Abstract
Background
Premature discontinuation of P2Y12 inhibitor therapy has been associated with adverse cardiac events, which might be preventable by improving medication persistence. Current risk models have limited ability to predict patients at risk of P2Y12 inhibitor nonpersistence.
Methods and Results
ARTEMIS (Affordability and Real‐World Antiplatelet Treatment Effectiveness after Myocardial Infarction Study) was a randomized, controlled trial testing the impact of a copayment assistance intervention on P2Y12 inhibitor persistence and outcomes. Among 6212 patients post myocardial infarction with a planned 1‐year course of P2Y12 inhibitor therapy, nonpersistence was defined as a gap in P2Y12 inhibitor filled >30 days by pharmacy fill data. We developed a predictive model for 1‐year P2Y12 inhibitor nonpersistence among patients randomized to usual care. P2Y12 inhibitor nonpersistence rates were 23.8% (95% CI, 22.7%–24.8%) at 30 days and 47.9% (46.6%–49.1%) at 1 year; the majority of these patients had in‐hospital percutaneous coronary intervention. Patients who received the copayment assistance intervention had nonpersistence rates of 22.0% (20.7%–23.3%) at 30 days and 45.3% (43.8%–46.9%) at 1 year. A 53‐variable multivariable model predicting 1‐year persistence had a C‐index of 0.63 (optimism‐corrected C‐index 0.58). Model discrimination did not improve with inclusion of patient‐reported perceptions about disease, medication‐taking beliefs, and prior medication‐filling behavior in addition to demographic and medical history data (C‐index 0.62).
Conclusions
Despite addition of patient‐reported variables, models predicting persistence with P2Y12 inhibitor therapy performed poorly, thereby suggesting the need for continued patient and clinician education on the importance of P2Y12 inhibitor therapy after acute myocardial infarction.
Registration
URL: https://www.clinicaltrials.gov; Unique identifier: NCT02406677.
Keywords: medical adherence, P2Y12 inhibitor therapy, post‐myocardial infarction
Subject Categories: Quality and Outcomes, Health Services
Nonstandard Abbreviations and Acronyms
- ARTEMIS
Affordability and Real‐World Antiplatelet Treatment Effectiveness After Myocardial Infarction Study
Clinical Perspective.
What Is New?
This study examines whether a predictive model can be developed to predict 1‐year nonpersistence to oral P2Y12 inhibitors using patient‐reported perceptions about disease, medication‐taking beliefs, and prior medication‐filling behavior in addition to demographic and medical history data.
What Are the Clinical Implications?
Further analyses are needed to understand what factors/variables are associated with well‐performing predictive models of nonpersistence.
In the United States, inappropriate medication use (ie, nonpersistence, nonadherence, premature discontinuation, or not taken as prescribed) accounts for half of medication‐related hospital admissions and several billions of dollars in potentially avoidable annual health care spending. 1 Among patients hospitalized with acute myocardial infarction (MI), less than half of patients persist with prescribed evidence‐based secondary prevention medications by 1 year, despite compelling evidence supporting these life‐saving treatments. 2 , 3 Persistence with P2Y12 inhibitor therapy following percutaneous coronary intervention is critical to prevent stent thrombosis and recurrent myocardial ischemia, yet many patients do not complete the 1‐year course of P2Y12 inhibitor therapy post MI currently recommended by guidelines. 4
Identifying patients who are less likely to adhere to their medications could allow for targeted interventions; however, prior attempts 3 , 5 , 6 to accurately predict medication adherence have been challenging, partially because they are limited to routinely collected health care data. The ARTEMIS (Affordability and Real‐World Antiplatelet Treatment Effectiveness After Myocardial Infarction Study) trial was a multicenter, cluster‐randomized clinical trial that randomized 301 hospitals to usual care or to the copayment assistance intervention, where patients were able to offset copayment costs when filling ticagrelor or clopidogrel prescriptions in the year post MI. The intervention was a voucher to cover the complete cost of clopidogrel or ticagrelor. At the time of each medication fill or refill, the copayment was charged to the study, resulting in zero out‐of‐pocket cost for the patient. In addition to demographic, medical history, and clinical data typically collected for patients post MI, ARTEMIS uniquely collected patient‐reported perceptions about disease, medication‐taking beliefs, and prior medication‐filling behavior that might influence posthospitalization medication persistence, as well as reasons for treatment change among those who reported P2Y12 inhibitor interruption or discontinuation. We leveraged the ARTEMIS data set in an attempt to optimize our prediction of patients at high risk of nonpersistence with guideline‐recommended P2Y12 inhibitor therapy.
