Introduction – Prediction of sudden cardiac death (SCD) using traditional clinical data has yielded limited success in risk stratification. Although multiple data modalities may associate with outcomes, their overlap can be substantial, and the independent contribution of each source should be quantified.
We analyzed a prospective clinical exercise testing cohort (FINCAVAS; n=2861) to assess the contribution and overlap of multiple data modalities in SCD prediction. SCD events (n=75) were adjudicated by in-depth review of medical records and death certificates. Contributions of different data domains were evaluated using a multimodal modeling framework by fitting Cox proportional hazards models to time-to-event data. Domain contributions were quantified using Shapley decomposition across all domain combinations with model R² as the value function. Domain overlap was derived from differences in model R² between individual and combined domain models. Preselected domains included demographic, physiological (and anthropometric), dynamic exercise, laboratory, medication, genetic (polygenic risk scores based on published genome-wide association studies), and electronic health record data (10–30 variables per domain).
Cross-validated are under curve for the full model was 0.79. Most predictive information was captured by dynamic exercise, demographic, and genetic data, accounting for 29%, 28%, and 17% of total contribution. Medication data and physiological measures contributed moderately (12% and 11%), whereas laboratory variables added only 3.6%. Overlap between domains was generally low (−0.001 to 0.018), with modest redundancy between demographic and exercise data. Genetic data showed minimal overlap with all other domains, indicating largely orthogonal information not captured by clinical or physiological variables. Limiting variables to the five most significant predictors per domain did not materially change the results. Adding left ventricular ejection fraction (available in 71%) increased the contribution of physiological measurements only from 11% to 13%.
Conclusion – Exercise data and genetic information provide the most independent predictive value beyond clinical demographics.