The ACEIs and TZDs curve applies reweighting to approximate the counterfactual of what ACEIs survival would look like had the ACEIs cohort been by including control research questions where the true exposed to ACEIs and TZDs instead. The shaded area denotes the 95% CI. eect size is known. 7. Evidence will be generated using best-practices. 8. LEGEND will not be used to evaluate methods. 9. Evidence will be updated on a regular basis. 10. No patient-level data will be shared between sites in the network, only aggregated data.

Study cohorts. Please see the LEGEND hypertension Study protocol (https://github.com/OHDSI/Legend/tree/master/ Documents) for complete speciﬁcation of the ACEIs, ACEIs and TZDs and chronic kidney disease cohorts using ATLAS (http://www.ohdsi.org/web/atlas).

4 | Schuemie et al. JACC VOL. 71,Chlorthalidone NO. 19, 2018 vs. hydrochlorothiazide Whelton et al. e165 MAY 15, 2018:e127– 248 2017 High Blood Pressure Clinical Practice Guideline

2017 ACC/AHA Guidelines

TABLE 18 Oral Antihypertensive Drugs

Usual Dose, Daily Class Drug Range (mg/d)* Frequency Comments

Primary agents

Thiazide or thiazide-type Chlorthalidone 12.5–25 1 n Chlorthalidone is preferred on the basis of prolonged half-life and diuretics proven trial reduction of CVD. Hydrochlorothiazide 25–50 1 n Monitor for hyponatremia and hypokalemia, uric acid and calcium Indapamide 1.25–2.5 1 levels. n Use with caution in patients with history of acute gout unless patient Metolazone 2.5–51is on uric acid–lowering therapy.

ACE inhibitors Benazepril 10–40 1 or 2 n Do not use in combination with ARBs or direct renin inhibitor. n There is an increased risk of hyperkalemia, especially in patients with Captopril 12.5–150 2 or 3 CKD or in those on Kþ supplements or Kþ-sparing drugs. Diuretic ComparisonEnalapril Project 5–40 (2022) 1 or 2 n There is a risk of acute renal failure in patients with severe bilateral renal artery stenosis. – Fosinopril 10 40 1 n Do not use if patient has history of angioedema with ACE inhibitors. Lisinopril 10–40 1 n Avoid in pregnancy.

Moexipril 7.5–30 1 or 2

Perindopril 4–16 1

Quinapril 10–80 1 or 2

Ramipril 2.5–20 1 or 2

Trandolapril 1–41

ARBs Azilsartan 40–80 1 n Do not use in combination with ACE inhibitors or direct renin inhibitor. n There is an increased risk of hyperkalemia in CKD or in those on Kþ Candesartan 8–32 1 supplements or Kþ-sparing drugs. Eprosartan 600–800 1 or 2 n There is a risk of acute renal failure in patients with severe bilateral renal artery stenosis. – Irbesartan 150 300 1 n Do not use if patient has history of angioedema Losartan 50–100 1 or 2 with ARBs. Patients with a history of angioedema with an ACE inhibitor can receive an ARB beginning 6 weeks after ACE inhibitor is Olmesartan 20–40 1 discontinued. Telmisartan 20–80 1 n Avoid in pregnancy.

Valsartan 80–320 1

CCB—dihydropyridines Amlodipine 2.5–10 1 n Avoid use in patients with HFrEF; amlodipine or felodipine may be used if required. Felodipine 2.5–10 1 n They are associated with dose-related pedal edema, which is more Isradipine 5–10 2 common in women than men.

Nicardipine SR 60–120 2

Nifedipine LA 30–90 1

Nisoldipine 17–34 1

CCB—nondihydropyridines Diltiazem ER 120–360 1 n Avoid routine use with beta blockers because of increased risk of bradycardia and heart block. Verapamil IR 120–360 3 n Do not use in patients with HFrEF. Verapamil SR 120–360 1 or 2 n There are drug interactions with diltiazem and verapamil (CYP3A4 major substrate and moderate inhibitor). Verapamil-delayed 100–300 1 (in the onset ER evening)

Secondary agents

Diuretics—loop Bumetanide 0.5–22n These are preferred diuretics in patients with symptomatic HF. They are preferred over thiazides in patients with moderate-to-severe CKD – Furosemide 20 80 2 (e.g., GFR <30 mL/min). Torsemide 5–10 1 Diuretics—potassium sparing Amiloride 5–10 1 or 2 n These are monotherapy agents and minimally effective antihyper- Triamterene 50–100 1 or 2 tensive agents. n Combination therapy of potassium-sparing diuretic with a thiazide can be considered in patients with hypokalemia on thiazide monotherapy. n Avoid in patients with signiﬁcant CKD (e.g., GFR <45 mL/min).