METHODS
Data Source and Patient Population
The design and primary results of the ARTEMIS trial have been previously published. 4 , 7 Briefly, ARTEMIS enrolled 11 001 patients from June 2015 to September 2016. Eligible patients were ≥18 years old, diagnosed with ST‐segment–elevation MI or non–ST‐segment–elevation MI, discharged on a P2Y12 inhibitor (clopidogrel, prasugrel, or ticagrelor) selected at the treating physicians' preference, had US‐based health insurance coverage with a prescription drug benefit, and were able to provide written informed consent for longitudinal follow‐up. Patients were excluded from enrollment in the trial who had a prior intracranial hemorrhage, contraindications to P2Y12 inhibitor therapy at discharge, enrollment in another research study that specified the type and duration of P2Y12 inhibitor in the 1 year post MI, life expectancy <1 year, or plans to move outside of the United States in the next year. The co‐primary end points of the trial were 1‐year P2Y12 inhibitor persistence and major adverse cardiac event (composite of death, recurrent MI, or stroke). All patients enrolled in ARTEMIS provided written informed consent, and the study protocol was approved by the institutional review board of each participating site. The Duke University Medical Center Institutional Review Board approved use of ARTEMIS data for this analysis. The authors are solely responsible for the design and conduct of the study, all study analyses, drafting and editing of the article, and its final contents. The authors declare that all supporting data are available within the article.
Patient‐reported persistence was validated using pharmacy fill data from Symphony Health Solutions, which captures pharmacy claims data from ≈90% of retail, 60% of mail order, and 70% of specialty pharmacies in the United States. Of the 11 001 patients at 287 sites originally enrolled in the ARTEMIS trial from June 2015 through September 2016, 25 patients died during their index hospitalization, 1 patient withdrew from the study during the index hospitalization, 4 were not discharged on a P2Y12 inhibitor, and 3419 had a planned P2Y12 inhibitor duration <1 year. An additional 1341 patients were excluded because they had no pharmacy medication fills for any medication class in the Symphony pharmacy data within 1 year, yielding a final analysis population of 6212 patients with a planned course of 1‐year P2Y12 inhibitor therapy with pharmacy fill data available to assess medication persistence.
Statistical Analysis
Nonpersistence was defined as a gap in P2Y12 inhibitor filled >30 days by pharmacy fill data. We stratified patients by 1‐year persistence versus nonpersistence and described characteristics of each group using proportions for categorical variables and medians with 25th and 75th quartiles for continuous variables. Differences between voucher use groups were tested using χ2 tests for categorical variables and Wilcoxon rank sum tests for continuous variables. Kaplan–Meier rates of nonpersistence were calculated at 3 months, 6 months, and 1 year post discharge. Among patients who were nonpersistent based on pharmacy fills, we examined patient‐reported reason for P2Y12 inhibitor discontinuation (answer options shown in Table 1). Time to first 30‐day gap in fill was reported for each patient‐reported reason.
Table 1.
Patient‐Reported Reasons for Nonpersistence
| Proportion of nonpersistence N (%) | Time to nonpersistence (days)* | |
|---|---|---|
| Patient‐reported persistence | 1215 (41.5%) | |
| Doctor told me I no longer needed to be on this medication | 536 (18.3) | 36 (1, 150) |
| Bleeding/bruising, with physician guidance | 83 (2.8) | 62 (1, 210) |
| Bleeding/bruising, without physician guidance | 28 (1.0) | 35 (1, 173) |
| Other medication side effects | 128 (4.4) | 62 (1, 132) |
| Procedure/surgery that required stopping | 153 (5.2) | 30 (1, 108) |
| Cost too much | 164 (5.6) | 30 (1, 107) |
| Ran out of pills and did not refill | 77 (2.6) | 1 (1, 91) |
| I felt it wasn't helping me | 30 (1.0) | 74 (28, 198) |
| Other/unknown | 516 (17.6) | 31 (1, 143) |
Expressed as median (25th, 75th percentiles).
We developed the predictive model for 1‐year P2Y12 inhibitor nonpersistence among patients randomized to usual care; restricting to these patients allowed us to assess nonpersistence in a population not confounded by presence of a copayment voucher that offset cost to fill P2Y12 inhibitor therapy. All 53 variables specified in the primary outcome model were included in this model (full list in Table 2). To estimate internal validity, we used 500 bootstrap samples drawn with replacement from the sample of usual care arm patients. A logistic regression model for nonpersistence was estimated in the original sample of usual care arm patients and in each bootstrap sample. Estimates of a model's predictive ability may be overly optimistic when estimated on the same data used to train the model. One simple technique to correct for this optimism is split‐sample methodology where the data are split into training and test samples. The model is developed on the training sample and predictive ability assessed on the test data. Split sample methodology is commonly used, but inefficient as it uses only a subset of the data to develop the model. An estimate of optimism was calculated as the average of the differences between the apparent performance in the bootstrap sample and the test performance in the original sample. 8 The predictive performance of the model was assessed using measures for discrimination and calibration. Discrimination was assessed using the naïve C‐statistic calculated from the original sample and the optimism‐corrected C‐statistic. We examined a locally estimated scatterplot smoothing‐based calibration curve as described by Austin and Steyerberg in 2013 based on the model estimated on the original data. 9 We also calculated optimism‐corrected calibration‐in‐the‐large (the intercept) and calibration slope as a measure of spread.
