Prospective Monitoring of New Drugs in Older Adults with and without Frailty: Near-Real-Time Assessment of Effectiveness and Safety of Oral Anticoagulants in Medicare Data

Abstract

Background and Objective:

Prospective sequential analyses after a new drug approval allow proactive surveillance of new drugs. In the current study, we demonstrate feasibility of frailty-specific sequential analyses for dabigatran, rivaroxaban, and apixaban versus warfarin.

Methods:

We partitioned Medicare data from 2011 to 2020 into datasets based on calendar year following the date of drug approval. Each calendar year of data was added sequentially for analysis. We used a new-user, active comparative design by comparing the initiators of dabigatran versus warfarin, rivaroxaban versus warfarin, and apixaban versus warfarin. Patients aged ≥65 years with atrial fibrillation without contraindication to the anticoagulants were included. Claims-based frailty index ≥0.25 was used to define frailty. The initiators of each direct oral anticoagulant were propensity-score matched to the initiators of warfarin within each frailty status. The effectiveness outcome was ischemic stroke or systemic thromboembolism. The safety outcome was major bleeding. For each calendar year, we estimated hazard ratios and 95% confidence intervals for from Cox proportional hazards models using all data available up to that year.

Results:

As an example of the results, in the 2020 dataset, compared to warfarin, apixaban was associated with a reduced risk of the effectiveness outcome (frail: HR 0.73, 95% CI 0.63-0.85; non-frail: HR 0.65, 95% CI 0.59-0.72) and major bleeding (frail: HR 0.63, 95% CI 0.57-0.69; non-frail: HR 0.59, 95% CI 0.56-0.63) in both frail and non-frail patients. We found evidence for apixaban’s effectiveness and safety within 1-2 years after the drug approval in frail older patients.

Conclusion:

Our frailty-specific sequential analyses can be applied to future near-real-time monitoring of newly approved drugs.

1. Introduction

Frailty affects up to 40% of older adults with atrial fibrillation (AF)¹ and is one of the important reasons for non-initiation of oral anticoagulation (OAC).² In our previous study, we found that among Medicare beneficiaries aged 65 years and older, frailty was associated with a 26% reduction in the odds of initiation of OAC after a new AF diagnosis.² Under-treatment of frail older adults with AF reflects the fear of OAC-related major bleeding in patients with advanced age and multimorbidity.³ Despite the high prevalence of frailty in AF and frail older adults having excess risks for both major bleeding and ischemic stroke,^4,5 frailty was not formally assessed in the premarket randomized controlled trials for direct oral anticoagulants (DOACs).^6-8 A post-hoc frailty subgroup analysis of the Effective Anticoagulation with Factor Xa Next Generation in Atrial Fibrillation-Thrombolysis in Myocardial Infarction 48 (ENGAGE AF–TIMI 48) trial used a frailty score generated post hoc using the variables available within the trial and was underpowered to determine heterogeneous treatment effect by frailty levels.⁴ To address the limitations of clinical trials, we have previously developed and validated a claims-based frailty index (CFI) using the United States (US) Medicare data.^9-11 CFI allows measurement of frailty in large older adult populations in routine care settings and facilitates comparative effectiveness and safety studies in older adults with and without frailty.

The US Food and Drug Administration (FDA) Sentinel Initiative proactively monitors safety of newly approved drugs by facilitating analysis of insurance claims and electronic health records (EHR). Prospective sequential analyses, in which analysis is refreshed at regular intervals as patient-level data are continuously generated from claims and EHR after drug approval, allows near-real-time surveillance of effectiveness and safety of new drugs.^12-15 Prior case studies involving rofecoxib, prasugrel, and dabigatran demonstrated potential utility of prospective sequential analyses for early detection of effectiveness and safety signals.¹⁵ Yet the focus of these studies was feasibility of conducting sequential analyses in the general population and not in frail older populations. Frailty-specific prospective monitoring enabled by incorporation of a database-specific frailty index would allow timely generation of data on effectiveness and safety of newly approved drugs in older adults with frailty.

