This analysis of a randomized clinical trial examines the clinical event data to determine which source of data (eg, electronic health records, claims, participant reports) made the greatest relative contribution to the primary end point rates.
Key Points
Question
Which source of data provides the greatest relative contribution for clinical end-point examination in pragmatic randomized clinical trials?
Findings
In this prespecified analysis of a randomized clinical trial, most events (92%-100%) were identified in the electronic health record (EHR) or in claims data. Participant-reported data contributed less than 10% of events not otherwise available from claims or EHR data.
Meaning
In a large pragmatic randomized clinical trial, claims and EHR data provided the most clinical end-point data when compared with participant-reported events.
Abstract
Importance
Pragmatic randomized clinical trials (RCTs) often use multiple data sources to examine clinical events, but the relative contribution of data sources to clinical end-point rates is understudied.
Objective
To assess the contribution of data sources (electronic health records [EHRs], public/private insurance claims, and/or participant-reported data) to clinical end points among ADAPTABLE participants who had available data.
Design, Setting, and Participants
The ADAPTABLE study was an open-label, pragmatic RCT from April 2016 through June 2019 conducted in research networks within clinical practice. Participants had existing atherosclerotic cardiovascular disease and available data to analyze. The characteristics of patients by combinations of data source availability were compared to examine the contribution of each of the data sources to end-point ascertainment. Data for this prespecified analysis were examined from January 2022 to June 2023.
Exposures
Randomized exposure to 81 mg or 325 mg of aspirin daily.
Main Outcomes and Measures
Number of events for the primary end point (composite of death, hospitalization for myocardial infarction, and hospitalization for stroke) that were contributed by EHR or claims data and then number of events contributed by each additional data source.
Results
Of 15 006 participants randomized with at least 1 other source of data available beyond participant-reported data, there were 8756 (58.3%) with participant-reported and EHR data; 4291 (28.6%) with participant-reported, EHR, and claims data; 1412 (9.4%) with EHR-only data; 262 (1.7%) with participant-reported and claims data; 202 (1.3%) with EHR and claims data; and 83 (0.6%) with claims-only data. Participants with EHR-only data were younger (median age, 63.7 years; IQR, 55.8-71.4) compared with the other groups (range, 65.6-71.9 years). Among participants with both EHR and claims data, with or without participant-reported data (n = 4493), for each outcome, most events (92%-100%) were identified in the EHR or in claims data. For all clinical end points, participant-reported data contributed less than 10% of events not otherwise available from claims or EHR data.
Conclusions and Relevance
In this analysis of a pragmatic RCT, claims and EHR data provided the most clinical end-point data when compared with participant-reported events. These findings provide a framework for collecting end points in pragmatic clinical trials. Further work is needed to understand the data source combinations that most effectively provide clinical end-point data in RCTs.
Introduction
There has been increasing interest over the past decade to innovate within the traditional clinical trial and improve efficiency, sustainability, and generalizability. One of the promises of the pragmatic clinical trial is extracting clinical events from both traditional and existing sources of data, as well as using streamlined methods of follow-up.1 The ADAPTABLE trial (Aspirin Dosing: A Patient-Centric Trial Assessing Benefits and Long-term Effectiveness) was a pragmatic trial that examined the optimal dose of aspirin therapy for secondary prevention in atherosclerotic cardiovascular disease.2 As with many pragmatic clinical trials, ADAPTABLE used participant-facing internet portals to gather patient-reported events, in addition to using claims data (private insurance and Centers for Medicare & Medicaid Services [CMS]) and electronic health record (EHR) data. While a recent analysis of ADAPTABLE demonstrated poor concordance between participant-reported events and claims and EHR data,3 little is known about the relative contribution of EHR, claims, and participant-reported data to event collection. We examined the contribution of event sources to event collection for each of the clinical end points within the ADAPTABLE trial.
Methods
Trial Design
The ADAPTABLE trial examined the optimal aspirin dose for secondary prevention of ischemic disease in participants with existing atherosclerotic cardiovascular disease. The trial randomized 15 076 participants with atherosclerotic cardiovascular disease in a 1:1 ratio to receive a dose of aspirin, 81 mg daily vs 325 mg daily, from April 2016 through June 2019. An independent data and safety monitoring committee approved the trial protocol. All patients provided electronic informed consent before enrollment in the trial. Data for this analysis were examined from January 2022 to June 2023.