Table 2.
Parameter Estimates for the Predictive Model of Nonpersistence to P2Y12 Inhibitor Use in the Usual Care Arm, Ranked by Chi‐Square Value
| Variable | OR (95% CI) | Wald chi‐square | P value |
|---|---|---|---|
| Male sex | 1.577 (1.251–1.988) | 14.8802 | 0.0001* |
| Hospital: government vs nongovernment, not‐for‐profit | 1.612 (1.198–2.169) | 9.9479 | 0.0016* |
| Financial hardship of paying for meds: moderate/extreme vs none | 1.541 (1.177–2.017) | 9.8912 | 0.0017* |
| Missed >1 dose of medication in past month | 1.262 (1.027–1.550) | 4.8921 | 0.027* |
| Hospital: teaching status | 0.781 (0.627–0.974) | 4.8305 | 0.028* |
| Drug‐eluting stent implanted | 0.695 (0.493–0.979) | 4.3227 | 0.0376* |
| Femoral access site | 1.199 (0.991–1.450) | 3.4866 | 0.0619* |
| Insurance: private vs nonprivate | 0.823 (0.666–1.017) | 3.2476 | 0.0715* |
| Hispanic vs non‐Hispanic | 0.679 (0.427–1.082) | 2.6466 | 0.1038 |
| Creatinine clearance per 5 increase | 0.984 (0.965–1.003) | 2.6409 | 0.1041 |
| Married | 0.850 (0.698–1.037) | 2.5679 | 0.109 |
| Culprit only PCI vs no PCI | 1.471 (0.907–2.387) | 2.4502 | 0.1175 |
| Prior PCI | 1.246 (0.929–1.672) | 2.158 | 0.1418 |
| Cardiogenic shock | 1.549 (0.843–2.846) | 1.9854 | 0.1588 |
| ST‐segment–elevation myocardial infarction | 0.877 (0.729–1.054) | 1.9586 | 0.1617 |
| Hospital: nongovernment, investor owned, for profit vs nongovernment, not for profit | 1.327 (0.889–1.980) | 1.9218 | 0.1657 |
| Multivessel disease | 1.134 (0.933–1.379) | 1.5929 | 0.2069 |
| CABG | 2.780 (0.540–14.308) | 1.4964 | 0.2212 |
| Health literacy: low vs high | 0.840 (0.634–1.114) | 1.4681 | 0.2256 |
| Age ≥65 | 0.834 (0.614–1.134) | 1.3387 | 0.2473 |
| Multivessel PCI vs no PCI | 1.336 (0.795–2.245) | 1.1965 | 0.274 |
| Baseline EuroQol‐5D visual analog scale per 10 increase | 1.025 (0.979–1.072) | 1.1264 | 0.2886 |
| Prior coronary artery bypass grafting | 0.842 (0.605–1.173) | 1.0306 | 0.31 |
| Financial hardship of paying for meds: low/some vs none | 1.111 (0.906–1.362) | 1.0254 | 0.3112 |
| Prior MI | 1.158 (0.864–1.551) | 0.9589 | 0.3275 |
| Diabetes | 1.108 (0.900–1.365) | 0.9322 | 0.3343 |
| Heart failure | 1.163 (0.834–1.624) | 0.7912 | 0.3737 |
| Home aspirin | 0.915 (0.742–1.129) | 0.6826 | 0.4087 |
| In‐hospital recurrent MI | 1.642 (0.502–5.369) | 0.6735 | 0.4118 |
| Hypertension | 0.916 (0.741–1.133) | 0.6509 | 0.4198 |
| Nadir hemoglobin per 5 increase | 0.899 (0.684–1.181) | 0.5878 | 0.4433 |
| Hospital: bed size per 100 increase | 1.013 (0.975–1.053) | 0.4564 | 0.4993 |
| Race: White vs non‐White | 0.909 (0.684–1.209) | 0.4307 | 0.5116 |
| Daily/weekly angina vs no angina | 0.917 (0.708–1.188) | 0.4297 | 0.5121 |
| Peripheral artery disease | 1.133 (0.771–1.665) | 0.4015 | 0.5263 |
| Home P2Y12 | 1.106 (0.803–1.523) | 0.3808 | 0.5372 |
| Weight per 5 kg increase | 0.993 (0.971–1.016) | 0.3208 | 0.5711 |
| Dialysis | 0.843 (0.439–1.619) | 0.2639 | 0.6075 |
| Smoker | 0.948 (0.771–1.165) | 0.2608 | 0.6096 |
| Prior stroke/transient ischemic attack | 1.095 (0.749–1.599) | 0.2186 | 0.6401 |
| Prior heart failure | 1.095 (0.746–1.609) | 0.2151 | 0.6428 |
| Age per 10‐y increase | 1.031 (0.892–1.192) | 0.1744 | 0.6763 |
| Hospital: member of a health care network | 0.961 (0.793–1.165) | 0.1643 | 0.6852 |
| Monthly angina vs no angina | 1.034 (0.846–1.263) | 0.1066 | 0.7441 |
| Hospital: surgery capability | 1.052 (0.771–1.434) | 0.1015 | 0.75 |
| Cardiac arrest | 1.090 (0.619–1.919) | 0.0887 | 0.7658 |
| Cardiac rehabilitation referral | 0.976 (0.807–1.179) | 0.0647 | 0.7991 |
| Transfer‐in | 1.020 (0.838–1.243) | 0.0402 | 0.841 |
| Education: college or higher degree | 1.018 (0.847–1.224) | 0.0369 | 0.8477 |
| Randomization scheme: 2:1 vs 1:1 | 1.012 (0.826–1.239) | 0.0124 | 0.9113 |
| Employed: full/part‐time | 1.012 (0.807–1.268) | 0.0101 | 0.9198 |
| Patient Health Questionnaire‐2 depression score >3 | 0.988 (0.726–1.344) | 0.0058 | 0.9391 |
| In‐hospital or prior bleeding | 0.987 (0.588–1.658) | 0.0024 | 0.961 |
MI indicates myocardial infarction; OR, odds ratio; and PCI, percutaneous coronary intervention.