In the current study, we demonstrate feasibility of frailty-specific sequential analyses for dabigatran, rivaroxaban, and apixaban versus warfarin using Medicare data and CFI. Our case study illustrates what frailty-specific prospective monitoring may have achieved had it been in place at the time of DOAC approval for AF.

2. Methods

2.1. Sequential Analyses Framework

Using Medicare fee-for-service claims data from 2011 to 2020, we emulated prospective monitoring of dabigatran (approved October 19, 2010), rivaroxaban (approved November 4, 2011), and apixaban (approved December 28, 2012) versus warfarin beginning at the Food and Drug Administration approval of each DOAC. To emulate prospective monitoring, we partitioned Medicare data into datasets based on calendar year following the date of drug approval (Figure 1). Each calendar year of data is added sequentially for analysis given Medicare data are released each year. We used a new-user, active comparator design by comparing initiators of dabigatran versus warfarin (hereafter referred to as “dabigatran cohort”), initiators of rivaroxaban versus warfarin (hereafter referred to as “rivaroxaban cohort”), and initiators of apixaban versus warfarin (hereafter referred to as “apixaban cohort”) in the same calendar year. As an example, the 2015 dabigatran cohort included patients who initiated dabigatran or warfarin in 2011 through 2015. The 2016 dabigatran cohort included patients who initiated dabigatran or warfarin in 2016 in addition to those included in the 2015 cohort. When a patient was eligible for the cohort (see the study population section below) more than once, we only kept the first treatment episode.

Figure 1.

Schematic diagram for creation of sequential datasets to emulate prospective monitoring. (A) Year 1 dataset includes initiators of study drugs in the first year following drug approval; (B) Year 2 dataset includes initiators of study drugs in the first and second years following drug approval; (C) Year 3 dataset includes initiators of study drugs in the first, second, and third years following drug approval

2.2. Study Population

The index date was the date of drug initiation. The assessment period for inclusion and exclusion criteria was the prior 365 days including the index date. Patients were included if they were age 65 years or older, had at least 1 inpatient or 2 outpatient diagnoses for AF, had no exposure to a DOAC or warfarin in the previous 365 days, and were continuously enrolled in Medicare Parts A, B, and D in the previous 365 days. Patients were excluded if they were missing data on age, sex, or race or had contraindications to a DOAC or warfarin (i.e., mechanical valves, rheumatic heart disease, mitral stenosis, end-stage renal disease, history of intracranial hemorrhage), or had another indication for requiring anticoagulation (i.e., venous thromboembolism or hip or knee replacement) (Supplemental Figure S1).

2.3. Frailty and Other Baseline Patient Characteristics

The assessment period for frailty and other baseline characteristics was the prior 365 days including the index date. Within each sequential dataset, patients were stratified by frail and non-frail using CFI ≥0.25 as the cut-off to define frailty.^9,16 CFI is a deficit-accumulation index that uses 93 variables defined by the International Classification of Diseases (ICD), Current Procedural Terminology, and Healthcare Common Procedure Coding System codes in Medicare claims data.⁹ CFI uses has been validated against clinical frailty measures, ¹⁰ physical performance metrics, ¹⁷ disability level, ^10,17 and risk of institutionalization.¹⁷ The CFI cutpoint of 0.25 has 62% sensitivity and 78% specificity for the Fried Physical Frailty Phenotype and 60% sensitivity and 86% specificity for the deficit accumulation frailty index 0.25 or higher.¹⁸ We measured dual eligibility for Medicare and Medicaid (health insurance for people with low income), comorbidities, outpatient prescription drug use, and healthcare use (e.g., hospital days, number of emergency room visits). We calculated the CHA₂DS₂-VASc score to estimate the risk of thromboembolic stroke and the modified HAS-BLED score (excluded labile international normalized ratio, which is not available in administrative claims) to estimate bleeding risk using claims-based algorithms. We calculated a combined comorbidity score to quantify overall comorbidity burden.¹⁹ The ICD codes used in the study have been previously published.⁵