Events were ascertained through several sources, including from the participant at scheduled intervals (self-reported hospitalization data from the participant portal), EHR-based queries of datamarts at participating sites and the National Patient-Centered Network Clinical Research Network (PCORnet), and linkage with administrative claims from both private health insurers and from the CMS Medicare program. The inclusion criteria have been previously described (trial protocol in Supplement 1 and statistical analysis plan in Supplement 2).4 The primary end point was a composite of all-cause death, hospitalization for myocardial infarction, and hospitalization for stroke. The primary safety end point was hospitalization for major bleeding with an associated blood transfusion.
Data Sources
The current analysis was prespecified. Detailed information about participant data sources and outcome extraction is provided in the supplement to the main results article.2 There were 3 sources of participant-level data and outcomes extracted during the study: EHR data, participant-reported (portal) data, and claims data from public and private insurance carriers. Enrolled participants were linked across data sources via the use of an ADAPTABLE study identifier. Electronic health record data were queried for participants approximately every 6 months during the study. Health systems extracted and loaded EHR data into a datamart using the PCORnet common data model, which serves as the foundation of the PCORnet data research network. Electronic health record queries involved programmed algorithms for clinical outcomes using International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM),5 ICD-10-CM,6 and ICD-10 Procedure Coding System7 codes. Medicare claims data and private health plan data were extracted with the use of ICD-9 and ICD-10 codes and loaded into the common data model.
Participant-reported health data were captured via an electronic patient portal/platform (originally designed by Mytrus, now Medidata). Enrolled participants were randomized to receive a link to the ADAPTABLE web portal every 3 or 6 months (or via telephone calls if the participant was uncertain they could use the internet patient portal). There were 5 domains of information captured within the participant-reported data: demographic and eligibility criteria, quality of life, nonstudy medications that participants were taking, aspirin use, and hospitalizations. Race and ethnicity were self-reported by the participants from the categories Asian, Black, Hispanic, White, and undetermined (“prefer not to say”). Participant-reported hospitalizations were not included as an event in the trial without a hospitalization record, which required release authorizations to be signed.
Definitions
We examined subgroups of participants based on data source availability. Data availability was defined using postrandomization information, where availability for a given source is defined as availability for at least half of the participant’s follow-up period. The follow-up period was defined as the time from randomization to the earliest of June 30, 2020, date of withdrawn consent, date of death, or the data source censoring date. The data source censoring dates were defined as follows: date of the last observed participant-reported visit, date of latest EHR data available for a site within PCORnet, CMS censor date (CMS-specific censor date or end of fee-for-service enrollment date), or private claims censor date (health plan censor date or end of enrollment date).
The primary efficacy end point was a composite of death, hospitalization for myocardial infarction (MI), and hospitalization for stroke. The secondary efficacy end points included all-cause mortality, hospitalization for MI, and hospitalization for stroke. The primary safety end point was hospitalization for major bleeding with associated blood transfusion. Electronic health record and claims data were considered the gold standard for end-point data during the trial.
Statistical Analysis
We compared the baseline characteristics of patients by combinations of data source availability (eg, EHR/claims/patient-reported, EHR/patient-reported, EHR only). Data availability for claims data was defined by enrollment start and stop dates. For patient-reported outcome data, it was defined by the first and last portal visit. For EHR data, availability was based on randomization date and the earliest of end of EHR data on the site level (for most sites was June 30, 2020), date consent was withdrawn, or death. Continuous variables were presented as median (IQR) and categorical variables as frequencies and percentages. To examine the contribution of each of the data sources to end-point ascertainment, we then restricted the population to those participants with both EHR and claims data, with or without patient-reported data. We then described the number of events for the primary and secondary end points identified by each data source or combination of data sources. We described the number of events contributed by EHR or claims data, and then the number of additional events that were contributed by each additional data source (ie, if starting with EHR data, how many additional events were contributed by claims or patient-reported data). Additionally, to account for the age of the CMS population, we repeated the analysis further restricting the population to participants 65 years and older. We also conducted a sensitivity analysis in which the definition of data availability was modified to require that the data source was available for 90% of the participant’s follow-up.