*P<0.05.
Model discrimination was tested in several secondary analyses. In case patients filled medications at pharmacies not captured by the Symphony data set, models 2 and 3 subset to usual care arm patients with at least 1 P2Y12 inhibitor or at least 1 fill of any medication in every quarter. Occasionally, hospitals may sponsor the first 30 days of medication, so model 4 tested model performance ignoring the first 30 days post hospital discharge. We externally validated the model in the copayment intervention arm in model 5, although we expected that the presence of the copayment voucher would change the accuracy of the predictions and clinical usefulness of the risk score model, so model 6 reestimated the beta‐coefficients in this population. Model 7 applied a definition of persistence in the intervention arm that leveraged both pharmacy fills and copayment voucher use to more optimally capture medication fills (ie, even if the patient did not fill at pharmacy reporting Symphony data, study copayment voucher use documented a fill). Finally, model 8 added several patient‐reported variables, including prior mail‐order pharmacy use, prior difficulty accessing pharmacy for prescription filling, prior history of not filling a prescription due to cost, worry about heart failure, patient‐perceived impact of medications (on reducing risk of future heart attack, likelihood of improving symptoms, possible side effects), convenience and cost of taking medication, impact of doctor recommendation to take medications, and prescription of any cardiac medication on discharge with more than once daily frequency.
We imputed missing socioeconomic variables, laboratory values, and weight to age‐, sex‐, and race‐specific modes for categorical variables and medians for continuous variables. Missing data related to medical history, home medications, admission features, and in‐hospital events were imputed to the mode. For each variable included in models, <5% of patients had missing data for all variables. All analyses were performed by statisticians at the Duke Clinical Research Institute (Durham, NC) using SAS version 9.4.
RESULTS
Among 6212 patients discharged with a planned course of 1‐year P2Y12 inhibitor therapy post MI, the 1‐year nonpersistence rate was 47.9% (95% CI, 46.6%–49.1%). Nonpersistent patients were more likely to be Black (11.3% versus 7.7%, P<0.0001) and have noncommercial insurance (Table 3). Nonpersistent patients had higher rates of medical comorbidities, including cardiovascular risk factors, as well as prior MI and revascularization (Table 3). Although patients with nonpersistence had higher rates of P2Y12 inhibitor use before their index hospitalization (16.0% versus 9.8%, P<0.0001), they did not have a significant difference in bleeding rates in the year before their hospitalization (P=0.15).
Table 3.