2.4. Outcomes and Follow-up

The effectiveness outcome was ischemic stroke or systemic thromboembolism (SEE), and the safety outcome was major bleeding. Ischemic stroke or SEE was defined using the diagnosis codes in the primary position of the inpatient discharge diagnosis. Major bleeding was defined using the diagnosis codes for bleeding in any critical site (e.g., intracranial, intraocular, intraspinal) in the primary position of the inpatient discharge diagnosis or the diagnosis codes for any bleeding in the primary position of the inpatient discharge diagnosis combined with procedure codes for transfusion. These claims-based algorithms have positive predictive values of 96% for ischemic stroke,²⁰ 87% for SEE,²¹ and 86-96% for major bleeding.²²

The follow-up began on the day after the index date until the occurrence of the study outcomes, death, disenrollment from Medicare part A and B, discontinuation of the index drug, or switching to another OAC. Discontinuation was defined as a treatment gap (“grace period”) of more than 60 days for warfarin and 14 days for DOAC. Patients were considered exposed to the index drug (“exposure risk window”) until 30 days and 14 days after the discontinuation of the drug for warfarin and DOAC, respectively. Setting different grace periods and exposure risk windows for a DOAC and warfarin is necessary because extended warfarin prescriptions are often provided to accommodate dosing adjustments based on international normalized ratio (INR).²³

2.5. Statistical Analysis

For each DOAC cohort, propensity score (PS) estimation and 1:1 matching between DOAC initiators and warfarin initiators were performed within each frailty status (frail and non-frail) and the calendar year of the index date. We estimated PS using logistic regression that included 96 variables (age, sex, race, year of drug initiation, insurance status, 37 comorbidities, 38 drug classes, 17 measures of healthcare utilization) and implemented 1:1 nearest neighbor matching without replacement with a caliper of 0.2 of the standard deviation of the logit of PS.²⁴ PS-matched pairs from previous years were kept in the new year. As an example, the 2015 dabigatran cohort included newly PS-matched pairs of dabigatran initiators and warfarin initiators in 2015 as well as PS-matched pairs from 2011, 2012, 2013, and 2014. The adequacy of PS matching was evaluated based on a standardized mean difference of 0.10 or less. For each calendar year, we estimated hazard ratios (HRs) and 95% confidence intervals (CIs) for each outcome from Cox proportional hazards models using all data available up to that year. Additionally, using the 2020 dataset, we tested heterogeneity of treatment effects between the frail and non-frail groups.²⁵ As a sensitivity analysis, we conducted a intention-to-treat (ITT) analyses in which patients were not censored for discontinuation of the index drug or switching to another OAC and followed until 365 days after the index date. Analyses were conducted in the Aetion Evidence Platform (including R version 4.1.1 [R Foundation for Statistical Computing])^26,27 and in SAS version 9.4 (SAS Institute Inc.). A 2-sided p-value less than 0.05 was considered statistically significant.

3. Results

3.1. Baseline Patient Characteristics

Table 1 (selected characteristics) and Supplemental Tables S1-S3 (complete list of characteristics) describe the baseline characteristics of the 2020 PS-matched dataset for each DOAC-vs-warfarin comparison stratified by frailty status. At baseline, compared to non-frail patients, frail patients were older, more likely to be women, and more likely to be dually eligible for Medicare and Medicaid; had higher combined comorbidity score, CHA₂DS₂-VASc score, and HAS-BLED score; had higher prevalence of both cardiovascular and non-cardiovascular comorbidities including stroke and dementia; had greater use of cardiovascular and non-cardiovascular drugs; and had greater health care utilization and greater healthcare costs. The characteristics were well balanced between DOAC and warfarin initiators.

Table 1.