Continued on the next page tial prescriptions that have fewer than 30 days gap between prescriptions. Exposed: chlorthalidone: n = 451706 Statistical analysis. We conduct our cohort study using the hydrochlorothiazide: n = 9385478 open-source OHDSI CohortMethod R package (Schuemie et al., Y 2018c) , with large-scale analytics achieved through the Cy- chlorthalidone: n = 429260 clops R package (Suchard et al., 2013) . We use propensity Mono−therapy new users hydrochlorothiazide: n = 8818043 scores (PSs) – estimates of treatment exposure probability N conditional on pre-treatment baseline features in the one year Y prior to treatment initiation – to control for potential measured confoudning and improve balance between the target chlorthalidone: n = 8296 Having indication hydrochlorothiazide: n = 278666 (chlorthalidone) and comparator (hydrochlorothiazide) cohorts N (Rosenbaum and Rubin, 1983) . We use an expansive PS model that includes all available patient demographics, drug, Y condition and procedure covariates generated through the Fea- chlorthalidone: n = 0 tureExtraction R package (Schuemie et al., 2018d) instead Restricted to common period hydrochlorothiazide: n = 13 of a prespecified set of investigator-selected confounders. We N perform PS stratification or variable-ratio matching and then Y estimate comparative chlorthalidone-vs-hydrochlorothiazide hazard ratios (HRs) using a Cox proportional hazards model. Restricting duplicate chlorthalidone: n = 0 subjects to first cohort hydrochlorothiazide: n = 0 Detailed covariate and methods informations are provided in N the Supporting Information. We present PS and covariate Y balance metrics to assess successful confounding control, and CTD vs. HCTZ: risk of MI in CCAE provide HR estimates and Kaplan-Meier survival plots for chlorthalidone: n = 27 No prior outcome hydrochlorothiazide: n = 299 the outcome of acute myocardial infarction. We additionally N estimate HRs for pre-specified subgroups to evaluate interactions with the treatment eect. For eciency reasons, we fit Y subgroup Cox models using PS stratification only. Table 1. Patient cohorts. Target (T) cohort is chlorthalidone new-users. Comparative (C) cohort is hydrochlorothiazide new-users. We report chlorthalidone: n = 46 Have at least 1 days at risk Residual study bias from unmeasured and systematictotal number of patients, follow-uphydrochlorothiazide: time (in years), n = 1365 number of acute myocardial infarction events, and event incidence rate (IR) per 1,000 sources can exist in observational studies after controlling forpatient years (PY) in patient cohorts,N as well as the their minimum detectable relative risk (MDRR). Note that the IR does not account for any measured confounding (Schuemie et al., 2014, 2016).Toesti-stratification or matching.Y mate such residual bias, we conduct negative control outcome Study population: experiments with 187 negative control outcomes identified chlorthalidone: n = 14077 Patients PYs Events IR through a data-rich algorithm (Voss et al., 2017). We fit the hydrochlorothiazide: n = 287092Design T C T C T C T C MDRR negative control estimates to an empirical null distribution that characterizes the study residual bias and is an important PS stratification, on-treatment 14,077 287,092 8,513 200,715 11 348 1.29 1.73 2.01 artifact from which to assess the study design (Schuemie et al., Fig. 1. Attrition diagram for selecting new-users of chlorthalidone and hy- drochlorothiazidePS from stratification, the CCAE database. intent-to-treat 14,098 288,073 32,565 888,405 72 1,806 2.21 2.03 1.36 2018a) . Using the empirical null distribution and synthetic positive controls (Schuemie et al., 2018b) , we additionally calibrate all HR estimates, their 95% confidence intervals (CIs) and the p-value to reject the null hypothesis of no dierential chlorthalidone hydrochlorothiazide eect (HR = 1). Empirical calibration serves as an importantTable 2. Patient demographics. We report the standardized differ- diagnostic tool to evaluate if residual systematic error is suf-ence of population means (StdDiff) before and after stratification for ficient to cast doubt on the accuracy of the unknown eectselected base-line3 patient characteristics. estimate. Before stratification After stratification Characteristic T (%) C (%) StdDiff T (%) C (%) StdDiff 0.2 2 Results Age group Cohort stratification / balance: 15-19 0.4 0.6 -0.03 0.4 0.6 -0.03