Baseline Characteristics of Persistent and Nonpersistent Patients
| Variable | Overall (N=6212) | Persistent (N=3282) | Nonpersistent (N=2930) | P value |
|---|---|---|---|---|
| Age, y | 62 (54–70) | 62 (54–69) | 62 (53–70) | 0.4820 |
| Sex | 0.2316 | |||
| Male | 4249 (68.4%) | 2223 (67.73%) | 2026 (69.15%) | |
| Female | 1963 (31.6%) | 1059 (32.27%) | 904 (30.85%) | |
| Race | ||||
| White | 5549 (89.33%) | 2991 (91.13%) | 2558 (87.3%) | <0.0001 |
| Black | 584 (9.40%) | 254 (7.74%) | 330 (11.26%) | <0.0001 |
| Asian | 106 (1.71%) | 57 (1.74%) | 49 (1.67%) | 0.8449 |
| American Indian or Alaska Native | 70 (1.13%) | 32 (0.98%) | 38 (1.30%) | 0.2302 |
| Native Hawaiian or other Pacific Islander | 29 (0.47%) | 18 (0.55%) | 11 (0.38%) | 0.3180 |
| Hispanic ethnicity | 198 (3.19%) | 109 (3.32%) | 89 (3.04%) | 0.2316 |
| Insurance status | ||||
| Commercial | 4084 (65.74%) | 2261 (68.89%) | 1823 (62.22%) | <0.0001 |
| Medicaid | 500 (8.05%) | 226 (6.89%) | 274 (9.35%) | 0.0004 |
| Military health | 164 (2.64%) | 40 (1.22%) | 124 (4.23%) | <0.0001 |
| Home P2Y12 receptor inhibitor use before admission | 790 (12.72%) | 322 (9.81%) | 468 (15.97%) | <0.0001 |
| Past medical history | ||||
| Hypertension | 4227 (68.05%) | 2153 (65.6%) | 2074 (70.78%) | <0.0001 |
| Dyslipidemia | 3609 (58.1%) | 1816 (55.33%) | 1793 (61.19%) | <0.0001 |
| Diabetes | 1965 (31.63%) | 945 (28.79%) | 1020 (34.81%) | <0.0001 |
| Prior myocardial infarction | 1198 (19.29%) | 537 (16.36%) | 661 (22.56%) | <0.0001 |
| Prior percutaneous coronary intervention | 1505 (24.23%) | 644 (19.62%) | 861 (29.39%) | <0.0001 |
| Prior coronary artery bypass grafting | 608 (9.79%) | 266 (8.1%) | 342 (11.67%) | <0.0001 |
| Prior stroke/transient ischemic attack | 370 (5.96%) | 170 (5.18%) | 200 (6.83%) | 0.006 |
| Peripheral arterial disease | 367 (5.91%) | 161 (4.91%) | 206 (7.03%) | 0.0004 |
| End‐stage renal disease | 112 (1.8%) | 43 (1.31%) | 69 (2.35%) | 0.002 |
| Bleeding in the past year | 66 (1.06%) | 29 (0.88%) | 37 (1.26%) | 0.15 |
Substantial nonpersistence developed early after discharge, with 23.8% (95% CI, 22.7%–24.8%) demonstrating nonpersistence at 30 days and 38.4% (95% CI, 37.2%–39.6%) at 6 months. Among the 1475 patients with nonpersistence within 30 days, 1331 (90.2%) had in‐hospital percutaneous coronary intervention, and the majority of these (n=1224) underwent drug‐eluting stent implantation. Nonpersistence rates were consistently lower among patients enrolled in the copayment intervention arm than the usual care arm at all time points (P=0.01). Even with copayment assistance, nonpersistence rates were 22.0% (20.7%–23.3%), 36.0% (34.5%–37.5%), and 45.3% (43.8%–46.9%) at 30 days, 6 months, and 1 year, respectively.
Among patients found to be nonpersistent by pharmacy fills, 41.5% self‐reported they were persistent (ie, discordance between patient report and pharmacy fills). The most common patient‐reported reason for nonpersistence was being told by their physician that they no longer needed to be taking this medication (18.3%). Bleeding/bruising accounted for 3.8% of nonpersistence with the majority of these (74.7%) discontinued under clinician guidance. Time to nonpersistence varied with reason for nonpersistence (Table 1).
A logistic regression model was fit within the usual care arm using the full variable list from the primary outcome propensity model of the original ARTEMIS trial, inclusive of demographic, medical history, and hospital characteristics. Parameter estimates for the model are shown in Table 2, with male sex, discharge hospital type, and financial hardship shown to be the strongest factors associated with nonpersistence. However, the naïve C‐index for this model was only 0.630, with an optimism corrected C‐index of 0.583. Optimism‐corrected calibration plots showed an intercept of 0.025 with a slope of 0.64 (Figure). Table 4 shows model performance in sensitivity cohorts to account for variations in pharmacy fills, validated in the copayment intervention arm cohort, and with additional patient‐reported factors that might predict medication‐taking behavior. None of these models substantially improved model performance with C‐indices <0.65.
Figure . Optimism‐Corrected Calibration Plots.
Optimism‐corrected calibration plots for ROC curve for model (A); and calibration plot for model among usual care (B). A, This displays the model's ability to discriminate between medication persistence and nonpersistence. Discrimination will be assessed using the naïve C‐statistic calculated from the original sample and the optimism‐corrected C‐statistic (defined as naïve C‐statistic—optimism). The C‐statistic represents the probability that a randomly selected patient having the event of interest (nonpersistence) will have a higher risk score than the patient who did not experience the event of interest. C‐statistic ranges from 0.5 to 1, with higher numbers indicating better discrimination. B, This displays calibration on the original data set. Calibration refers to whether predicted probabilities agree with observed end points. The calibration plot has predicted probabilities on the x axis, observed probabilities on the y axis, a LOESS curve, and a diagonal line which indicates perfect calibration (slope=1 and intercept=0). If the LOESS curve closely approximates the diagonal line, then the model is well calibrated (to the original data). The curve is close to the dashed line of perfect calibration for predicted probabilities ranging from about 0.3 to 0.9. LOESS indicates locally estimated scatterplot smoothing; and ROC, receiver‐operating characteristic.
