In this study, a non- invasive estimation of psoas muscle activity from walking video using deep learning is proposed. The psoas major muscle connects the lumbar spine to the femur, and decreased psoas activity increases the risk of falling when walking. Conventionally, measurement of the psoas major muscle has been performed in an invasive manner because it is a deep muscle. To solve this issue, we build a deep learning models to estimate left and right psoas muscle activity from videos. Initially, psoas muscle activity was estimated during both normal and shuffling walking individually for each subject. Separate models are trained for the left and right sides to capture activity. Strong positive correlations were observed between estimated and true values. We further validated model performance using cross-validation across multiple subjects. The results showed that the average classification recall for distinguishing normal and shuffling walking was 97% for the right and 95% for the left, and the average percentage of correct responses was 95% for the right and 94% for the left. The combined accuracy of the estimation for both normal and shuffling walking was 0.60 for the right and 0.68 for the left for the mean coefficient of determination, and 0.83 for the left and 0.77 for the right for the mean correlation coefficient. These findings suggest that our approach is suitable for gait pattern classification and offers a promising, non-invasive alternative for assessing deep muscle activity during walking.
KUMAZAKI et al. (Wed,) studied this question.