Stroke is the fifth most prevalent cause of death in the U.S., afflicting nearly 800,000 per year. About three quarters of strokes are first events, underscoring the importance of primary prevention. Designing optimal preventive strategies requires identification of risk factors and estimation of the risk of stroke. However, the most recent American Heart Association (AHA)/American Stroke Association Guidelines for the Primary Prevention of Stroke conclude that “an ideal stroke risk assessment tool that is simple, is widely applicable and accepted, and takes into account the effects of multiple risk factors does not exist.” Although multiple stroke risk prediction models have been developed, their generalizability is limited for a number of reasons, including that they rely on data from one study, focus on any stroke, do not differentiate between different etiologies, and only incorporate a small number of predictors.

In this NIH-supported project, researchers from Duke AI Health and the American Heart Association aim to address these limitations and enhance stroke risk prediction modeling by aggregating and harmonizing existing patient-level data collected as part of four longitudinal cohort studies. This broad range of data from different patient cohorts strengthens the evidence for the scientific questions being investigated and, as part of the study, we have introduced an open metadata repository designed to give researchers access to the processes and methods used to harmonize the datasets. In addition, the research team is developing and validating a new machine learning-based risk prediction model for primary stroke and comparing its performance with existing models. The model development includes adding new risk factors, including genetic information, and applying clustering of individuals for the prediction of stroke etiology (ischemic versus hemorrhagic).

This research is supported by NIH/NINDS R61-NS120246.