Digital Twins (DTs) are emerging as a transformative paradigm in healthcare, promising personalized, predictive, and proactive care by creating dynamic virtual replicas of patients. However, their integration into clinical practice faces a fundamental "trust gap". Most AI models driving these twins are opaque "black-boxes" that provide predictions without auditable justifications or honest estimates of their own uncertainty. This paper introduces Trust-DT, a modular and model-agnostic technological framework designed to treat trust as a first-class engineering principle. The framework natively integrates two synergistic pillars: Explainable AI (XAI): Utilizing SHAP and LIME to provide transparent, clinically plausible rationales for every prediction. Uncertainty Quantification (UQ): Implementing Conformal Prediction (via the MAPIE library) to deliver statistically calibrated reliability estimates. We instantiate Trust-DT in a maternal health risk case study using the UCI Maternal Health Risk dataset. Our empirical evaluation reveals a critical discovery: ensemble models (Random Forest and XGBoost) optimized for high accuracy (87.19%) exhibit catastrophic miscalibration when audited. Specifically, at a 90% target confidence level, the models demonstrated an empirical coverage of only ~24%, diagnosing a dangerous level of overconfidence. Trust-DT proves that accuracy alone is an insufficient and misleading metric for clinical trust. By surfacing these hidden reliability gaps, the framework provides the necessary tools to bridge the gap between high-performance AI and safe clinical practice.
Pedro Miguel Alves Ribeiro (Fri,) studied this question.