Researchers developed a Transformer-based survival model to predict long-term cardiovascular disease (CVD) risk using longitudinal health checkup records. The development cohort comprised 100,056 individuals without baseline CVD drawn from the Hokuriku Health Service Association (2010–2024). During follow-up, 4,113 incident CVD events (4.1%) were observed. An external evaluation cohort consisted of 79,756 participants from Kanazawa City, in which 21,179 CVD events (26.6%) occurred and were used to assess model generalizability.
The Transformer model was trained on routinely collected health checkup variables: anthropometric measurements, laboratory data, electrocardiogram findings, medication use, and self-reported lifestyle items. Incident CVD was defined as the first self-reported physician diagnosis of heart disease or stroke during up to 10 years of follow-up and was modeled as right-censored survival data.
The study compared the Transformer against multiple benchmarks: Cox proportional hazards regression, XGBoost using survival embeddings, a multilayer perceptron, and two established clinical risk scores—the Framingham Risk Score and the Hisayama Risk Score. Primary performance metrics were time-dependent ROC-AUC focused on the 10-year horizon, precision–recall area under the curve (PR-AUC), and concordance index (C-index). Interpretability was evaluated using SHapley Additive exPlanations (SHAP) and a Feature-level Attention Network (FAN).
In the development cohort, the Transformer produced the best internal discrimination among tested methods. Reported internal performance at 10 years was 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). These values indicate the model achieved higher time-dependent discrimination compared with the conventional statistical and machine-learning comparators evaluated in the study.
When applied to the Kanazawa City external cohort, the Transformer retained substantial predictive performance. External metrics reported were ROC-AUC 0.762, PR-AUC 0.500, and C-index 0.744. The preserved performance in an independent data set supports the model’s robustness across samples drawn from routine health checkup populations.
To probe model behavior and feature contributions, the authors combined global and attention-based approaches. SHAP values were calculated and the top 12 SHAP-ranked features were visualized to highlight key predictors and their marginal contributions. In addition, the Feature-level Attention Network (FAN) within the Transformer architecture was analyzed to reveal interactions and how features attend to each other across timepoints and modalities.
SHAP identified age, electrocardiogram abnormality, use of antihypertensive medication, and sex as the most important predictors of incident CVD in the model. The FAN analysis further clarified that self-reported lifestyle factors—specifically daily exercise and weight gain—carry prognostic information and modify the model’s assessment of age-related risk. Within the attention network, age functioned as a central hub linking behavioral habits with physiological measurements, illustrating how the model integrates multiple variable types to adjust risk estimates.
The authors report that a Transformer-based survival model trained on routine health checkup data can outperform conventional risk scores and statistical models for 10-year CVD prediction in these Japanese cohorts. The findings emphasize that self-reported lifestyle variables available in standard checkups can provide incremental prognostic value when integrated with physiological and laboratory data by a model capable of capturing nonlinear and time-dependent relationships. These results suggest potential utility for population-level risk stratification and personalized preventive strategies based on routinely collected data.
This work is a medRxiv preprint and has not undergone peer review. The study protocol received approval from the Medical Ethics Committee of Kanazawa University (Research Protocol Number: 2019-080 [113122]), and the requirement for informed consent was waived due to the retrospective use of deidentified secondary data. The authors state that the data are not publicly available because of potentially identifiable health information and data-use agreements; access requires approval from the data providers and relevant ethics committees. The manuscript’s scope, results, and limitations reflect the content reported in the preprint; no additional methodological or numerical details beyond those presented in the source were invented or inferred here.
In summary, within the cohorts studied, the Transformer-based survival model achieved superior discrimination for long-term CVD risk compared with standard statistical and machine-learning benchmarks. Interpretability analyses reinforced the importance of established clinical predictors (age, ECG abnormality, antihypertensive therapy, sex) and demonstrated that routinely collected self-reported lifestyle measures meaningfully interact with physiological features to influence predicted risk. Because this is a preprint, further peer review and external replication are necessary before clinical implementation.