Perception-Edge-Cloud Architecture for Post-Stroke Rehabilitation: A Biosensor-Based Intelligent System for Movement Recognition and Compliance Assessment

Abstract

To address the problem of distorted NC (Nursing Compliance) assessment for stroke patients during home rehabilitation, this paper proposes a four-level collaborative intelligent assessment and guidance system architecture: "Perception-Edge-Cloud-Feedback." First, upper limb movement and electromyography signals are synchronously collected using IMU (Inertial Measurement Unit) and sEMG (Surface Electromyography) sensors. Signal preprocessing, combined with wavelet denoising and SW (Sliding Window) segmentation, improves interference resistance in unstructured home environments. Second, multidimensional features are extracted from the time-frequency-spatial domain, and the input space is optimized. Finally, an attention-enhanced LSTM (Long Short-Term Memory)-Gated Recurrent Unit (GRU) hybrid architecture is constructed to adaptively capture key spatiotemporal features. Finally, a mapping mechanism between action semantics and rehabilitation prescriptions was constructed, along with a dynamically weighted compliance scoring model. The methodology integrates Daubechies wavelet denoising, sliding window segmentation, and an attention-enhanced LSTM-GRU network to process synchronous IMU and sEMG signals. Experimental validation involving 32 patients demonstrates that this system achieves an average recognition accuracy of 93.6%, an Intraclass Correlation Coefficient (ICC) of 0.89 between automatic and manual compliance scoring, and a significant improvement in Fugl- Meyer Assessment scores (p < 0.001) for the high-adherence group.