Sensory Data Classification for Gait Assessment using Deep Neural Networks: A Comparative Study with SGD and Adam Optimizer

The evaluation of human walking through sensor data is pivotal in diagnosing various neurological disorders. Despite the widespread use of deep learning methods to distinguish between healthy and pathological walking, the efficacy of already in-place approaches remains unclear. In this study, we have developed and implemented a Deep Neural Network (DNN) to identify healthy and pathological gaits based on a muscle sensory and knee movement dataset. The considered dataset consists of walking information from 22 subjects, with 11 classified as healthy and rest with knee pathology, sourced from the machine learning repository at the University of California, Irvine (UCI). This dataset encompasses five channels of sensor signals, with the four containing surface Electromyography (sEMG) data representing muscular initiations in lower limbs as Biceps Femoris (BF), Vastus Medialis (VM), Rectus Femoris (RF), and Semitendinosus (ST). The fifth channel comprises goniometer signals measuring the flexion/extension movements of the knee joint. To assess the classification process, two optimizers, namely Stochastic Gradient Descent (SGD) and Adaptive Moment (Adam), are considered. Subsequently, performance metrics such as precision, recall, F1-score, and confusion matrix are evaluated and discussed for both optimizers. The training accuracy in classifying healthy and pathological gaits for the two optimization algorithms is determined to be 79.83% and 90.81%, respectively. The Adam-DNN demonstrates superior performance compared to SGD-DNN in classifying healthy and pathological gaits.