Gaitformer: a spatial-temporal attention-enhanced network without softmax for Parkinson’s disease early detection

Abstract

Gait analysis is an increasingly expanding research field, characterized by the application of non-invasive sensors and machine learning techniques across various domains. Using these advanced technologies, researchers can deep dive into understanding human gait and movement patterns, providing robust support for applications such as medical diagnosis, rehabilitation, and sports optimization. In this paper, we focus primarily on analyzing the gait features of a large population and emphasize the study of representative features of gait in terms of both temporal and spatial dimensions. By analyzing parameters such as pressure distribution, gait cycles, and gait features of the foot sole, we aim to evaluate an individual’s gait function and detect and diagnose gait-related diseases such as Parkinson’s disease. By integrating spatiotemporal feature information and employing XNorm modules and sparse attention mechanisms in the spatiotemporal encoder to enhance gait feature extraction and model generalization capabilities, our experimental results show that our model achieves a classification accuracy of 92.3%. This also indicates that our Gaitformer demonstrates considerable potential in medical diagnosis models.