Phase-specific multimodal biomarkers enable explainable assessment of upper limb dysfunction in chronic stroke

Background Objective and precise assessment of upper limb dysfunction post-stroke is critical for guiding rehabilitation. While promising, current methods using wearable sensors and machine learning (ML) often lack interpretability and neglect underlying, phase-specific kinetic deficits (e. g. , muscle forces and joint torques) within functional tasks. This study aimed to develop and validate an explainable assessment framework that leverages musculoskeletal kinetic modeling to extract phase-specific, multimodal (kinematic and kinetic) biomarkers to assess upper limb dysfunction in chronic stroke. Methods Sixty-five adults with chronic stroke and 20 healthy controls performed a standardized hand-to-mouth (HTM) task. Stroke participants were allocated to a model-development cohort (n = 47) and an independent test cohort (n = 18). Using IMU and sEMG data, we employed musculoskeletal modeling to extract phase-specific kinematic (e. g. , inter-joint coordination, trunk displacement) and kinetic (e. g. , mechanical work, smoothness, co-contraction index) biomarkers from four task phases. A Lasso regression model was trained to predict FMA-UL scores, validated via 5-fold cross-validation and the independent test cohort. Explainable AI (SHAP) was used to identify key predictive features. Results Compared with controls, patients showed phase-specific alterations including greater trunk displacement and reduced inter-joint coordination and mechanical work (all p 0. 05). The Lasso model achieved strong performance in internal validation (R 2 = 0. 932; MAE = 0. 799) and generalized well to the independent test cohort (R 2 = 0. 881; MAE = 0. 954). SHAP identified trunk displacement in phase 2 (TD₂), elbow–shoulder coordination in phase 3 (ICₑlbₑlv₃), and trunk displacement in phase 3 (TD₃) as dominant predictors; larger trunk displacement contributed negatively to predicted FMA-UL scores. Conclusion Integrating phase-specific multimodal biomarkers with explainable ML yields an interpretable upper-limb dysfunction. By highlighting phase-specific kinetic and kinematic targets (e. g. , trunk compensation and inter-joint coordination), the framework supports individualized, precision rehabilitation.