IJCDS

Introduction

Recurrent stroke commonly follows an initial event, compounding disability and mortality risks. The risk of stroke recurrence is highest in the initial week following a minor non-cardioembolic stroke or TIA. This vulnerable period is a strategic target for secondary stroke prevention. Several major clinical trials, including SOCRATES [1], POINT [2], and THALES [3], have focused on this target with enhanced antiplatelet therapy, initiating treatment acutely within 12-24 hours following the ictus. These 3 trials, all testing addition or substitution of ticagrelor or clopidogrel to aspirin alone, share many similarities in trial structure, design, and core data elements, and show evidence of very similar temporal features of stroke recurrence [4], suggesting comparable clinical populations and risks. These similarities offer the opportunity for subject-level combination of the similar datasets for increased statistical power and generalizability. Combining these large data sets may provide the opportunity to detect which risk factors or medications affect the distinct processes that determine stroke recurrences over time. At the same time, there are some distinctions in definitions of clinical variables that call for careful harmonization in an effort at merging these datasets. Data harmonization, the practice of conforming various sources, types, and structures of data into formats that are compatible and comparable, can provide benefits in statistical power, generalizability, and data quality [5, 6]. As compared to aggregate data meta-analyses, harmonized datasets provide more opportunities for exploring interactions between clinical features and risk modifiers for outcomes of interest that may go beyond the primary outcomes reported in the original trials. Even individual participant-level meta-analysis, while similar in some ways, [7], usually take a narrower variable and outcome focus, and offer less flexibility of analytical questions and ready availability for repurposing and use by external researchers, than does a comprehensive harmonized merged data table, curated and webaccessible.  This project describes the approach to harmonization and merging of the 3 large datasets from the SOCRATES, POINT, and THALES trials, and the features of the resulting large merged dataset.

Methods

The studies- similarities and distinctives

The selected studies each targeted subjects with acute noncardioembolic minor ischemic stroke or high-risk TIA, randomizing subjects within 12-24 hours of index events. Each compared a form of enhanced anti-platelet therapy to treatment with aspirin alone.  There were minor differences in the definition of minor stroke in the studies. In SOCRATES and THALES, minor stroke was defined by NIHSS score of 5 or less, while in POINT it was restricted to NIHSS of 3 or less. The ABCD2 score, used to define the risk level of TIA, was required to be 4 or more in SOCRATES and POINT, and 6 or more in THALES. The time window for randomization following the qualifying ischemic event was within 24 hours of onset of symptoms in SOCRATES and THALES, and within 12 hours of the last time known to be free of new symptoms in POINT. The chosen enhanced anti-platelet regimen and the dosing of aspirin differed between the studies, with SOCRATES comparing ticagrelor 90 mg twice daily (following a loading dose of 180 mg) to aspirin 100 mg daily (following a loading dose of 300 mg); POINT comparing clopidogrel 75 mg daily (following loading doses of 600 mg) to placebo, in addition to aspirin 50 mg – 325 mg daily (following a loading dose of 325 mg); and THALES comparing ticagrelor 90 mg twice daily (following 180 mg loading dose) to placebo, added to aspirin 75 mg to 100 mg daily (following loading dose of 300 mg to 325 mg). Each of the trials were randomized with a 1:1 assignment of subjects to treatment groups, and each was double- blinded with placebo controls. SOCRATES and POINT continued study treatment and follow up for 90 days, while THALES continued study medication treatment for 30 days, with subject follow-up for up to 60 days.

Datasets

The POINT trial dataset was provided by the National Institute of Neurological Disease and Stroke, obtained from the NINDS data repository as a set of 22 de-identified SAS-compatible SYSdBat files. The datasets of the SOCRATES and THALES trials, from data contributor AstraZeneca, comprised 23 files and 13 files respectively, and were made available through Vivli, Inc. on a secure server as de-identified SAS-compatible SYSdBat files. Data encoding followed the ADaM (Analysis Data Model) standardized system.

The shared datasets included 4,881 subjects from the POINT trial, 12,018 subjects from the SOCRATES trial, and 9,565 subjects from the THALES trial, totaling 26,464 subjects (Figure 1). Data were not shared for subjects who had not given or had withdrawn consent (113 subjects in SOCRATES and 20 subjects in THALES) nor for subjects from countries not allowing data re-use based on the consent wording (1,182 subjects in SOCRATES and 1,489 subjects in THALES). From the shared datasets, 101 subjects from the SOCRATES trial and 44 subjects from the THALES trial were not assigned to a treatment group and largely lacked any clinical data, and these subjects were excluded from the final merged dataset. Intention-to-treat populations, definitions of outcome events, designation of event times, and data censoring methods were defined according to the trial protocols [1-3]. The POINT dataset served as the default harmonization standard for most clinical and outcome variables, as no participants were excluded from the shared POINT dataset and it was provided with an organized explanatory data dictionary.  Data assembly and analysis was carried out using SAS version 9.4 in the Vivli research environment.

