Explainable Digital Twin: Self-learning Cardiovascular risk prognosis through Explanations-as-AI framework with longitudinal Echocardiography

Abstract

The Human Digital Twins (HDTs) are virtual representations of real humans that reflect their physiological conditions. However, traditional HDTs lack the transparent decision-making mechanisms required in critical cardiovascular risk contexts. Cardiovascular disease is the leading cause of early death throughout the world. A reliable and trustworthy system forms the foundation of proactive risk stratification. The proposed “Explanations-as-AI” framework is an enhancement over the conventional Explainable Artificial Intelligence (XAI). XAI, combined with supervised learning models, builds an Explainable Digital Twin (XDT) to support adaptive cardiac risk prognosis. The Shapley Additive exPlanations (SHAP) serve as an active computational method that provides dynamic feedback for self-learning and adaptive risk prediction. The 3-layer XDT architecture leverages longitudinal clinical and echocardiographic measures from Touch Hospital in Telangana, India. The Data layer handles data preparation, and the Computation layer hosts the risk-prediction classifier. The Trust layer converts SHAP values into Bayesian feedback with uncertainty weights. The XDT performance is benchmarked with baseline prediction without feedback, Closed-loop feedback, and Direct XAI feedback. State-of-the-art comparisons include deep sequential classifiers, explanation-regularized models, and various versions of Reinforcement Learning (RL). XDT achieves an AUC-ROC of 0.843 (7.8% improvement over baseline), a Brier score of 0.115 (18.9% reduction), and a 27.7% improvement in model calibration with rapid convergence after 4 iterations and reduced circular dependencies (mutual information: 0.38 to 0.08 bits). The proposed XDT outperforms RL methods empirically and establishes itself as a more ethically appropriate tool for cardiac risk prediction in the context of limited episodic clinical data. SHAP analysis reveals that ventricular dimensions and age are the primary drivers. This research aligns closely with the United Nations’ Sustainable Development Goals 3.4, 9, and 10, using the novel XDT paradigm to enable scalable, interpretable, and personalized cardiac risk prediction.