Results
Seventy patients had neither EHR nor claims data available for at least 50% of follow-up and were excluded (final population = 15 006). Of 15 006 participants randomized with at least 1 other source of data available beyond just participant-reported data, there were 8756 (58.3%) with participant-reported and EHR data; 4291 (28.6%) with participant-reported, EHR, and claims data (Medicare or private insurance); 1412 (9.4%) with EHR-only data; 262 (1.7%) with participant-reported and claims data; 202 (1.3%) with EHR and claims data; and 83 (0.6%) with claims-only data. Participants with EHR-only data were younger (median, 63.7 years; IQR, 55.8-71.4) compared with the other groups (range, 65.6-71.9 years). Participants with EHR-only data or EHR and claims data were more likely Black (150/1412 [10.6%] and 28/202 [13.9%], respectively) compared with the other groups (range, 3.4%-9.7%; P < .001) (Table 1).
Table 1. Baseline Characteristics by Data Source Availabilitya.
| Characteristic | No. (%) | P value | |||||
|---|---|---|---|---|---|---|---|
| Participant-reported visits + EHR + claims (n = 4291) | Participant-reported visits + EHR (n = 8756) | Participant-reported visits + claims (n = 262) | EHR + claims (n = 202) | EHR only (n = 1412) | Claims only (n = 83) | ||
| Age, median (range), y | 70.7 (66.4-75.4) | 65.6 (59.1-72.1) | 71.9 (66.8-77.5) | 70.7 (65.7-76.4) | 63.7 (55.8-71.4) | 71.2 (63.1-78.3) | <.001 |
| Female | 1200 (28.0) | 2807 (32.1) | 67 (25.6) | 78 (38.6) | 509 (36.0) | 31 (37.3) | <.001 |
| Race and ethnicityb | <.001 | ||||||
| Asian | 27 (0.6) | 106 (1.2) | 1 (0.4) | c | 10 (0.7) | c | |
| Black | 270 (6.3) | 847 (9.7) | 9 (3.4) | 28 (13.9) | 150 (10.6) | c | |
| Hispanic | 69 (1.6) | 306 (3.5) | 4 (1.5) | c | 88 (6.2) | c | |
| White | 3866 (90.1) | 6919 (79.0) | 231 (88.2) | 152 (75.2) | 774 (54.8) | c | |
| Smoking status | |||||||
| Current | 277 (6.5) | 888 (10.4) | 19 (7.3) | 22 (11.6) | 169 (21.0) | c | <.001 |
| BMI, median (range)d | 29.7 (26.4-33.7) | 30.2 (26.8-34.5) | NA | 28.9 (24.9-34.1) | 30.2 (26.5-35.2) | <.001 | |
| Medical history | |||||||
| CAD | 4042 (94.2) | 8190 (93.5) | 242 (92.4) | 190 (94.1) | 1292 (91.5) | 77 (92.8) | <.001 |
| MI | 1362 (31.7) | 3316 (37.9) | 115 (43.9) | 64 (31.7) | 563 (39.9) | 38 (45.8) | <.001 |
| CABG | 1060 (24.7) | 2076 (23.7) | 23 (8.8) | 52 (25.7) | 339 (24.0) | c | .70 |
| PCI | 1697 (39.5) | 3597 (41.1) | 42 (16.0) | 72 (35.6) | 580 (41.1) | c | .23 |
| Cerebrovascular disease | 804 (18.7) | 1517 (17.3) | 83 (31.7) | 39 (19.3) | 264 (18.7) | 25 (30.1) | .28 |
| Hypertension | 3667 (85.5) | 7454 (85.1) | 242 (92.4) | 173 (85.6) | 1218 (86.3) | 79 (95.2) | .13 |
| Hyperlipidemia | 3803 (88.6) | 7751 (88.5) | 251 (95.8) | 168 (83.2) | 1224 (86.7) | 77 (92.8) | .002 |
| Atrial fibrillation | 400 (9.3) | 709 (8.1) | 37 (14.1) | 16 (7.9) | 108 (7.6) | 12 (14.5) | .14 |
| Congestive heart failure | 986 (23.0) | 2033 (23.2) | 63 (24.0) | 63 (31.2) | 422 (29.9) | 25 (30.1) | <.001 |
| Peripheral artery disease | 1094 (25.5) | 1945 (22.2) | 83 (31.7) | 62 (30.7) | 392 (27.8) | 34 (41.0) | <.001 |
| Diabetes | 1575 (36.7) | 3438 (39.3) | 118 (45.0) | 91 (45.0) | 572 (40.5) | 47 (56.6) | .006 |