Figure1: Flow diagram of subjects included in the merged dataset.

General Approach to Harmonization

Existing clinical trial datasets were pooled and harmonized through standardized procedures to create a single integrated dataset designed to support future analyses with increased sample size and statistical power. As a foundational step, a target data schema was developed defining standardized variable names, data types, coding conventions, and allowable values. This schema served as the reference framework for mapping and transforming variables from each source dataset, ensuring consistent alignment and integration across the pooled trials. Across the contributing clinical trials, a number of core variables were identified that used broadly similar clinical concepts but differed in naming conventions, coding schemes, measurement, and visit structures. Prior to harmonization, all source datasets were systematically reviewed to identify common variables and assess differences in definitions, formats, and allowable values. While key domains such as demographic characteristics, clinical outcomes, and outcome timing were represented across studies, variability existed in variable labeling, categorical encodings, and measurement structures. These differences informed the development of harmonization rules used to map study-specific variables to the standardized target data schema.

Harmonization strategies - Demographic and clinical data

Clinical data including demographic information, medical history, and baseline evaluations, as recorded in each of the trials, were culled from the trial data sets. Some potentially identifying data such as data of birth, dates of randomization or study visits, and country of origin were excluded in the de-identification process by the data owners prior to provision in all 3 datasets. Specific subject ages were not provided in the SOCRATES and THALES trials; instead, age ranges were specified. Specific ages (in years) up to 89 were provided in POINT. Body mass index (BMI) values were only provided as ranges in SOCRATES and THALES, and were not provided in the POINT dataset.

Unknown data values were of two types. Certain variables were simply not recorded in a given trial. Examples include smoking history, not included in SOCRATES, or history of prior stroke, not included in POINT. Such variable values were considered ‘not reported’ and in the harmonization process were designated as such by blank values for character variables and by entry of ‘-999’ for numeric variables (known values for these variables are all non-negative). Other variable values were within the scope of a study’s protocol but not found; these were designated with the entry ‘.’ to indicate ‘missing’ data values.

Specific flags for certain medical history conditions, such as history of hypertension, history of diabetes, were recorded in all 3 trials. Other medical history conditions were discoverable in SOCRATES and THALES trials by search of the provided medical history files for listing of the specific condition name; if not found listed for a subject, the history of that medical condition was considered not present and in harmonization the corresponding flag was imputed a value of ‘N’.

Harmonization of Medication data

All 3 trials included flags for aspirin exposure immediately prior to randomization. Prior exposures to certain other antiplatelet agents, to statins, and to other lipid-lowering medications were indicated by specific flags in POINT and were available in the concomitant medication data files for SOCRATES and THALES. The POINT trial also provided a limited information about exposure on the dates of each study visit to statin medications, to NSAIDS, to oral anticoagulant medications, and to proton pump inhibitors, but did not provide start and end dates regarding these medications, and did not report other concomitant medication use. SOCRATES and THALES in contrast included detailed concomitant medication data files with comprehensive reporting of all medications used during the trial by each subject, with specific medication names and Anatomic Therapeutic Chemical (ATC) codes, start and end dates, and dosages.

Twenty medications or medication groups of interest were selected for inclusion in the harmonized merged database. Searches for these concomitant medication exposures in SOCRATES and THALES datasets were generally based on ATC codes, with certain exceptions (see Table 1). In certain cases, medication names were used. For example, cilostazol use was pulled by name, to separate it from other antiplatelets included under the same ATC code. For prior antiplatelet use variables, searches were by medication names.

Table 1: Search and data summarizing strategies for concomitant medications. Full variable names are given for flags for prior antiplatelet medication use. For other medication variables, a 3 letter code designates the medication group, and a 2 letter suffix defines the parameter pertaining to that medication that the variable describes.

Several challenges arose in characterizing the detailed concomitant medication exposure information in a data table format with a single row assigned for each subject. Challenges included multiple exposures to different medications of the same class (i.e., different statins), multiple short-term exposures with different start and stop dates, overlapping exposure periods, and missing data for start and stop dates. This complex information was condensed into a set of 10 harmonized variables for each medication group summarizing the exposures, including start and end days, number of exposures, and total exposure duration. Details of the medication -related variables and of the naming conventions followed can be found in the explanatory READ ME file included in Supplementary materials.