Population characteristics. Figure 1 diagrams the inclusion20-24 Density 1.6 1.6 0.00 1.4 1.6 -0.02 of study subjects from the CCAE database under the on-25-29 3.0 3.6 -0.03 3.4 3.5 -0.01 Achieved across all 8,773 treatment with stratiﬁcation design. We augment these counts30-34 1 6.0 6.6 -0.02 6.0 6.6 -0.02 with cohort sizes we identify for the remaining designs in Table35-39 8.9 9.9 -0.03 9.5 9.8 -0.01

After stratification 0.1 1. This table also reports total patient follow-up time, numbers40-44 13.1 13.4 -0.01 13.2 13.4 0.00 baseline characteristics 45-49 15.5 16.4 -0.02 15.9 16.4 -0.01 of acute myocardial infarction events these patients experience50-54 0 18.4 17.6 0.02 18.1 17.7 0.01 and unadjusted incidence rates. Table 2 compares base-line55-59 0.00 0.25 0.5017.2 16.00.75 0.03 1.0017.1 16.1 0.03 Sufficient overlap in cohorts characteristics between patient cohorts. 60-64 Preference14.5 score13.1 0.04 13.8 13.2 0.02 65-69 1.3 1.1 0.02 1.1 1.1 0.00 Gender:Fig. female 2. Preference score distribution51.8 for chlorthalidone61.1 -0.19 and hydrochloroth-60.9 60.6 0.00 Patient characteristics balance. Figure 2 plots the preferenceMedicaliazide history: new-users. General The preference score is a transformation of the propensity score 0.0 score distributions, re-scalings of PS estimates to adjust forAcutethat respiratory adjusts disease for prevalence differences between23.4 populations.26.3 -0.07 A higher25.9 overlap26.1 indi- 0.00 0.0 0.1 0.2 dierential treatment prevalences, for patients treated withAttentioncates deficit that hyperactivity subjects in the disorder two populations are1.5 more similar1.1 in terms0.04 of their1.0 predicted1.1 0.00 Before stratification chlorthalidone and hydrochlorothiazide. We assess character-Chronicprobability liver disease of receiving one treatment over the1.1 other. 1.1 0.00 1.0 1.1 -0.01 istics balance achieved through PS adjustment by comparingChronic obstructive lung disease 1.2 1.4 -0.02 1.4 1.4 0.00 Dementia 0.1 0.1 -0.02 0.1 0.1 -0.02 Fig. 3. Patient characteristics balance before and after stratification. As a rule- Depressive disorder 8.3 8.1 0.00 7.7 8.2 -0.02 of-thum, all values < 0.1 is generall considered well-balance (Austin, 2009). 2 | Diabetes mellitus 4.5 4.6 0.00 Schuemie4.8 et4.6 al. 0.01 Gastroesophageal reflux disease 8.1 7.5 0.02 7.3 7.5 -0.01 Human immunodeficiency virus infection 0.3 0.3 0.01 0.2 0.3 -0.02 Hyperlipidemia 27.0 25.3 0.04 26.6 25.4 0.03 all characteristics’ standardized dierence (StdDi)between Lesion of liver 0.2 0.2 0.00 0.2 0.2 -0.01 treatment group means before and after PS trimming and Obesity 13.7 9.9 0.12 9.9 10.1 -0.01 Osteoarthritis 11.3 10.6 0.02 10.8 10.6 0.01 stratification (Table 2). Figure ?? plots StdDi for all 8773 Pneumonia 1.4 1.4 -0.01 1.7 1.4 0.02 base-line patient features that serve as input for the PS model. Psoriasis 1.0 0.9 0.02 0.7 0.9 -0.02 Renal impairment 1.0 0.5 0.06 0.7 0.5 0.02 Before stratification, 37 features have a StdDi > 0.1. After Rheumatoid arthritis 0.9 0.8 0.01 1.2 0.8 0.04 stratification, the count is 0. Ulcerative