Presentation (CTI)

OBJECTIVES: Stroke is difficult to predict using single-time-point risk measures. We developed a Bayesian path-based transition-state framework to estimate future stroke risk from long-term health screening and claims histories and support preventive intervention planning.
METHODS: We used Korean National Health Insurance claims and health screening data from 2002-2020. Incident stroke cases were individuals with first stroke in 2019-2020 and no stroke-related outpatient or inpatient diagnosis during the previous 10 years (n=111,895). Non-stroke controls (n=111,895) were matched by sex, one-year age, and insurance premium level. Longitudinal histories were divided into eight biennial periods. In each period, health-management states were defined using eight components: uncontrolled hypertension, uncontrolled diabetes, dyslipidemia history, heart disease history, kidney disease history, non-participation in screening, obesity or insufficient physical activity, and current smoking or ≥20 pack-years. Individuals with multiple components were assigned to composite states reflecting all applicable conditions. Sex-specific transition matrices were constructed separately for cases and controls. Smoothed case/control transition probabilities were estimated for each observed state-to-state transition, transformed into log-likelihood ratios, stabilized using count-based shrinkage and occupancy terms, accumulated across periods, and converted into Bayesian posterior stroke probabilities using a population prior risk.
RESULTS: In an extreme-risk contrast analysis, the framework showed good discrimination for distinguishing individuals in the highest versus lowest predicted-risk deciles. The AUC was 0.7343 in women and 0.7545 in men. The model also enables scenario-based estimation; for example, recent 6-year health-management histories can be combined with alternative 4-year future scenarios to identify modifiable risk pathways that may influence future stroke risk.
CONCLUSIONS: This Bayesian transition-state framework uses longitudinal real-world health histories to estimate future stroke risk and identify modifiable prevention pathways. Further calibration, external validation, and evaluation of implementation utility are warranted.