Abstract
Introduction/Aim: The aim of the work is to present a methodological framework for a ten-year prediction of multimorbidity using a generative, sequential model based on transformer architecture, with full reliance on publicly available and synthetic data, without processing identifiable patient data.
Methods: A methodological study was conducted with primary training and evaluation on the Synthea synthetic cohort (≥100,000 records), with technical checks on MIMIC-III (de-identified real intensive care data). Outcomes (≥1,000) were defined by mapping ICD-10/ICD-10-CM codes to PheCode categories. The model is a transformer with multitask outputs (one for each target) and time embeddings, with evaluation of discrimination (AUPRC, AUROC), calibration (Brier, intercept/slope) and clinical utility (Decision Curve Analysis). Sensitivity analyzes and basic fairness checks (gender, age) were conducted.
Results: The model achieved the best results in cardiometabolic and oncological domains, moderate in respiratory/renal, and more modest in mental/infectious outcomes. Calibration was good in the intermediate risk ranges; DCA showed a positive net benefit at thresholds relevant for opportunistic screening (≈5–15% 10-year risk). Sensitivity analyzes confirmed the stability of ranking performance across changes in rarity thresholds and history length, with no evidence of significant label leakage.
Conclusion: A reproducible and ethically acceptable approach to long-term multi-disease risk prediction using a generative transformer on public/synthetic sets is presented. This "health barometer" can support triage and personalized prevention, while recommending mandatory external validation, local re-calibration and equity monitoring prior to clinical application.