Transformer-Based Survival Model for Cardiovascular Risk Prediction from Longitudinal Health Checkup Data

Background: Cardiovascular disease (CVD) is a leading global health concern. Traditional models often miss nonlinear dependencies among physiological and behavioral factors. We hypothesized that a Transformer-based deep learning model, which excels at capturing complex patterns in structured data trained on large-scale health check-up records, would improve long-term CVD risk prediction. Methods: We analyzed longitudinal health records (2010?2024) from the Hokuriku Health Service Association (n = 100,056 without baseline CVD; development cohort). Incident CVD was defined as the first self-reported physician diagnosis of heart disease or stroke during the 10-year follow-up and was modeled as right-censored survival data. An external evaluation cohort comprised 79,756 Kanazawa City participants with health records. A Transformer model was trained using anthropometric, laboratory, and self-reported lifestyle data. Benchmarks included Cox regression, XGBoost survival embeddings, multilayer perceptron, the Framingham Risk Score, and the Hisayama Risk Score. Performance was evaluated using time-dependent area under the receiver operating characteristic curve (ROC-AUC) with a primary focus on the 10-year ROC-AUC, precision?recall AUC (PR-AUC), and concordance index (C-index). Interpretability was assessed through SHapley Additive exPlanations (SHAP) and a Feature-level Attention Network (FAN), visualizing the top 12 SHAP-ranked features to highlight key interactions. Results: In the development cohort, 4,113 CVD events (4.1%) occurred. The Transformer model achieved the best internal performance: 10-year ROC-AUC 0.821 (95% confidence interval [CI], 0.816?0.826), PR-AUC 0.427 (CI, 0.419?0.435), and C-index 0.781 (CI, 0.775?0.787). Performance remained robust externally (21,179 CVD events, 26.6%): ROC-AUC, 0.762; PR-AUC, 0.500; and C-index, 0.744. Regarding interpretability, SHAP identified age, electrocardiogram abnormality, antihypertensive medication, and sex as the most critical predictors. Notably, FAN elucidated the prognostic value of self-reported lifestyle factors. For example, daily exercise and weight gain modulated the model?s assessment of age-related risk. Within the attention network, age served as a central hub, linking these behavioral habits with physiological features. Conclusion: The Transformer-based model outperformed conventional methods in predicting long-term CVD risk. Model interpretation demonstrated the predictive utility of self-reported lifestyle factors, such as weight gain and daily exercise. These findings may support personalized CVD prevention and population-level risk stratification using routinely collected health checkup data.