PREDICTING FUTURE CARDIOVASCULAR ADVERSE EVENTS AT POPULATION SCALE USING A REAL-WORLD DARA FOUNDATION MODEL: IMPLICATIONS FOR CLINICAL TRIAL AND LAUNCH PLANNING
Author(s)
Richard Gliklich, MD, Ashish Deshpande, MD, Ombretta Palucci, MS, Costas Boussios, PhD.
OM1, Inc., Boston, MA, USA.
OM1, Inc., Boston, MA, USA.
OBJECTIVES: Cardiovascular adverse events - including MI, stroke, heart failure, atrial fibrillation, coronary artery disease, and revascularization - are critical safety endpoints in clinical trials and pharmacovigilance. Population-level prediction from real-world data (RWD) could enable sponsors to characterize CV risk, improve trial feasibility, and support safety planning. We evaluated PhenOM, a foundation model trained on billions of longitudinal patient records, for predicting major adverse cardiovascular events (MACE) in a US real-world population.
METHODS: PhenOM was applied to the OM1 Real World Data Cloud, a de-identified US longitudinal dataset encompassing more than 370 million patients. Seven new-onset CV outcomes were evaluated - acute MI, ischemic stroke, hemorrhagic stroke, heart failure, atrial fibrillation, coronary artery disease, and coronary revascularization - each with a 12-month prediction horizon. PhenOM generates patient-level risk scores from full longitudinal event sequences. Performance was assessed by AUROC, calibration slope, and lift at the top 10%, 1%, and 0.1% risk thresholds (TRIPOD+AI), stratified by age cohort and census region.
RESULTS: PhenOM demonstrated robust discrimination across all seven outcomes, with AUROCs ranging from 0.73 (Hemorrhagic Stroke) to 0.87 (Heart Failure), median 0.79. Calibration slopes were close to 1.0, indicating well-calibrated absolute risk estimates. At the top 1% threshold, lift ranged from 8-fold (CAD) to 15-fold (Revascularization) above baseline, enabling substantial enrichment through model-guided stratification. Performance was consistent across census regions and age strata (18-64 and 65+ years). These results support reductions in sample size and monitoring burden for trial safety planning
CONCLUSIONS: PhenOM provides well-calibrated, population-scale prediction of cardiovascular adverse events from RWD with clinically meaningful discrimination and high lift at narrow risk thresholds. Key applications include pre-trial CV risk characterization, high-risk subgroup identification for safety monitoring, background event rate estimation for external comparator arms, and pharmacovigilance population sizing - supporting a scalable, RWD-based complement to traditional approaches for trial planning and launch readiness
METHODS: PhenOM was applied to the OM1 Real World Data Cloud, a de-identified US longitudinal dataset encompassing more than 370 million patients. Seven new-onset CV outcomes were evaluated - acute MI, ischemic stroke, hemorrhagic stroke, heart failure, atrial fibrillation, coronary artery disease, and coronary revascularization - each with a 12-month prediction horizon. PhenOM generates patient-level risk scores from full longitudinal event sequences. Performance was assessed by AUROC, calibration slope, and lift at the top 10%, 1%, and 0.1% risk thresholds (TRIPOD+AI), stratified by age cohort and census region.
RESULTS: PhenOM demonstrated robust discrimination across all seven outcomes, with AUROCs ranging from 0.73 (Hemorrhagic Stroke) to 0.87 (Heart Failure), median 0.79. Calibration slopes were close to 1.0, indicating well-calibrated absolute risk estimates. At the top 1% threshold, lift ranged from 8-fold (CAD) to 15-fold (Revascularization) above baseline, enabling substantial enrichment through model-guided stratification. Performance was consistent across census regions and age strata (18-64 and 65+ years). These results support reductions in sample size and monitoring burden for trial safety planning
CONCLUSIONS: PhenOM provides well-calibrated, population-scale prediction of cardiovascular adverse events from RWD with clinically meaningful discrimination and high lift at narrow risk thresholds. Key applications include pre-trial CV risk characterization, high-risk subgroup identification for safety monitoring, background event rate estimation for external comparator arms, and pharmacovigilance population sizing - supporting a scalable, RWD-based complement to traditional approaches for trial planning and launch readiness
Conference/Value in Health Info
2026-11, ISPOR Europe 2026, Vienna, Austria
Value in Health, Volume 29, Issue 12S
Code
RWD117
Topic
Real World Data & Information Systems
Disease
Cardiovascular Disorders (including MI, Stroke, Circulatory)