Harmonization of Outcome Events

A wide variety of pre-specified outcome events was available for each study. Flags for each study’s primary outcome event, which was defined differently for the 3 studies (any stroke, myocardial infarction, or death for SOCRATES; ischemic stroke, myocardial infarction, or ischemic vascular death for POINT; and any stroke or death for THALES), were mapped to a single merged primary outcome event flag. Ischemic stroke, the most commonly occurring outcome event, was similarly defined in each trial and mapped to a single event flag. Multiple other important outcomes were tracked in all 3 trials, though with some minor differences of definition, and were mapped to common harmonized event flags, including flags for hemorrhagic stroke, any stroke, death, and major hemorrhage events. A total of 20 different outcome events common to at least 2 of the trials were assigned to harmonized event flags and event time variables, following the convention of the SOCRATES and THALES trials with failure events coded as ‘0’ and censored subjects assigned ‘1’. The harmonized timing flags were assigned values of the nearest number of days elapsed from day of randomization, which was assigned a value of ‘0’. This required rounding to the nearest whole day of the timing variable in the POINT trial, which were recorded as decimal numbers of days, and subtraction of 1 day in the timing variables of outcomes in the SOCRATES and THALES trial datasets, which followed a convention calling the day of randomization as day 1.

Further details of harmonization conventions followed and variable naming conventions are described in the README file that is included in supplementary materials and on the Vivli.org platform. Full descriptions of each variable and how the harmonized variable relates to the source data of the individual trial datasets are found in the SPT Data Dictionary, also included in the supplementary materials and on the Vivli platform.

Results

Harmonization of demographic and clinical variables

The harmonized merged data table includes a total of 26,319 subjects, comprising 90% of the total numbers of subjects enrolled in the three trials. A total of 311 variables were included, as described in Table 2.

Table 2: Summary of variables included in SPT datatable: Detailed descriptions of all included variables are available in the SPT Data Dictionary, included among the supplementary materials

A high level of data completeness was achieved, with numbers of subjects with known variable values exceeding 80% for 55 of 75 demographic and clinical variables, for 30 of 40 outcome event variables, and for 111 of 196 concomitant medication parameters (Figure 2, and Table 3). Concomitant medication data were not collected in the POINT trial (save for antiplatelet agents and statins), preventing a high level of >80% completeness for most of these variables. Other notable gaps in collected data included lack of baseline laboratory data in the SOCRATES and THALES datasets, lack of smoking status and of carotid or intracranial artery stenosis data in the SOCRATES trial dataset, and lack of end-of-study vital signs in POINT.

Figure 2: Data completeness. Numbers of included variables, and variable completeness at 50%, 80%, and 90% thresholds.

Table 3: Data Completeness metrics – Numbers of variables with known values above 50%, 80%, and 90% thresholds

Concomitant medication usage

Exploratory studies of effects of concomitant medications on stroke-related outcomes are of great interest. Such analyses will require adequate exposures of subjects to the medications in question. Exposures (based on the concomitant use variable) varied considerably across medication groups (Table 4). As expected, most patients (64%) in the trials were treated with statin medications but only a small minority were treated with other lipid-lowering medications. Most patients were on at least one antihypertensive medication, with angiotensin -converting enzyme (ACE) inhibitors or angiotensin receptor blocker (ARB) medications, beta-blocking medications, and Calcium channel-blocking medications being the most common. Many were on medications for diabetes, but glucagon-like peptide-1 (GLP-1) agonist use was rare during these trials. A number of subjects were exposed to NSAIDs, but few were treated with other antiinflammatory medications. A large number of subjects were concurrently treated with nootropic medications.

Table 4: Concomitant medication use: Completeness and exposure levels for the different medication groups for subjects included in the SPT merged datatable

Outcome events: completeness and event numbers

Most outcome events were included in data recording in all 3 trials, resulting in data that are complete or near-complete for most outcomes of interest. All 3 trial data sets tracked most of the same outcome events of interest including stroke subtypes, both ischemic and hemorrhagic, major hemorrhages, and fatal events. The THALES trial dataset did not report outcomes of ‘Any Stroke’, TIA’, “MI’, or ‘Vascular death’, so these outcomes are not reported for the 9521 subjects from this trial. On-treatment status was not flagged in 108 of the 4881 subjects in the POINT trial, so that the As-treated outcomes for these subjects are recorded as missing data. Event numbers, crucial for statistical power estimation for any proposed analyses, were moderately high for outcomes that included ischemic stroke, but were only modest or low for most other outcomes of interest. (See Table 5).

Table 5: Outcome events included, with incompleteness and event numbers (ITT- intention to treat; AT-as treated)

Data curation and availability.

The harmonized merged dataset, as well as an explanatory README file, the code used for its production, the individual trial harmonized data tables, metadata, and an extensively annotated data dictionary, are all housed on the Vivli organization (vivli.org) platform, and are available to qualified investigative teams for analysis. The platform allows users access to common software tools including SAS, R, and Excel, and provides for uploading of users’ code into the platform. Access is available through application to Vivli as vetted by the Vivli Independent Review Panel and by approval of the data provider (for SOCRATES and THALES) Astra-Zeneca.