colitis 0.3 0.2 0.01 0.2 0.2 0.00 Urinary tract infectious disease 5.3 6.4 -0.04 6.4 6.3 0.00 Viral hepatitis C 0.3 0.3 0.00 0.3 0.3 -0.01 Outcome assessment. Table 3 details the time to first acute Visual system disorder 15.1 14.9 0.00 14.6 14.9 -0.01 Medical history: Cardiovascular disease myocardial infarction or censoring distributions for patients Atrial fibrillation 0.3 0.2 0.02 0.3 0.2 0.00 in the chlorthalidone and hydrochlorothiazide cohorts. We Cerebrovascular disease 1.0 1.0 0.00 1.2 1.0 0.02 Coronary arteriosclerosis 1.1 1.0 0.02 1.2 1.0 0.02 report in Table 4 estimated HRs comparing chlorthalidone Heart disease 7.2 6.4 0.03 6.9 6.4 0.02 to hydrochlorothiazide for the on-treatment and intent-to- Heart failure 0.4 0.3 0.01 0.3 0.3 0.01 Peripheral vascular disease 3.9 3.2 0.04 3.5 3.2 0.02 treat designs with stratification or matching. Figure 4 plots Pulmonary embolism 0.2 0.2 0.00 0.2 0.2 -0.01 Kaplan-Meier survival curves for patients under the intent- Medical history: Neoplasms Hematologic neoplasm 0.5 0.4 0.01 0.4 0.4 -0.01 to-treat design. To examine possible subgroup dierences in Malignant lymphoma 0.2 0.2 0.02 0.3 0.2 0.02 treatment-eect, we include Table{tab:subgroups} that reports Malignant neoplasm of anorectum 0.1 0.1 -0.01 0.1 0.1 0.00 Malignant neoplastic disease 3.9 3.8 0.01 4.0 3.8 0.01 HR estimates separately for children (age < 18), the elderly Malignant tumor of breast 0.9 1.0 -0.01 1.2 1.0 0.02 (age 65), female patients, pregnant women, patients with Malignant tumor of colon 0.1 0.2 -0.01 0.1 0.2 -0.02 Malignant tumor of urinary bladder 0.1 0.1 0.00 0.1 0.1 0.01 hepatic impairment and patients with renal impairment, using Primary malignant neoplasm of prostate 0.4 0.3 0.00 0.3 0.3 0.00 PS stratification. Medication use Antibacterials for systemic use 46.1 50.9 -0.10 50.9 50.6 0.00 Antidepressants 17.7 19.2 -0.04 18.9 19.2 -0.01 ## Warning: package ’dplyr’ was built under R version 3.4.4 Antiepileptics 6.1 6.0 0.00 6.2 6.0 0.01 Antiinflammatory and antirheumatic products 24.6 26.4 -0.04 26.4 26.3 0.00 ## Antineoplastic agents 1.5 1.5 0.01 1.7 1.5 0.02 Antipsoriatics 0.4 0.4 0.00 0.4 0.4 0.00 ## Attaching package: ’dplyr’ Antithrombotic agents 2.7 2.2 0.03 2.7 2.2 0.03 Beta blocking agents 0.5 0.4 0.01 0.4 0.4 -0.01 ## The following objects are masked from ’package:stats’: Drugs for acid related disorders 13.4 14.0 -0.02 13.9 14.0 0.00 Drugs for obstructive airway diseases 21.0 20.3 0.02 20.5 20.3 0.00 ## Drugs used in diabetes 2.9 3.1 -0.01 3.2 3.0 0.01 ## filter, lag Immunosuppressants 1.7 1.5 0.02 1.4 1.5 -0.01 Lipid modifying agents 14.2 13.5 0.02 14.2 13.6 0.02 Opioids 15.6 16.0 -0.01 16.2 16.0 0.01 ## The following objects are masked from ’package:base’: Psycholeptics 18.3 18.2 0.00 18.8 18.2 0.01 ## ## intersect, setdiff, setequal, union

Schuemie et al. OHDSI version 1.0 | October 9, 2018 | 3 CTD vs. HCTZ: risk of MI in CCAE

Table 4. Hazard ratio (HR) estimates and their conﬁdence intervals (CIs) and p-value to reject the null hypothesis of no difference (HR = 1) under various designs.

