TF-GaitDT²: Trustworthy Personalized Federated Gait Twin Learning Under Heterogeneous Modalities for Parkinson's...

Gait impairment is a key motor manifestation of Parkinson's disease and can be assessed using wearable, plantar force, and vision-based sensing. In prac tice, however, clinical deployments are characterized by heterogeneous sensor availability and strict privacy con straints that preclude centralized data sharing. This paper proposes TF-GaitDT², a trustworthy and personalized fed erated digital-twin framework for multimodal PD gait anal ysis under partially observed sensing configurations. The framework integrates modality-specific encoders, masked multimodal fusion, and recurrent twin-state dynamics to learn a longitudinal representation of motor function while preserving data locality. Multi-task heads jointly support PD detection and severity estimation, with uncertainty calibration to improve reliability under non-IID federated settings. Experimental evaluation in a heterogeneous federated setting using public gait datasets shows that TF GaitDT2 improves upon standard federated baselines for PD detection and severity estimation, while approaching the performance obtained under centralized training. The proposed framework also yields more stable digital-twin representations and improved uncertainty calibration, with an associated increase in communication overhead. Over all, the results indicate that TF-GaitDT2 is suitable for privacy-preserving multimodal PD assessment in federated clinical settings.