Table 4.
Model Performance in Sensitivity Cohorts
| Model | Description | C‐index |
|---|---|---|
| 1 | Primary model (usual care arm) | 0.630 naïve/0.583 optimism corrected |
| 2 | Model tested in usual care arm among patients with at least 1 P2Y12 inhibitor fill in Symphony database | 0.629 |
| 3 | Model tested in usual care arm with at least 1 fill of any medication every quarter in Symphony database | 0.654 |
| 4 | Model tested in usual care arm with 30‐day grace period | 0.629 |
| 5 | Model tested in intervention arm | 0.565 |
| 6 | Model refit in intervention arm | 0.635 |
| 7 | Model tested in intervention arm with persistence defined by both fills and voucher use | 0.649 |
| 8 | Model 1+additional patient‐reported variables for medication affordability, depression, prior medication adherence, and health literacy | 0.621 |
DISCUSSION
Although practice guidelines recommend 1 year of dual antiplatelet therapy for patients post MI, we observed that P2Y12 persistence was poor, with almost half of all patients discontinuing their therapy within 1 year, even with prescription drug coverage. Nonpersistence began early after discharge, with 23.8% of patients discontinuing therapy by 30 days post MI. Of concern, the most common reason for P2Y12 inhibitor discontinuation reported by patients was instruction by their physician that they no longer needed to be taking this medication. Although we were able to identify multiple factors associated with P2Y12 persistence, multivariable models had at best modest‐to‐low ability to predict patients at increased risk of P2Y12 inhibitor nonpersistence.
The ARTEMIS trial randomized hospitals to usual care versus an intervention to eliminate patients' P2Y12 inhibitor copayments after acute MI. In intention‐to‐treat analyses, the intervention improved guideline‐adherent treatment selection and longitudinal medication persistence, 4 echoing other studies showing that financial assistance can enhance access to health care. 10 Nevertheless, 28% of patients in the intervention arm chose not to use the copayment vouchers provided, and these patients had lower medication persistence rates and higher major adverse cardiac event rates compared with intervention arm patients who opted to use the copayment voucher. These results suggest that although copayment assistance programs may improve persistence of evidence‐based mediation use for a subset of patients, other patients may require other interventions to support continued use of critical treatments and preventive health visits. 11
Models predicting medication‐taking behavior have been a “holy grail” for clinicians. Being able to predict which patients might benefit from extra attention and resources to improve medication persistence has potential for optimizing clinical outcomes, particularly after acute MI. Several predictive models of medication nonadherence and nonpersistence risks have been developed in other disease states. 12 , 13 Nevertheless, the results of our predictive modeling were disappointing. In a prior study, psychosocial factors, and health belief‐related factors were associated with patient self‐reported medication nonpersistence. 14 This study examined pharmacy fill‐based medication persistence, which is generally more accurate than patient self‐reported persistence, 15 with several sensitivity analyses to account for potential limitations of pharmacy fill records given the multipayer, multipharmacy structure that exists in the United States today. Beyond inclusion of the typical demographic, financial hardship, and clinical history variables, we added variables that were uniquely collected in this study, such as prior medication‐filling behavior, patient‐reported worry about future cardiovascular disease, patient‐perceived impact of medications, patient‐reported influence of clinician recommendation, and convenience of the prescribed medication regimen. Nonetheless, model performance remained poor despite the enriched variable list.
One reason the model may not have accurately predicted medication‐taking behavior is that medication persistence does not appear to be entirely driven by the patient. Almost 1 in 5 patients with nonpersistence attributed medication discontinuation to physician guidance. Although it is reasonable for clinicians to pause or discontinue P2Y12 inhibitor therapy in the setting of bleeding, we found that bleeding, bruising, or other side effects comprised only a small fraction of patients who reported their clinician told them to discontinue treatment. Therefore, some of the early discontinuation of treatment may reflect clinician beliefs and practice patterns. Although practice guidelines continued to recommend a 1‐year duration of P2Y12 inhibitor post MI, contemporary clinical studies during the ARTEMIS study period explored shorter duration therapy and single antiplatelet therapy regimens for patients treated with drug‐eluting stents, 16 , 17 , 18 , 19 which may have influenced clinician decision‐making, despite guideline recommendations. Of concern, 1 in 4 premature discontinuations of P2Y12 inhibitor therapy occurred within the first 30 days after hospital discharge in patients with acute MI treated with and without percutaneous coronary intervention. To date, most studies of single antiplatelet therapy have only examined antiplatelet discontinuation after 30 days, as the early period post MI is regarded as a high‐risk period warranting intensive antiplatelet treatment. Continued attention to close follow‐up after acute MI discharge and clinician education on importance of P2Y12 inhibitor treatment is warranted in hopes of improving patient persistence with guideline recommend medications following MI. As there is increasing information in the electronic health record around prescription fills, this information could be an opportunity for clinicians to discuss in real time with patients during follow‐up what the barriers to persistence are.