Table 3. Time-at-risk distributions as percentiles in the targetUncalibrated and comparator cohorts after Calibratedstratification. Design HR (95% CI) p HR (95% CI) p min 10% 25% 50% 75% 90% max PS stratification,chlorthalidone on-treatment2 29 0.6529 (0.3391 - 1.14)424 0.17569 5,4180.66 (0.37 - 1.19) 0.18 PS stratification,hydrochlorothiazide intent-to-treat1 29 1.0029 (0.7894 - 1.26)482 0.99673 6,1011.01 (0.79 - 1.30) 0.93

Table 5. Subgroup analyses. We chlorthalidone report hydrochlorothiazide HR estimates, their 95% CIs4. andEvidence uncalibrated will be generated and calibrated using a (cal)pre-specifiedp-values analysis to reject the null hypothesis of no difference in1.000 five pre-specified patient subgroups. design. 5. Evidence will be generated using open source software that is freely available to all. Subjects On-treatment6. Evidence generation process Intent-to-treat will be empirically evaluated 0.999 Subgroup T C HR (95% CI)by includingp cal-p controlHR (95% research CI) questionsp cal- wherep the true eect size is known. 7. Evidence will be generated using best-practices. 0.998 8. LEGEND will not be used to evaluate methods. • ConditionsSurvival probability (condition occurrence in lookback window)9. EvidenceOverhage will be J, updated Ryan P, on Reich a regular C, Hartzema basis. A, Stang P (2012). “Validation of 10. No patient-levelNota common shown: data data will model be shared Calibration for active between safety sites surveillance in the plots research.” Journal – in 3650.997 days prior to index date network,of only the aggregatedAmerican Medical data. Informatics Association, 19, 54–60. – in 30 days prior to index date demonstrateReboussin DM, Allen (and NB, Griswold control ME, Guallar for) onlyE, Hong 0 250 500 Study cohorts. Please see the LEGEND hypertension • Condition aggregation Time in days Study protocolY, (https://github.com/OHDSI/Legend/tree/master/ Lackland DT, Miller EPR, Polonsky T, Thompson-Paul Number at risk Documentsminimal)AM, for complete Vupputuri specification systematic S (2017). of the chlorthalidone, “Systematic residual review for bias the 2017 – SMOMEDchlorthalidone 14,098 10,563 7,804 hydrochlorothiazide 288,073 227,706 179,330 hydrochlorothiazideACC/AHA/AAPA/ABC/ACPM/AGS/APhA/ASH/ASPC/NMA/PCNA and acute myocardial infarction cohorts • Drugs (drug occurrence0 in lookback250 window)500 using ATLASguideline (http://www.ohdsi.org/web/atlas for the prevention, detection,). evaluation, and management Time in days of high blood pressure in adults: a report of the American College of – in 365 days prior to index date Negative controls. We selected negative controls using a pro- Fig. 4. Kaplan Meier plot of acute myocardial infarction-free survival. This plot Cardiology/American Heart Association Task Force on Clinical Practice is adjusted– in for30 the days propensity prior score to stratification; index thedate chlorthalidone curve shows cess similar to that outlined by Voss et al. (2017). We first the actual observed survival. The hydrochlorothiazide curve applies reweighting to construct aGuidelines.” list of all conditionsJournal of that the satisfy American the College following of Cardiology, p. 24428. • Drugapproximate aggregation the counterfactual of what chlorthalidone survival would look like had the criteria withRosenbaum respect to PR, all Rubin drug exposures DB (1983). in “The the LEGENDcentral role of the propensity score chlorthalidone cohort been exposed to hydrochlorothiazide instead. The shaded area hypertensionin study: observational studies for causal effects.” Biometrika, 70(1), 41–55. denotes– Ingredient the 95% CI. Ryan PB, Schuemie MJ, Gruber S, Zorych I, Madigan D (2013). “Empirical – ATC class • No Medline abstract where the MeSH terms suggests a drug-conditionperformance association of a new (Winnenburg user cohortet method: al., 2015), lessons for developing a risk • Risk## Warning: Scores package (Charlson ’bindrcpp’ comorbidity was built index) under R version• 3.4.4No mentionidentification of the drug-condition and analysis pair system.” on a USDrug product Safety, 36(1), 59–72. We exclude all covariates that occur in fewer than 0.1% oflabelSchuemie in the “Adverse M, Ryan Drug P, Reactions” Hripcsak G, or Madigan “Postmarketing” D, Suchard M (2018a). “Improv- section (Duke et al., 2013), patientsResidual within systematic the target error. In and thecomparator absense of bias, cohorts we expect prior to ing reproducibility by using high-throughput observational studies with 95% of negative and positive control estimate 95% confidence • No US spontaneousempirical calibration.” reports suggestingPhilosophical that the Transactions pair is in of the Royal Society A, model fitting for computational eciency. an adverse event relationship (Evans et al., 2001; Banda intervals to include their presumed HR. In the case of negative 376, 20170356. et al., 2016), controls, the presumed HR = 1. Figure 5 describes the negative Schuemie MJ, Hripcsak G, Ryan PB, Madigan D, Suchard MA (2016). “Ro- and positive control estimates under the on-treatment