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

Area of Science:

• Biomechanics
• Neurorehabilitation
• Machine Learning

Background:

• Accurate assessment of upper limb dysfunction post-stroke is crucial for effective rehabilitation.
• Current wearable sensor and machine learning (ML) methods often lack interpretability and fail to capture phase-specific kinetic deficits.

Purpose of the Study:

• To develop and validate an explainable assessment framework using musculoskeletal kinetic modeling.
• To extract phase-specific, multimodal (kinematic and kinetic) biomarkers for assessing upper limb dysfunction in chronic stroke patients.

Main Methods:

• Sixty-five chronic stroke adults and 20 healthy controls performed a hand-to-mouth task.
• Inertial Measurement Units (IMU) and surface electromyography (sEMG) data were used with musculoskeletal modeling to extract biomarkers.
• A Lasso regression model predicted Fugl-Meyer Assessment-Upper Limb (FMA-UL) scores, validated using cross-validation and an independent cohort; SHAP (SHapley Additive exPlanations) identified key features.

Main Results:

• Stroke patients exhibited phase-specific alterations, including increased trunk displacement and reduced inter-joint coordination and mechanical work compared to controls.
• The Lasso model demonstrated strong predictive performance (R²=0.932 internally, R²=0.881 externally) for FMA-UL scores.
• SHAP analysis highlighted trunk displacement and elbow-shoulder coordination as dominant predictors, with greater trunk displacement negatively impacting FMA-UL scores.

Conclusions:

• An interpretable framework integrating phase-specific multimodal biomarkers with explainable ML effectively assesses upper limb dysfunction post-stroke.
• The framework identifies specific targets for rehabilitation, such as trunk compensation and inter-joint coordination, enabling individualized, precision rehabilitation strategies.