An In-Silico Investigation of Machine Learning for Integrating Genomic and Digital Biomarker Data in Cardiovascular Risk Stratification

Conventional models for stratifying cardiovascular disease (CVD) risk have limitations. The integration of static genomic data and dynamic digital biomarkers from wearable technology holds theoretical promise, but its potential quantitative impact remains poorly defined. This study aimed to develop and validate an in-silico framework to quantify the theoretical maximum predictive gain of an integrated risk model under idealized conditions. We developed a sophisticated data generating process (DGP) to create a synthetic dataset of 5,000 individuals. The DGP incorporated demographic and clinical variables with distributions and correlations based on epidemiological literature. It included a simulated polygenic risk score (PRS) for coronary artery disease and advanced digital biomarkers derived from wireless health monitoring data, such as heart rate variability (HRV) and time in moderate-to-vigorous physical activity (MVPA). The 10-year risk of Major Adverse Cardiovascular Events (MACE) was generated via a defined logistic function incorporating these variables plus stochastic noise. We compared the performance of the ACC/AHA Pooled Cohort Equations (PCE) against several machine learning models (Logistic Regression, Random Forest, XGBoost) using the area under the receiver operating characteristic curve (AUC-ROC), precision, recall, and F1-score. In this simulated environment, the integrated XGBoost model achieved near-optimal predictive performance with an AUC-ROC of 0.92 (95% CI, 0.90-0.94), significantly outperforming the benchmark PCE model (AUC-ROC 0.76; 95% CI, 0.73-0.79; p < 0.001). The inclusion of the PRS and, most notably, dynamic digital biomarkers like HRV, provided substantial incremental improvements in risk discrimination over traditional factors alone. In conclusion, this in-silico study demonstrates the substantial theoretical potential of integrating genomic and advanced digital biomarker data through machine learning for CVD risk stratification. While these idealized results are not directly generalizable, they provide a quantitative rationale for pursuing real-world data collection and validation studies. This work establishes a methodological proof-of-concept and highlights the potential for a paradigm shift toward more dynamic and personalized cardiovascular risk assessment.