with PS • OMOP vocabulary does not suggest that the drug is References bust empirical calibration of p-values using observational data.” Statistics stratification design. Before calibration, negative and positive indicated for the condition, ALLHATcontrols Officers demonstrate and Coordinators poor coverage. for the After ALLHAT calibration, Collaborative controls Research• Vocabularyin Medicine conditional, 35 concepts(22), 3883–3888. are usable (i.e., not too Groupdemonstrate (2002). “Major acceptable outcomes coverage. in high-risk hypertensive patients ran-broad,Schuemie not suggestive MJ, Hripcsak of an adverse G, Ryan event PB,relationship, Madigan no D, Suchard MA (2018b). domized to angiotensin-converting enzyme inhibitor or calcium channelpregnancy“Empirical related), confidence and interval calibration for population-level effect es- • Exact condition concept itself is used in patient level data. blockerConclusions vs diuretic: The antihypertensive and lipid-lowering treatment to timation studies in observational healthcare data.” Proceedings of the preventWe find heart that attack chlorthalidone trial (ALLHAT).” has a similarJournal risk of of the acute American myocar- MedicalWe optimizeNational remaining Academy condition of Sciences concepts,, such p. 201708282. that parent Associationdial infarction, 288, as 2981–2997. compared to hydrochlorothiazide within the concepts removeSchuemie children MJ, asRyan defined PB, by DuMouchel the OMOP W, vocabulary Suchard MA, Madigan D (2014). Austinpopulation PC (2009). that “Balance the CCAE diagnostics represents. for comparing the distributionand perform“Interpreting manual review observational to exclude any studies: pairs that why may empirical still calibration is needed of baseline covariates between treatment groups in propensity-scorebe in a causalto relationship correct p-values.” or too similarStatistics to the in studymedicine outcome., 33(2), 209–218. For hypertension, this process led to a candidate list of 76 matchedSupporting samples.” InformationStatistics in medicine, 28(25), 3083–3107. Schuemie MJ, Suchard MA, Ryan PB (2018c). CohortMethod: New-user negative controls for which table can be found in study proto- cohort method with large scale propensity and outcome models.R BandaHere JM, Evanswe enumerate L, Vanguri the RS, guiding Tatonetti principles NP, Ryan of LEGEND PB, Shah and NH (2016).col (https://github.com/OHDSI/Legend/tree/master/Documents). “A curatedprovide and linking standardized details on adversestudy cohorts drug eventand design. resource to accelerateIn the comparisonpackage of chlorthalidone version 2.5.1. and hydrochlorothiazide drug safety research.” Scientific data, 3, 160026. in the CCAESchuemie database, MJ, Suchard 76 negative MA, controls Ryan PB, had Reps su Jcient (2018d). FeatureExtraction: DukeLEGEND J, Friedlin principles. J, Li X (2013). “Consistency in the safety labelingoutcomes of toGenerating return estimable Features HRs. for We a Cohort list these. R conditions package version 2.1.5. bioequivalent1. Evidence medications.” will be generatedPharmacoepidemiology at large-scale. and Drug Safetyin Table, 6Suchard M, Simpson S, Zorych I, Ryan P, Madigan D (2013). “Massive (3),2. 294–301.Dissemination of the evidence will not depend on the parallelization of serial inference algorithms for a complex generalized 22 Covariate sets. Evans S, Wallerestimated PC, eDavisects. S (2001). “Use of proportional reporting ratios linear model.” Transactions on Modeling and Computer Simulation, 23, (PRRs)3. Evidence for signalwill generation be generated from spontaneous by consistently adverse applying drug a reaction• Demographics10. (age in 5-year bands, gender, index year, systematic approach across all research questions. index month) reports.” Pharmacoepidemiology and Drug Safety, 10(6), 483–486. Voss EA, Boyce RD, Ryan PB, van der Lei J, Rijnbeek PR, Schuemie MJ Forouzanfar MH, Liu P, Roth GA, Ng M, Biryukov S, Marczak L, Alexander (2017). “Accuracy of an automated knowledge base for identifying drug 4 | Schuemie et al. L, Estep K, Abate KH, Akinyemiju TF, et al. (2017). “Global burden of adverse reactions.” J Biomed Inform, 66, 72–81. hypertension and systolic blood pressure of at least 110 to 115 mm Whelton PK, Carey RM, Aronow WS, Casey DE, Collins KJ, Himmelfarb Hg, 1990-2015.” Journal of the American Medical Association, 317(2), CD, DePalma SM, Gidding S, Jamerson KA, Jones DW, et al. (2018). 165–182. “2017 ACC/AHA/AAPA/ABC/ACPM/AGS/APhA/ASH/ASPC/NMA/PCNA Hripcsak G, Duke J, Shah N, Reich C, Huser V, Schuemie M, Suchard guideline for the prevention, detection, evaluation, and management M, Park R, Wong I, Rijnbeek P, et al. (2015). “Observational Health of high blood pressure in adults: a report of the American College of Data Sciences and Informatics (OHDSI): Opportunities for Observational Cardiology/American Heart Association Task Force on Clinical Practice Researchers.” Studies in health technology and informatics, 216, 574– Guidelines.” Journal of the American College of Cardiology, 71(19), 578. e127–e248.

