Development of a Pressure Sensing System Coupled with Deployable Machine Learning Models for Assessing Residual Limb Fit in Lower Limb Prosthetics

Lower limb amputations pose significant challenges for patients, with over 150,000 cases annually in the U.S., leading to a high demand for effective prosthetics. However, only 43% of lower limb prosthetic users report satisfaction, primarily due to issues with socket fit, which is critical for comfort, stability, and preventing injury. This study presents a deployable sensing system for potentially real-time monitoring of prosthetic socket fit by using pressure sensors and convolutional neural networks (CNNs) to analyze the pressure distribution within the socket. A novel CNN architecture, utilizing both dilated and strided convolutions, is proposed to effectively capture spatial-temporal patterns in multivariate timeseries data, which is processed as an image. The system was designed for edge deployment on the Sony Spresense microcontroller, maintaining a small model size while achieving high accuracy. Results show that the CNN models, particularly those optimized with the stochastic gradient descent (SGD), demonstrated robustness and high transferability. This system provides a cost-effective, portable solution to improve prosthetic fit, enhancing patient care and preventing gait-related injuries.