Neuromuscular Adaptation and Low-Latency Sensorimotor Integration in Stroke Rehabilitation: A Broad Learning System- Driven Exoskeleton Study

Abstract

Stroke and spinal cord injury (SCI) are two leading causes of permanent upper-limb impairment worldwide, and a large majority of stroke survivors retain hand motor deficits that restrict activities of daily living. Robotic exoskeletons can deliver the high-repetition, intent-driven movement practice associated with activity-dependent neuroplasticity, but their therapeutic value depends on how promptly sensory feedback follows the patient’s own motor intent. Closed-loop neurorehabilitation studies indicate that afferent feedback should reach the cortex within a few hundred milliseconds of volitional intent for associative, plasticity-inducing pairing to occur; the classification stage of an sEMG-driven controller is the portion of that budget most directly under the designer’s control, and convolutional neural network (CNN) classifiers typically consume 20–50 ms of it on embedded hardware. This paper presents a low-cost hand rehabilitation exoskeleton intended as a neuroplasticity-oriented intervention platform for stroke and SCI rehabilitation. A Broad Learning System (BLS) classifier decodes three hand gestures—power grasp, lateral pinch, and palm-up support—from surface electromyography (sEMG) with a mean inference latency of 7.8 ms on an STM32F405RGT6 microcontroller, approximately three times faster than a compact CNN baseline (23.5 ms) evaluated on the same feature set, at an overall classification accuracy of 94.7%. The hardware platform combines the STM32F405RGT6 controller, carbon-fibre-reinforced PLA (CF-PLA) structural links produced by fused-deposition modelling, and a lightweight wireless interface, for a worn assembly mass of 148 g excluding the battery and controller board; we refer to the complete system as BLS-ExoHand. Validation combined hardware-in-the-loop multibody simulation with bench testing of the assembled third-generation prototype and sEMG recordings from ten healthy participants; the two validation paths were used to cross-check one another, and the provenance of every reported figure is stated explicitly in Section 5.1. The assisted condition reached a task success rate of 88.5% and a fingertip repeatability error of 1.8 ± 0.4 mm. At a bill-of-materials cost below $85 USD, the system offers an accessible alternative to commercial rehabilitation exoskeletons priced between $6,000 and $100,000.