CardioTwin-AI: development and internal validation of a digital twin-based adaptive risk-aware learning framework for heart disease prediction

Heart disease is one of the most common causes of death in the world and early identification of individuals at elevated cardiovascular risk requires reliable risk-assessment methods that can support clinical decision-making. Many conventional machine-learning models focus primarily on disease classification or risk prediction and do not explicitly provide patient-specific virtual representations. This study introduces CardioTwin-AI, a Digital Twin-based Adaptive Risk-Aware Learning Framework (ARLF) to predict the presence or absence of heart disease. We propose a data-driven, patient-specific Digital Twin representation based on clinical attributes, in which the patient state is mapped to a corresponding virtual state and evaluated using state-level synchronization measures, including Twin Simulation Error (TSE) and Twin Fidelity Score (TFS) to quantify representation consistency. The deep feature representations are generated by feature engineering process based on autoencoder, and then analysed by the ARLF prediction model to estimate the probability of the presence of heart disease. It achieved high patient-state representation consistency between the source clinical state and its virtual representation, with a Twin Fidelity Score of 99.76% and a Twin Simulation Error of 0.002402. The proposed ARLF model achieved accuracy of 90.21%, precision of 90.32%, recall of 94.11%, F1-score of 92.18%, and ROC-AUC of 95.13% in the fixed held-out evaluation. The results indicate the potential of CardioTwin-AI for predicting the presence or absence of heart disease and assessing patient-state representation consistency, subject to further external validation and prospective clinical evaluation.