A Convolution Neural Network Approach to Access Knee Joint Angle Using Foot Pressure Mapping Images: A Preliminary Investigation

The use of various footwear results in change in knee biomechanics, altering gait pattern. Monitoring the joint angle using machine or deep learning approaches is challenging due to input data type selection. This study introduces a convolution neural network (CNN) relied on foot pressure mapping to assess knee joint angle during use of various footwear. A pilot study was conducted with six healthy participants walking at a comfortable gait on a treadmill wearing three types of footwear. We used the F-scan system to measure foot pressure and saved the information to video as input data. Kinematic data of knee joint angle was calibrated by Kinovea software using video recording and served as reference data. We extracted 10 gait cycles from input and reference data. The CNN model was trained with 12000 images and validated with 3000 images. Results of our study suggest that the CNN image-based can accurately predict the knee joint angle with mean absolute error lower than 10.0 (deg), mean relative error less than 10.0%, and the correlation coefficient between 70.0% and 90.0% for the three types of tested footwear. The CNN image-based generated a similar knee joint angle pattern for a gait cycle (GC) of all tested shoes; less information about the image pattern during the swing phase is available. Integration between CNN and foot pressure mapping image provided a strong correlation in estimating knee joint angle for a GC during use of three footwear types.