Identification of Motion Factors for Knee Joint Contact Force during Walking Using Convolutional Neural Network

Abstract

The analysis of knee motion during walking is essential for understanding the mechanisms of knee joint contact force (KJCF), a factor associated with knee joint pain and related joint diseases. A comprehensive analysis of whole-body motion can provide valuable insights and practical strategies for mitigating KJCF. This study aimed to identify body segments whose motion is related to KJCF using a Convolutional Neural Network (CNN). We used a gait database to obtain three-dimensional motion data and calculated their KJCF using a musculoskeletal model. In addition, the peak values of KJCF during gait were classified into five classes to develop a learning model with CNN. Visualization using Gradient-weighted Class Activation Mapping revealed that the regions of interest identified by the trained CNN model were the bases of the middle and ring fingers of the right hand, the outside of the right thigh, the left side of the hip, and the thumb side of the left wrist. These findings suggest that the motions of the hand markers, which change with the arm swings, impact the variation of KJCF.