| History of bleeding | 374 (8.7) | 746 (8.5) | 39 (14.9) | c | 137 (9.7) | 12 (14.5) | .22 |
| Significant bleeding disorder | 58 (1.4) | 103 (1.2) | 3 (1.1) | c | 13 (0.9) | c | .63 |
| Significant GI bleed | 276 (6.4) | 550 (6.3) | 32 (12.2) | c | 116 (8.2) | c | .04 |
| Intracranial hemorrhage | 66 (1.5) | 125 (1.4) | 4 (1.5) | c | 16 (1.1) | c | .58 |
| Prior medications | |||||||
| Prior aspirin use | <.001 | ||||||
| None | 116 (2.7) | 368 (4.3) | 12 (4.6) | c | 52 (6.5) | c | |
| 81 mg | 3542 (82.8) | 6956 (82.0) | 204 (78.5) | 149 (78.8) | 645 (80.4) | c | |
| 162 mg | 101 (2.4) | 177 (2.1) | c | c | 20 (2.5) | c | |
| 325 mg | 520 (12.2) | 984 (11.6) | 39 (15.0) | 25 (13.2) | 85 (10.6) | c | |
| P2Y12 inhibitor | 921 (21.6) | 1818 (22.0) | 46 (18.1) | 44 (23.3) | 207 (26.4) | c | .04 |
| Trial adherence, % | |||||||
| Visits completed, median (range) | 100.0 (80.0-100.0) | 90.9 (70.0-100.0) | 100.0 (80.0-100.0) | 33.3 (18.2-50.0) | 16.7 (0.0-33.3) | 0.0 (0.0-0.0) | <.001 |
| Category of visit completion | <.001 | ||||||
| 0 | NA | 100 (1.1) | NA | 12 (6.0) | 575 (40.8) | 69 (83.1) | |
| 1-25 | 18 (0.4) | 175 (2.0) | NA | 71 (35.3) | 379 (26.9) | c | |
| 26-50 | 207 (4.8) | 800 (9.1) | 19 (7.3) | 68 (33.8) | 321 (22.8) | c | |
| 51-75 | 674 (15.7) | 1759 (20.1) | 25 (9.5) | 27 (13.4) | 97 (6.9) | c | |
| 76-99 | 965 (22.5) | 1746 (19.9) | 60 (22.9) | c | 0 (0.0) | c | |
| 100 | 2425 (56.5) | 4176 (47.7) | 154 (58.8) | 23 (11.4) | 37 (2.6) | c | |
Abbreviations: BMI, body mass index; CABG, coronary artery bypass graft procedure; CAD, coronary artery disease; EHR, electronic health record; GI, gastrointestinal; MI, myocardial infarction; NA, not applicable; PCI, percutaneous coronary intervention.
This table includes patients with the 50% data availability definition, so patients with patient-reported or no data were excluded (n = 70).
Race and ethnicity were self-reported by the participants from the categories Asian, Black, Hispanic, White, and undetermined (“prefer not to say”).
Data cells suppressed for participants with claims data where n < 11.
Calculated as weight in kilograms divided by height in meters squared.
Table 2 and the Figure describe the event counts by data source for participants who had EHR and claims data, with or without participant-reported visit data. Among these participants, for each outcome, most events (92%-100%) were identified in the EHR or in claims data, with the participant-reported data adding only a small number of events missed by other sources. Of the 417 composite end-point events identified in any source, 294 (71%) were identified in EHR data, while claims data contributed 117 (28%) events (Figure, A). Similarly, 342 composite events (82%) were identified in claims data, while EHR data contributed 69 events (17%) (Figure, B). For all-cause death, 157 events (65%) were identified in EHR data, and claims data contributed 80 events (33%) (Figure, A). However, EHR data contributed few events on top of claims data for all-cause death (3%). For the MI end point, claims data contributed 29% of events on top of EHR data, and EHR data contributed 33% of MI events on top of claims data. Similar trends were seen for the stroke and major bleed outcomes.