Schuemie et al. OHDSI version 1.0 | October 9, 2018 | 5 CTD vs. HCTZ: risk of MI

Risk estimates and meta-analysis across LEGEND databases: + CTD

1/4 ●

● ● ● ● ● ● 1/2 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ●●● ● ● ● ● ●● ● ● ● ● ●●● ●● ● ● ●● ● ● ● ● ●● ● ●● ● ● ● ●● ● ● ●● ● ● ●● ● ●●● ● ●● ● ●● 1 ● ● ● ● ● ●● ●● ● ● ● ● ●● ● ● ●● ●● ●●● ● ● ●● ● ● ● ● ●● ● ● ●● ● ●●● ●● ● ● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ●●● ● ● ●● ● ●●● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●

HR ● ● ● ● 2 ● ● ● ● ● 4 ●

CV Safety

+ HCTZ Meta-analysis CCAE Optum Panther MDCR

CTD vs. HCTZ: safety proﬁle

+ CTD 1/4

1/2 ● ● weight gain ● diabetes ● 1 ●● ● ● ● ● ●● ● ● ● ● ● ● ● ● ●● ●● ● ● ● ● ● ● ● HR ● ● ● 2

● ● hypokalemia ● 4 ● CV Safety

+ HCTZ Meta-analysis CCAE Optum Panther MDCR

Safety favors HCTZ – electrolyte imbalance CTD is more potent, longer half-life CTD vs. HCTZ: prediction

DCP-RCT LEGEND CTD-users 6,750 25,566 HCTZ-users 6,750 528,202 avg/median follow-up 3 years 2 years maximum follow-up 4.5 years 16 years HR for MI (anticipated in 2022) 0.81 (0.56 - 1.17) NS

25% of LEGEND patients have follow-up > 4 years

Caveat: design differences (similar vis-a-vis ALLHAT)