Summary Characteristics of 1:1 Propensity Score-Matched Patients in the 2020 Dataset with Atrial Fibrillation Treated with Dabigatran versus Warfarin (2011-2020), Rivaroxaban versus Warfarin (2012-2020), and Apixaban versus Warfarin (2013-2020), Stratified by Frailty Status

	Sample size	Mean age (SD), y	Women (%)	Race (%)			Dual status (%)	Mean CCS (SD)	CHA2DS2-VASc (SD)	HAS-BLEDa (SD)	Mean frailty index (SD)	Stroke (%)	Dementia (%)
				Black	Other	White
Frail
Dabigatran	13 511	78.9 (7.5)	63.6	5.6	5.3	89.1	15.5	5.26 (2.77)	5.39 (1.46)	2.72 (0.71)	0.32 (0.06)	21.2	32.3
Warfarin	13 511	78.9 (7.5)	62.9	5.2	5.6	89.2	15.7	5.25 (2.71)	5.39 (1.43)	2.72 (0.70)	0.32 (0.06)	21.4	32.5
Non-frail
Dabigatran	54 205	76.6 (6.9)	47.7	3.3	5.2	91.6	3.6	2.19 (2.05)	3.81 (1.27)	2.14 (0.61)	0.17 (0.04)	6.4	2.2
Warfarin	54 205	76.4 (6.8)	47.6	3.4	5.3	91.4	3.6	2.19 (2.02)	3.81 (1.24)	2.14 (0.61)	0.17 (0.04)	6.4	2.2
Frail
Rivaroxaban	31 598	79.0 (7.7)	61.6	5.9	4.6	89.5	14.6	5.58 (2.79)	5.34 (1.44)	2.72 (0.71)	0.31 (0.06)	18.7	34.2
Warfarin	31 598	79.0 (7.7)	62.1	6.0	4.6	89.4	14.6	5.56 (2.80)	5.34 (1.43)	2.72 (0.71)	0.31 (0.06)	18.8	34.4
Non-frail
Rivaroxaban	116 932	77.1 (7.0)	47.3	3.3	4.6	92.1	3.2	2.37 (2.15)	3.85 (1.27)	2.14 (0.61)	0.17 (0.04)	5.7	2.4
Warfarin	116 932	77.0 (6.9)	47.6	3.4	4.7	91.9	3.2	2.35 (2.14)	3.84 (1.24)	2.15 (0.62)	0.17 (0.04)	5.7	2.4
Frail
Apixaban	30 960	79.5 (7.7)	59.1	5.4	3.7	90.9	13.7	5.81 (2.81)	5.42 (1.44)	2.76 (0.71)	0.31 (0.06)	20.4	33.1
Warfarin	30 960	79.5 (7.7)	59.5	5.6	3.7	90.8	13.6	5.79 (2.82)	5.41 (1.43)	2.76 (0.71)	0.31 (0.06)	20.3	33.3
Non-frail
Apixaban	112 434	77.8 (7.3)	46.8	3.3	4.0	92.7	3.1	2.55 (2.25)	3.93 (1.30)	2.16 (0.63)	0.18 (0.04)	6.2	2.6
Warfarin	112 434	77.8 (7.2)	47.5	3.3	4.1	92.7	3.1	2.50 (2.22)	3.92 (1.26)	2.16 (0.64)	0.18 (0.04)	6.1	2.5

CHA₂DS₂-VASc 1 point for congestive heart failure, 1 point for hypertension, 2 points for 75 years or older, 1 point for age 65-74, 1 point for diabetes, 2 points for history of stroke or transient ischemic attack, or systemic thromboembolism, 1 point for vascular disease including myocardial infarction or peripheral arterial disease, and 1 point for female sex, CCS combined comorbidity score, HAS-BLED 1 point each for hypertension, renal disease, liver disease, prior stroke, prior history of bleeding, age greater than 65, use of aspirin and other antiplatelets and alcohol use disorder, SD standard deviation.

^aExcluded labile international normalized ratio component

3.2. Sequential analyses DOACs versus Warfarin for Ischemic Stroke or SEE

The mean (standard deviation) follow-up times for ischemic stroke or SEE in the 2020 cohort were as follows: 237 (360) days for dabigatran and 532 (640) days for warfarin in the dabigatran cohort; 275 (376) days for rivaroxaban and 485 (546) days for warfarin in the rivaroxaban cohort; and 293 (376) days for apixaban and 459 (496) days for warfarin in the apixaban cohort. The most common reason for censoring was discontinuation of the index drug (Supplemental Table S4). The number of patients, number of ischemic stroke or SEE events, and follow-up times for each sequential dataset are reported in Supplemental Tables S5-S7.