Table 2. Event Counts by Data Source for 4493 Participants Who Had EHR and Claims Data With or Without Portal-Visit Dataa.
| No. (%) | |||||
|---|---|---|---|---|---|
| Compositeb | All-cause death | MI | Stroke | Major bleed | |
| Events identified in either EHR or claims (Medicare and private) | 411 (99) | 237 (99) | 162 (98) | 79 (100) | 34 (92) |
| EHR onlyc | 69 (17) | 6 (3) | 55 (34) | 30 (38) | 10 (29) |
| Claims onlyc | 117 (28) | 80 (34) | 48 (30) | 21 (27) | 10 (29) |
| EHR and claimsc | 225 (55) | 151 (64) | 59 (36) | 28 (35) | 14 (41) |
| Events added by participant report (portal) | 6 (1) | 3 (1) | 3 (2) | 0 | 3 (8) |
| Total events identified by any source | 417 | 240 | 165 | 79 | 37 |
Abbreviations: EHR, electronic health record; MI, myocardial infarction.
Data availability was defined as availability for at least 50% of the participant’s follow-up period.
The composite end point included death, hospitalization for MI, and hospitalization for stroke.
The denominator is the number of events identified in either EHR or claims.
Figure. Event Counts by Data Source Added to Electronic Health Record (EHR) Data (n = 4493) and Claims Data (n = 4493).
A, We started with EHR data and examined the events added by claims data and participant portal data. B, We started with claims data and examined the events added by EHR data and participant portal data. MI indicates myocardial infarction.
These trends were similar when data availability was defined by 90% availability during follow-up (eTable in the Supplement). Table 3 describes event counts by data source for this population, restricted to participants 65 years and older. Similarly, for each outcome, most events (92%-100%) were identified in the EHR or in claims data, with the participant-reported data adding only a small number of events missed by other sources (1%-8%). Participant-reported data contributed 6 composite events and 3 all-cause death events.
Table 3. Event Counts by Data Source for 3700 Participants 65 Years and Older Who Had EHR and Claims Data With or Without Portal-Visit Dataa.
| No. (%) | |||||
|---|---|---|---|---|---|
| Compositeb | All-cause death | MI | Stroke | Major bleed | |
| Events identified in either EHR or claims (Medicare and private) | 321 (99) | 193 (99) | 121 (98) | 59 (100) | 23 (92) |
| EHR onlyb | 49 (15) | 3 (2) | 40 (33) | 23 (39) | 9 (39) |
| Claims onlyb | 91 (28) | 64 (33) | 34 (28) | 18 (31) | 2 (9) |
| EHR and claimsb | 181 (56) | 126 (65) | 47 (39) | 18 (31) | 12 (52) |
| Events added by participant report (portal) | 3 (1) | 1 (1) | 2 (2) | 0 | 2 (8) |
| Total events identified by any source | 324 | 194 | 123 | 59 | 25 |
Abbreviations: EHR, electronic health record; MI, myocardial infarction.
Data availability was defined as availability for at least 50% of the participant’s follow-up period.
The composite end point included death, hospitalization for MI, and hospitalization for stroke.
The denominator is the number of events identified in either EHR or claims.
Discussion
In the current analysis, we examined the relative data source contribution for various clinical end points in a large pragmatic trial. We demonstrated that (1) claims and EHR data contributed 92% to 100% of the composite end point and secondary end point events among participants with EHR and claims data; (2) this trend was consistent among older participants (≥65 years); and (3) for participants with available EHR and claims, patient-reported data contributed relatively little in addition to the other event sources among patients with available EHR and claims data. Lastly, EHR data add little to capturing all-cause death when there are available claims data.
Importantly, we showed that the vast majority of events for each of the clinical end points were extracted from either EHR or claims data with very little contribution from participant-reported data. A recently published analysis from ADAPTABLE demonstrated that participant reports of clinical events had very low sensitivity compared with EHR data for MI, stroke, and bleeding.3 Significant discrepancies between claims data and self-reported events for acute MI and for rehospitalizations have been previously shown8; our analysis adds to these findings and suggests that participant-reported data may not successfully capture a significant number of events outside of traditional approaches using claims and EHR data. This finding raises a question regarding the utility of participant-reported data given (1) its lack of concordance with other data event sources and (2) its limits in terms of meaningfully adding events that were not captured by claims or EHR. It is very important to note that hospitalization events reported by participants in the ADAPTABLE trial required authorized releases, which were often difficult to obtain. Therefore, one of the important lessons, if considering participant-reported data sources in a clinical trial, is to ensure that there are robust processes in place to obtain authorization from participants so hospitals can be queried about participant-reported events.