Limitations
Although the data source used in this study covers a wide range of physical and mail‐order pharmacies used in the United States, P2Y12 inhibitor persistence may be underestimated if patients filled their prescription at a non‐data source‐linked pharmacy. A prior study showed major adverse cardiac event rates to be higher in patients who self‐reported persistence but were found to be nonpersistent by pharmacy fills when compared with patients who were concordantly persistent by both self‐report and pharmacy fills; such a finding suggests that not all nonpersistence can be attributed to gaps in pharmacy data. Patients enrolled in ARTEMIS needed to have US‐based health insurance with a prescription drug benefit, so study results may not be generalizable to patients without health insurance or prescription coverage. Additionally, the data source is not able to detect gaps in persistence <30 days and is not able to capture provider type or specialty caring for patients who discontinued/were nonpersistent. Furthermore, the extent of bleeding or bruising that led to treatment discontinuation or nonpersistence cannot be determined from this data source.
Conclusions
Models predicting persistence with P2Y12 inhibitor therapy performed poorly even when these models were composed of diverse demographic, clinical, and patient‐reported variables, including prior medication‐filling behavior. Many patients attribute premature P2Y12 inhibitor discontinuation to physician guidance that they no longer need to be taking this class of medication even when a 1‐year treatment course was planned at hospital discharge, and many premature discontinuations occur within the first 30 days after MI hospitalization. These results suggest the need for continued patient and clinician education on the importance of P2Y12 inhibitor therapy after acute MI.
Sources of Funding
This article was funded by AstraZeneca.
Disclosures
E. Peterson reports research grants from Amgen, BMS, Janssen, and Esperion; advisory board for Novartis, BI, Bayer, NovoNordisk, and Cerner. T. Wang reports research grants to the Duke Clinical Research Institute from Abbott, AstraZeneca, Bristol Myers Squibb, Boston Scientific, Artivion, Chiesi, Merck, Portola, and Regeneron; consulting honoraria from AstraZeneca, Bristol Myers Squibb, Artivion, Novartis, and CSL Behring. The remaining authors have no disclosures to report.
This article was sent to Ajay K. Gupta, MD, MSc, PhD, FRCP, FESC, Senior Associate Editor, for review by expert referees, editorial decision, and final disposition.
For Sources of Funding and Disclosures, see page 8.
REFERENCES
- 1. Osterberg L, Blaschke T. Adherence to medication. N Engl J Med. 2005;353:487–497. doi: 10.1056/NEJMra050100 [DOI] [PubMed] [Google Scholar]
- 2. Choudhry NK, Avorn J, Glynn RJ, Antman EM, Schneeweiss S, Toscano M, Reisman L, Fernandes J, Spettell C, Lee JL, et al. Full coverage for preventive medications after myocardial infarction. N Engl J Med. 2011;365:2088–2097. doi: 10.1056/NEJMsa1107913 [DOI] [PubMed] [Google Scholar]
- 3. Doll JA, Hellkamp AS, Thomas L, Fonarow GC, Peterson E, Wang TY. The association of pre‐ and posthospital medication adherence in myocardial infarction patients. Am Heart J. 2019;208:74–80. doi: 10.1016/j.ahj.2018.11.004 [DOI] [PubMed] [Google Scholar]
- 4. Wang TY, Kaltenbach LA, Cannon CP, Fonarow GC, Choudhry NK, Henry TD, Cohen DJ, Bhandary D, Khan ND, Anstrom KJ, et al. Effect of medication co‐payment vouchers on P2Y12 inhibitor use and major adverse cardiovascular events among patients with myocardial infarction: the ARTEMIS randomized clinical trial. JAMA. 2019;321:44–55. doi: 10.1001/jama.2018.19791 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. Steiner JF, Ho PM, Beaty BL, Dickinson LM, Hanratty R, Zeng C, Tavel HM, Havranek EP, Davidson AJ, Magid DJ, et al. Sociodemographic and clinical characteristics are not clinically useful predictors of refill adherence in patients with hypertension. Circ Cardiovasc Qual Outcomes. 2009;2:451–457. doi: 10.1161/CIRCOUTCOMES.108.841635 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Chan DC, Shrank WH, Cutler D, Jan S, Fischer MA, Liu J, Avorn J, Solomon D, Brookhart MA, Choudhry NK. Patient, physician, and payment predictors of statin adherence. Med Care. 2010;48:196–202. doi: 10.1097/MLR.0b013e3181c132ad [DOI] [PubMed] [Google Scholar]
