Personalized Stroke Rehabilitation via Stratified Interpretable Modeling With Wearable IMUs

Accurate assessment of post-stroke upper-limb impairment using the Fugl-Meyer Assessment of the Upper Extremity (FMA-UE) is vital for rehabilitation, but clinical administration is burdensome and infrequent. Recent studies have attempted to automate FMA-UE scoring with wearable sensors, yet most use a "one-size-fits-all" model and lack clinical interpretability. We propose a novel framework that integrates patient stratification and interpretable modeling for FMA-UE estimation using Generalized Additive Model (GAM). Kinematic features extracted from two Inertial Measurement Units (IMUs) on the affected wrist and trunk when performing four standardized upper-limb tasks are used in the experiment. K-means clustering is applied to stratify participants into "Mild" and "Moderate" impairment subgroups based on extracted movement features. This framework, hereafter referred to as Cluster-specific GAM (CGAM), is trained for each subgroup to predict FMA-UE scores while maintaining model transparency. We evaluate the model using Leave-One-Out Cross-Validation (LOOCV) on data from 30 stroke patients. The CGAM achieves a low estimation error (RMSE = 5.79, R ${^{{2}}} =0.75$ ) and outperforms a non-stratified global model (RMSE = 6.78, R ${^{{2}}} =0.66$ ) as well as standard regression models. Importantly, the CGAM's intrinsic interpretability provides insight into how specific movement features influence the FMA-UE score differently for Mild vs. Moderate groups. This work demonstrates a feasible solution for remote, personalized stroke assessment that addresses patient heterogeneity and provides transparent decision support for rehabilitation planning.