Limitations
There are several limitations to note. We reported the contribution of the data sources among participants with all 3 data sources available or among participants with claims and EHR data. It is possible that there were unreported events among this population that were not reported by any of the data sources and would therefore result in an underestimate of the number of clinical events overall. It should also be noted that this population was older and more likely to have CMS claims data; thus, the findings may not be generalizable to a younger population or one without health insurance. Additionally, as the site often asks similar questions of patients that are asked of participants during study follow-up, site-based interviews and discussions with participants should be incorporated with data collected from participant-reported events. Finally, as participant-reported events had to be queried with the use of patient authorizations, it may be that some participant-reported events were missed if authorization was not granted. In the future, we will request preauthorization of all participants for future events and for linkage to private and public insurance data. Lastly, it is likely that the potential yield of EHR data has increased since ADAPTABLE was conducted. With the expansion of Care Everywhere in the Epic EHR to allow for seamless sharing of data across institutions and the evolving capabilities of artificial intelligence and natural language processing software, there may be an increased capability to capture events outside the common data model. This capability might minimize the importance of collecting patient-reported events and facilitate a more pragmatic ascertainment of events.
Conclusions
In a large pragmatic randomized clinical trial, claims data contributed the most clinical events for all-cause death. When compared with other available data sources, few events were contributed by participant-reported data that were not captured by claims or EHR data. Future studies with access to EHR and claims data should consider the respondent burden relative to the low yield of participant portal (or other self-reported) data.
Trial protocol
Statistical analysis plan
eTable. Event Counts by Data Source for 2,098 Participants Who Have EHR and Claims Data, With or Without Portal-visit Data
Data sharing statement
References
- 1.NIH Pragmatic Trials Collaboratory . Using electronic health record data in pragmatic clinical trials. Published August 25, 2017. Updated August 26, 2022. Accessed September 1, 2023. https://rethinkingclinicaltrials.org/chapters/design/using-electronic-health-record-data-pragmatic-clinical-trials-top/using-electronic-health-record-data-in-pragmatic-clinical-trials-introduction/
- 2.Jones WS, Mulder H, Wruck LM, et al. ; ADAPTABLE Team . Comparative effectiveness of aspirin dosing in cardiovascular disease. N Engl J Med. 2021;384(21):1981-1990. doi: 10.1056/NEJMoa2102137 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.O’Brien EC, Mulder H, Jones WS, et al. Concordance between patient-reported health data and electronic health data in the ADAPTABLE trial. JAMA Cardiol. 2022;7(12):1235-1243. doi: 10.1001/jamacardio.2022.3844 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Marquis-Gravel G, Roe MT, Robertson HR, et al. Rationale and design of the aspirin dosing: a patient-centric trial assessing benefits and long-term effectiveness (ADAPTABLE) trial. JAMA Cardiol. 2020;5(5):598-607. doi: 10.1001/jamacardio.2020.0116 [DOI] [PubMed] [Google Scholar]
- 5.World Health Organization . International Classification of Diseases, Ninth Revision (ICD-9). World Health Organization; 1977. [Google Scholar]
- 6.World Health Organization . International Statistical Classification of Diseases, Tenth Revision (ICD-10). World Health Organization; 1992. [Google Scholar]
- 7.Centers for Medicare & Medicaid Services . 2020. ICD-10 Procedure Coding System. Accessed September 1, 2023. https://www.cms.gov/medicare/coding-billing/icd-10-codes/2020-icd-10-pcs
- 8.Yasaitis LC, Berkman LF, Chandra A. Comparison of self-reported and Medicare claims-identified acute myocardial infarction. Circulation. 2015;131(17):1477-1485. doi: 10.1161/CIRCULATIONAHA.114.013829 [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Trial protocol
Statistical analysis plan
eTable. Event Counts by Data Source for 2,098 Participants Who Have EHR and Claims Data, With or Without Portal-visit Data
Data sharing statement