In the 2020 datasets with the frail and non-frail populations combined, the overall treatment effects on the risk of ischemic stroke or SEE for dabigatran, rivaroxaban, and apixaban versus warfarin were HR 0.78 (0.67-0.90), HR 0.79 (95% CI 0.73-0.86), and HR 0.67 (0.62-0.73), respectively. The treatment effects by frailty status for dabigatran were HR 0.72 (0.54-0.95) for frail and 0.81 (0.68-0.95) for non-frail (P_{heterogeneity}=0.47); for rivaroxaban the treatment effects were HR 0.88 (0.75-1.04) for frail and HR 0.76 (0.69-0.84) for non-frail (P_{heterogeneity}=0.14); for apixaban the treatment effects were HR 0.73 (0.63-0.85) for frail and HR 0.65 (0.59-0.72) for non-frail (P_{heterogeneity}=0.22).

For the dabigatran cohort with frailty, the HR remained stable throughout the study period from 2011 to 2020 (HR 0.72, 95% CI 0.54-0.95 in 2020) (Figure 2). The 95% CI precluded the null for the first time in 2017 (6 years after FDA approval). For the rivaroxaban cohort with frailty, the HR settled to be similar to the 2020 estimate (HR 0.88, 95% CI 0.75-1.04) in 2015 (3 years after FDA approval). The 95% CI precluded the null only in one year, 2017. For the apixaban cohort with frailty, the 95% CI precluded the null the first time in 2015 (2 years after the FDA approval), and the HR settled to be similar to the 2020 estimate (HR 0.73, 95% CI 0.63-0.85) in 2016.

Figure 2.

Prospective Monitoring of (A) Dabigatran, (B) Rivaroxaban, and (C) Apixaban for Ischemic Stroke and Systemic Thromboembolism Stratified by Frailty Status

For the dabigatran cohort without frailty, the HR remained stable throughout the study period from 2011 to 2020 (HR 0.81, 95% CI 0.68-0.95 in 2020). The 95% CI precluded the null for the first time in 2016 (5 years after FDA approval). For the rivaroxaban cohort without frailty, the HR settled to be similar to the 2020 estimate (HR 0.76, 95% CI 0.69-0.84) in 2016 (4 years after FDA approval). The 95% CI precluded the null throughout the study period. For the apixaban cohort without frailty, the HR settled to be similar to the 2020 estimate (HR 0.65, 95% CI 0.59-0.72) in 2016 (3 years after FDA approval). The 95% CI precluded the null for the first time in 2014.

3.3. Sequential analyses of DOACs versus Warfarin for Major Bleeding

The mean (standard deviation) follow-up times for major bleeding in the 2020 cumulative cohort were as follows: 237 (359) days for dabigatran and 528 (647) days for warfarin in the dabigatran cohort; 274 (375) days for rivaroxaban and 482 (544) days for warfarin in the rivaroxaban cohort; and 293 (376) days for apixaban and 457 (494) days for warfarin in the apixaban cohort. The most common reason for censoring was discontinuation of the index drug (Supplemental Table S4). The number of patients, number of major bleeding events, and follow-up times for each sequential dataset are reported in Supplemental Tables S8-S10.

In the 2020 datasets with the frail and non-frail populations combined, the overall treatment effects on the risk of major bleeding for dabigatran, rivaroxaban, and apixaban versus warfarin were HR 0.86 (0.80-0.93), HR 1.05 (95% CI 1.01-1.10), and HR 0.60 (0.57-0.64), respectively. The treatment effects by frailty status for dabigatran were HR 0.97 (0.86-1.11) for frail and 0.82 (0.75-0.89) for non-frail (P_{heterogeneity}=0.037); for rivaroxaban the treatment effects were HR 1.00 (0.91-1.10) for frail and HR 1.10 (1.03-1.18) for non-frail (P_{heterogeneity}=0.42); for apixaban the treatment effects were HR 0.62 (0.56-0.68) for frail and HR 0.59 (0.55-0.63) for non-frail (P_{heterogeneity}=0.25).

