MFSI-Net: a weakly supervised framework for continuous estimation of expert-annotated functional states from multisource wearable signals in type 2 diabetes

Continuous assessment of functional states from wearable signals is challenging because dense longitudinal signals are paired with sparse expert annotations. We developed MFSI-Net, a weakly supervised framework that estimates a study-specific continuous functional-state trajectory aligned with expert-defined categories in adults with type 2 diabetes. The framework combines fold-contained dependency-aware feature screening, Gaussian-process-derived soft-label generation, and a GRU–Transformer temporal model. In a 28-day observational study of 90 participants, the model showed agreement with GPR-derived soft labels (RMSE 0.17, Pearson r 0.72; temporal label-availability sensitivity: RMSE 0.22, r 0.61). These values quantify internal agreement with model-generated soft labels rather than agreement with an independent clinical reference. Label-associated group and scale analyses are therefore interpreted as exploratory within-cohort evidence. MFSI-Net provides a methodological approach for studying continuous trajectories derived from sparse expert annotations; independent clinical, physiological, and external validation remains necessary.