AI-Based Personalized Rehabilitation Planning for Post-Stroke Patients

Abstract

The retraining process following stroke is crucial in restoring motor functions and stimulating neuroplasticity development of the cortex. Yet conventional physical therapy methods suffer from such drawbacks as high labor cost, qualitative subjectivity and rigidity of their rehabilitation schemes that fail to adjust to the rapid changes in recovery pace of their patients. This paper will discuss a new safety-critical multi-modal artificial intelligence system to design and continuously refine patient-specific motor rehabilitation program. The proposed system accepts inputs from sEMG, inertial motion tracking and EEG sensors to form an overall 'state vector' of the patient that includes muscle activation intensity, joint coordination error and the level of cognitive involvement. The asymmetrical paretic joint coordinate interactions will be modeled by a Graph Convolutional Network (GCN) to establish non-linear dependencies between the corresponding joints. Finally, this state vector will be employed to build a Soft Actor-Critic (SAC) reinforcement learning agent to optimize torque assistance, task complexity and duration of the exercise when utilizing a robotic exoskeleton. Safety of the patient during the learning process will be secured by the implementation of a Control Barrier Function (CBF) limiting torque/force to physiological levels. The clinical efficacy of the proposed model was validated on a cohort of twenty simulated post-stroke subjects representing mild, moderate, and severe motor impairments. The experimental findings demonstrate that the proposed framework accelerates motor recovery by 34.2% on the Fugl-Meyer Assessment scale compared to conventional physical therapy, while eliminating safety violations. Semicolons are used to demarcate our primary metric categories; the comparative systems are evaluated using joint tracking deviation, muscular fatigue, cognitive engagement index, and safety violation counts.