For the dabigatran cohort with frailty, the HR settled to be similar to the 2020 estimate (HR 0.97, 95% CI 0.86-1.11) in 2014 (3 years after FDA approval) (Figure 3). The lower limit of 95% CI crossed the null in 2012. For the rivaroxaban cohort with frailty, the HR remained stable around the null throughout the study period from 2012 to 2020 (HR 1.03, 95% CI 0.95-1.11). For the apixaban cohort, the HR remained stable throughout the study period from 2013 to 2020 (HR 0.63, 95% CI 0.57-0.69 in 2020). The 95% CI precluded the null for the first time in 2014 (1 year after FDA approval).

Figure 3.

Prospective Monitoring of (A) Dabigatran, (B) Rivaroxaban, and (C) Apixaban for Major Bleeding Stratified by Frailty Status

For the dabigatran cohort without frailty, the HR settled to be similar to the 2020 estimate (HR 0.82, 95% CI 0.75-0.89) and the 95% CI precluded the null for the first time in 2014 (3 years after FDA approval). For the rivaroxaban cohort without frailty, the HR and the 95% CI remained above the null throughout the study period from 2012 to 2020 (HR 1.07, 95% CI 1.02-1.13). For the apixaban cohort without frailty, the HR settled to be similar to the 2020 estimate (HR 0.59, 95% CI 0.56-0.63) in 2015 (2 years after FDA approval). The 95% CI precluded the null throughout the study period.

3.4. Sensitivity analysis - ITT analyses

For ischemic stroke or SEE, the 2020 point estimates from the ITT analyses for the dabigatran and apixaban cohorts were qualitatively consistent with the 2020 point estimates from the as-treated analyses (Supplemental Table S11). For the rivaroxaban cohorts, there was a trend toward an increased risk of ischemic stroke or SEE with rivaroxaban compared to warfarin in the ITT-analyses. For major bleeding, the 2020 point estimates from the ITT analyses for the dabigatran, rivaroxaban, and apixaban cohorts were qualitatively consistent with the 2020 point estimates from the as-treated analyses.

4. Discussion

In the current study, we demonstrate feasibility of frailty-specific sequential analyses for dabigatran, rivaroxaban, and apixaban versus warfarin using Medicare data and CFI. Our case study demonstrates what frailty-specific prospective monitoring may have achieved had it been in placed at the time of DOAC approval for AF. For apixaban, which is now the most commonly prescribed OAC in the United States,² we found evidence for the drug’s effectiveness within 2 years after the drug approval and evidence for the drug’s safety within 1 year after the drug approval in frail older adults. If prospective monitoring had been in place real-time, we would have been able to generate timely evidence on the benefits of apixaban in frail older adults from routine practice, and the early safety data may have led to a greater initial uptake of the drug among frail older adults. Both the prescribers and patients may have felt more confident about the applicability of the superior safety of apixaban compared to warfarin demonstrated in the landmark trial. ⁶The FDA’s Sentinel Initiative does not study how treatment effectiveness and safety can be influenced by a patient’s frailty status, which increases the risk of treatment-related adverse events including the risk of bleeding from OAC.

A typical pharmacoepidemiologic study is conducted when an investigator writes a study protocol to evaluate the effectiveness and safety of a drug after some time the drug has been in the market. The detection of the effectiveness or safety signals in routine care population depends on when the analysis is conducted, which could be several years after the drug approval. The purpose of our prospective monitoring is to prospectively design a pharmacoepidemiologic study and conduct analysis repeatedly as new data become available so that we can identify the signals of effectiveness and safety as early as we can. Initially, the number of outcome events is low because the number of new DOAC users is low, which is expected following the approval of a new drug; it takes time for the drug to be adopted into clinical practice. As more people begin to use the drug over longer periods, pharmacoepidemiologic studies like ours can generate more precise and stable effect estimates with narrower confidence intervals.