DCP LEGEND Clinical lessons for hypertension co − amilozide

hydralazine

minoxidil

guanfacine

methyldopa

chlorthalidone

hydrochlorothiazide

bendroflumethiazide

clonidine indapamide

metolazone terazosin benazepril

LEGEND evidence is concordant with RCTs: prazosin captopril

doxazosin enalapril

fosinopril Aliskiren

lisinopril labetalol

moexipril Where RCT results exist, but many carvedilol perindopril

pindolol quinapril

penbutolol ramipril

unanswered questions remains carteolol trandolapril

acebutolol azilsartan

propranolol candesartan

More outcomes, comparisons, data sources nadolol eprosartan irbesartan nebivolol

losartan metoprolol

olmesartan

bisoprolol telmisartan

st betaxolol valsartan

atenolol Not all 1 -line agents are equivalent: amlodipine

felodipine

spironolactone isradipine

epleronone nicardipine

nifedipine triamterene nisoldipine

amiloride lacidipine diltiazem

BBs, TZDs > ACEIs torsemide

furosemide

Cardiovascularverapamil disease Meta−analysis # bumetanide

Chlorthalidone HCTZ Indapamide Benazepril Captopril Enalapril Fosinopril Combo-therapy initiation: Lisinopril Quinapril Ramipril Azilsartan Candesartan Irbesartan Losartan evidence that combo-therapy is better Olmesartan Telmisartan Valsartan Amlodipine # Felodipine Nicardipine Nifedipine Diltiazem Verapamil Evidence of safety risk Furosemide Spironolactone Atenolol " Bisoprolol Metoprolol Nebivolol Nadolol Propranolol Carvedilol Labetalol Hydralazine Doxazosin DCP trial prediction: Terazosin Clonidine Methyldopa HCTZ Nadolol Atenolol Ramipril Enalapril Losartan Lisinopril Captopril Quinapril Labetalol Nebivolol Diltiazem Clonidine Valsartan Terazosin Fosinopril Bisoprolol Verapamil Nifedipine Azilsartan Carvedilol Irbesartan Felodipine Doxazosin Metoprolol Benazepril Amlodipine Nicardipine Propranolol Methyldopa Telmisartan Olmesartan Indapamide Hydralazine CTD vs. HCTZ – no efﬁcacy difference Furosemide Candesartan Chlorthalidone Spironolactone Acknowledgments co − amilozide

hydralazine

minoxidil

guanfacine

methyldopa

chlorthalidone

hydrochlorothiazide

bendroflumethiazide

clonidine indapamide

metolazone terazosin benazepril

prazosin captopril

doxazosin enalapril

fosinopril Aliskiren

lisinopril LEGEND Scientiﬁc Group: labetalol moexipril carvedilol

perindopril

pindolol quinapril

penbutolol Martijn J. Schuemie David Madigan ramipril

carteolol trandolapril

acebutolol azilsartan Patrick B. Ryan George Hripcsak propranolol candesartan nadolol eprosartan

irbesartan nebivolol

losartan metoprolol

olmesartan Seng Chan You Marc A. Suchard bisoprolol telmisartan

betaxolol valsartan

atenolol amlodipine

felodipine

spironolactone isradipine

epleronone nicardipine

Nicole Pratt nifedipine triamterene nisoldipine

amiloride lacidipine diltiazem

torsemide

furosemide

Cardiovascularverapamil disease Meta−analysis bumetanide

Chlorthalidone HCTZ Indapamide Benazepril Captopril Enalapril Fosinopril Clinical Advisory Team: Lisinopril Quinapril Ramipril Azilsartan Candesartan Irbesartan Losartan Olmesartan RuiJun Chen Christian Reich Telmisartan Valsartan Amlodipine Felodipine Nicardipine Nifedipine Diltiazem Verapamil Furosemide Jon Duke Peter R. Rijnbeck Spironolactone Atenolol Bisoprolol Metoprolol Nebivolol Nadolol Propranolol Carvedilol Rae Woong Park Labetalol Hydralazine Doxazosin Terazosin Clonidine Methyldopa HCTZ Nadolol Atenolol Ramipril Enalapril Losartan Lisinopril Captopril Quinapril Labetalol Nebivolol Diltiazem Clonidine Valsartan Terazosin Fosinopril Bisoprolol Verapamil Nifedipine Azilsartan Carvedilol Irbesartan Felodipine Doxazosin Metoprolol Benazepril Amlodipine Nicardipine Propranolol Methyldopa Telmisartan Olmesartan Indapamide Hydralazine Furosemide Candesartan Chlorthalidone Spironolactone