- 7. Doll JA, Wang TY, Choudhry NK, Cannon CP, Cohen DJ, Fonarow GC, Henry TD, Bhandary DD, Khan N, Davidson‐Ray LD, et al. Rationale and design of the Affordability and Real‐world Antiplatelet Treatment Effectiveness after Myocardial Infarction Study (ARTEMIS): a multicenter, cluster‐randomized trial of P2Y12 receptor inhibitor copayment reduction after myocardial infarction. Am Heart J. 2016;177:33–41. doi: 10.1016/j.ahj.2016.04.008 [DOI] [PubMed] [Google Scholar]
- 8. Harrell FE. Regression Modeling Strategies: With Applications to Linear Models, Logistic and Ordinal Regression, and Survival Analysis. Switzerland AG: Springer, 2015. [Google Scholar]
- 9. Austin PC, Steyerberg EW. Predictive accuracy of risk factors and markers: a simulation study of the effect of novel markers on different performance measures for logistic regression models. Stat Med. 2013;32:661–672. doi: 10.1002/sim.5598 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10. Bellows B, Bulaya C, Inambwae S, Lissner CL, Ali M, Bajracharya A. Family planning vouchers in low and middle income countries: a systematic review. Stud Fam Plann. 2016;47:357–370. doi: 10.1111/sifp.12006 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11. Fanaroff AC, Peterson ED, Kaltenbach LA, Anstrom KJ, Fonarow GC, Henry TD, Cannon CP, Choudhry NK, Cohen DJ, Atreja N, et al. Copayment reduction voucher utilization and associations with medication persistence and clinical outcomes: findings from the ARTEMIS Trial. Circ Cardiovasc Qual Outcomes. 2020;13:e006182. doi: 10.1161/CIRCOUTCOMES.119.006182 [DOI] [PubMed] [Google Scholar]
- 12. Wu XW, Yang HB, Yuan R, Long EW, Tong RS. Predictive models of medication non‐adherence risks of patients with T2D based on multiple machine learning algorithms. BMJ Open Diabetes Res Care. 2020;8:e001055. doi: 10.1136/bmjdrc-2019-001055 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Gu Y, Zalkikar A, Liu M, Kelly L, Hall A, Daly K, Ward T. Predicting medication adherence using ensemble learning and deep learning models with large scale healthcare data. Sci Rep. 2021;11:18961. doi: 10.1038/s41598-021-98387-w [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Crowley MJ, Zullig LL, Shah BR, Shaw RJ, Lindquist JH, Peterson ED, Bosworth H. Medication non‐adherence after myocardial infarction: an exploration of modifying factors. J Gen Intern Med. 2015;30:83–90. doi: 10.1007/s11606-014-3072-x [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Rocha RV, Tam DY, Karkhanis R, Wang X, Austin PC, Ko DT, Gaudino M, Royse A, Fremes SE. Long‐term outcomes associated with total arterial revascularization vs non‐total arterial revascularization. JAMA Cardiol. 2020;5:532–539. doi: 10.1001/jamacardio.2019.6104 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Feres F, Costa RA, Abizaid A, Leon MB, Marin‐Neto JA, Botelho RV, King SB III, Negoita M, Liu M, de Paula JET, et al. Three vs twelve months of dual antiplatelet therapy after zotarolimus‐eluting stents: the OPTIMIZE randomized trial. JAMA. 2013;310:2510–2522. doi: 10.1001/jama.2013.282183 [DOI] [PubMed] [Google Scholar]
- 17. Valgimigli M, Patialiakas A, Thury A, McFadden E, Colangelo S, Campo G, Tebaldi M, Ungi I, Tondi S, Roffi M, et al. Zotarolimus‐eluting versus bare‐metal stents in uncertain drug‐eluting stent candidates. J Am Coll Cardiol. 2015;65:805–815. doi: 10.1016/j.jacc.2014.11.053 [DOI] [PubMed] [Google Scholar]
- 18. Schulz‐Schüpke S, Byrne RA, Ten Berg JM, Neumann F‐J, Han Y, Adriaenssens T, Tölg R, Seyfarth M, Maeng M, Zrenner B, et al. ISAR‐SAFE: a randomized, double‐blind, placebo‐controlled trial of 6 vs. 12 months of clopidogrel therapy after drug‐eluting stenting. Eur Heart J. 2015;36:1252–1263. doi: 10.1093/eurheartj/ehu523 [DOI] [PubMed] [Google Scholar]
- 19. Franzone A, McFadden E, Leonardi S, Piccolo R, Vranckx P, Serruys PW, Benit E, Liebetrau C, Janssens L, Ferrario M, et al. Ticagrelor alone versus dual antiplatelet therapy from 1 month after drug‐eluting coronary stenting. J Am Coll Cardiol. 2019;74:2223–2234. doi: 10.1016/j.jacc.2019.08.1038 [DOI] [PubMed] [Google Scholar]