A multimodal machine learning approach to predict Fugl-Meyer scores and motor recovery potential in stroke rehabilitation: Toward precision-based therapies

Stroke is a leading cause of long-term disability, with highly variable recovery trajectories and challenges in prediction and monitoring. Frequently used measures (e.g., National Institute of Health Stroke Scale (NIHSS) and Fugl-Meyer (FM) assessment of motor impairment) have significant limitations. As the societal burden of stroke increases, developing robust methodologies for assessing and predicting recovery is essential to optimize treatment plans and improve outcomes. This paper presents our Integrated Motion Analysis Suite (IMAS), which leverages multi-modal data (clinical, sensor, and neuroimaging inputs) and multimodal-machine-learning (MML) to predict FM scores and motor recovery in stroke. Its potential is demonstrated via analysis of 28 S patients in acute and subacute phases of recovery, where features extracted from a set of motor tasks were used to predict FM scores and motor recovery, achieving a coefficient of determination (R 2 ) of 0.75 and Mean Absolute Error (MAE) of 2.83 and R 2 = 0.83 and MAE = 2.6 %, respectively. IMAS is designed to continuously improve through its integration with a Big Data database, allowing for ongoing refinement of predictive algorithms as new data is collected in real-world clinical environments. Its ability to complement inherently limited clinical scales, handle incomplete data, and adapt to diverse applications highlights its potential for broader use in recovery after stroke, including long-term monitoring and precision rehabilitation.