Automatic classification of forefoot (FFS), midfoot (MFS), or rearfoot (RFS) strike patterns from video data remains challenging. This study aimed to explore the ability of machine learning classifiers to determine foot strike patterns of runners and compare their performance to cut-point methods described in the literature. Video data from 767 U.S. Army trainees were used for this study. Participants were recorded while running on a treadmill at 2.68-2.90 m/s. The location of the heel and 5th metatarsal head were identified from video recordings using DeepLabCut and used to compute foot strike angles, angular excursion, and angular velocity. Three machine learning algorithms were trained to classify FFS, MFS, and RFS using 60% of the data. The remaining 40% of the data was held out to test model performance. Foot strike patterns were determined from foot strike angles using three different cut-point methods. The highest performing machine learning classifier had an F1 score of 70.0% compared to 66.7%, 68.9%, and 53.0% for the investigated cut-point methods. The machine learning classifier performed better at classifying MFS compared to the cut-point methods on most evaluation metrics. Machine learning algorithms may be an accurate and practical tool for automatic classification of foot strike.
Dæhlin et al. (Mon,) studied this question.