FedEC-Dia: Explanation-Calibrated Federated Learning for Patient-Level Diabetes Digital Twin Risk Prediction Under Non-IID Healthcare Data.

Patient-level digital twins for diabetes oriented smart healthcare require risk models that can learn from distributed clinical data while remaining reliable across heterogeneous environments. Federated learning enables collaborative training without centralizing raw patient records, but non-independent and identically distributed (non-IID) data can cause weak-client prediction failure and inconsistent feature-use patterns. Site-specific differences in disease prevalence, clinical measurements, and risk profiles can therefore undermine stable patient state assessment. We propose FedEC-Dia, an explanation calibrated federated learning framework for digital-twin oriented diabetes risk prediction. FedEC-Dia uses a lightweight client-specific affine calibration layer to adjust local risk scale and bias, and constructs differentiable gate based attribution summaries that are softly regularized toward a server-maintained global explanation prototype. Explanation calibration here denotes training-time regularization of gate-based attribution summaries rather than causal clinical explanation. Controlled simulated federated validation on PIMA and MIMIC-III covers label-, feature-, quantity-, and mixed-skew settings. FedEC Dia improves weak-client performance and probability calibration while producing more stable gate-based attribution summaries, supporting its role as a reliability oriented risk-estimation component for distributed patient-level diabetes digital twins.