Feasibility and utility of prospective monitoring of newly approved drugs in frail older adults is dependent on several factors. First, there cannot be a long lag period between when data become available and when new drugs are released. Medicare data are released once a year and are available typically 2 years later. Such lag time is shorter when Medicare data are accessed through the CMS Virtual Research Data Center or for EHR data and commercial insurance claims data. Second, a large amount of data for older adults with various degrees of frailty are needed. EHR database from Veterans Health Administration are the only other national data for older adults in the US, but only 2% of its patients are women.²⁸ Claims data from countries other than the US—Denmark, Korea, Taiwan, for example—may be used. Third, since frailty is not measured directly in real world data, a validated frailty index has to be available for the database being analyzed,²⁹ similar to CFI in Medicare data.

Our prospective monitoring program for DOACs can be expanded in the future to include factor XIa inhibitors, which are currently being tested against apixaban³⁰ or placebo³¹ in phase 3 trials and against placebo in a phase 2 trial.³² The factor XIa inhibitors have been shown to reduce risk of bleeding compared to apixaban or rivaroxaban in phase 2 trials.³³ If proven to be as effective as the DOACs for prevention of ischemic stroke or SEE, this novel class of drugs will likely address the clinical need of frail older adults who cannot tolerate long-term anticoagulation due to unacceptable bleeding risk despite having higher risk of ischemic stroke than non-frail patients. The clinical trials will likely be underpowered to determine whether there is heterogeneity in treatment effects by frailty status. Our prospective monitoring program is feasible and promises to generate timely evidence for frail older adults if and when the drugs are approved.

Our study has several important limitations. First, in this claims-based study, we were unable to measure important clinical variables (e.g., body mass index, creatinine level, aspirin use, INR) that can affect the effectiveness and safety of DOACs or warfarin. Although we included nearly 100 variables in the PS models, imbalance in these unmeasured variables could have led to residual confounding. Second, we allowed a longer grace period (i.e., the gap between prescriptions) to define discontinuation of warfarin (60 days) compared to discontinuation of DOAC (14 days). Although this was necessary to reflect the provision of extended warfarin prescriptions to accommodate dosing adjustments based on INR,²³ this resulted in much longer follow-up period for the warfarin group than the DOAC group in the as-treated analysis. The ITT analyses, in which patients were followed for 1 year regardless of the treatment discontinuation, suggest that some of our results (rivaroxaban cohort) were sensitive to the length of the grace period. Third, because the mean follow-up period in our study was 237 days for dabigatran, 274-275 days for rivaroxaban, and 293 days for apixaban, we were unable to determine long-term effectiveness and safety of the study drugs. Fourth, because our goal was estimation of the treatment effect rather than hypothesis testing, we did not have a specific alternative hypothesis (a threshold for safety or effectiveness), precluding the group sequential test procedure. Fifth, we did not study the potential impact of OAC use during hospitalizations on the study outcomes.

In conclusion, in our study emulating prospective monitoring of DOACs in older adults, we found evidence for apixaban’s effectiveness and safety within 1-2 years of its approval in frail older adults. Our novel approach that incorporates frailty and prospective monitoring can be applied to future claims-based comparative effectiveness and safety studies of other newly approved cardiovascular drugs in older adults.

KEY POINTS.

• Frailty is prevalent in older adults with atrial fibrillation and is an important risk factor for under-treatment with anticoagulation.

• Frailty-specific prospective monitoring of newly approved drugs enables proactive surveillance of the drugs’ effectiveness and safety in frail older adults.

• The case studies using dabigatran, rivaroxaban, and apixaban versus warfarin comparisons demonstrate feasibility of frailty-specific prospective monitoring using Medicare data and claims-based frailty index. Had the system in place at the time of apixaban approval, we would have had evidence for its effectiveness and safety within 1-2 year of apixaban’s approval.

• Frailty-specific prospective monitoring can be applied to generate evidence near-real-time for newly approved cardiovascular drugs in the future