Smart Prosthetics with Sensor Integration

Abstract

Upper-limb prosthetics have evolved from passive cosmetic devices to motorised myoelectric hands, yet current commercial systems remain limited by coarse grip control (typically 2-6 predefined patterns), absence of sensory feedback, and rejection rates exceeding 35% among transradial amputees. Smart prosthetics integrating distributed sensor arrays with machine learning-driven intent decoding promise to restore dexterous manipulation and tactile perception. This study developed and evaluated a sensorised prosthetic hand platform incorporating five sensing modalities -- surface electromyography (sEMG, 8-channel armband), barometric tactile sensors (24 taxels across fingertips and palm), inertial measurement units (IMUs, 6-axis per digit), flex sensors (resistive, per proximal interphalangeal joint), and force-sensitive resistors (FSRs, per fingertip) -- interfaced with four classification architectures: linear discriminant analysis (LDA), support vector machine (SVM), convolutional neural network (CNN), and a proposed recurrent attention network (RANet) for real-time grasp intent decoding. Evaluation was conducted with 15 transradial amputees performing 12 Activities of Daily Living (ADL) tasks from the Southampton Hand Assessment Procedure (SHAP). RANet achieved the highest grasp classification accuracy (94.6 +- 2.8% across 8 grasp types, 150 ms decision latency), significantly outperforming LDA (76.4 +- 5.2%; p < 0.001) and CNN (88.2 +- 3.4%; p = 0.004). Integration of tactile feedback via electrotactile stimulation on the residual limb improved SHAP functional scores by 18.4% (p = 0.002) and reduced object drop rate from 24.6% to 8.2% (p < 0.001). These results establish the multi-modal sensor fusion approach as a viable pathway toward intuitive, sensorised prosthetic hands with near-natural grasp control