Classification of functional movement screening scores and reported fall-incident history from deep squat videos among logistics service workers using deep learning

Logistics service workers (LSWs) face a high risk of fall-related injuries owing to the physically demanding nature of their work. This study aimed to develop a deep-learning model to classify Functional Movement Screening (FMS) deep-squat scores and reported fall-incident history among LSWs by analyzing sagittal-plane squat movement patterns captured through video recordings. A total of 403 LSWs participated in this study, and 378 videos were used for model development after excluding missing or inadequate-quality recordings. Among participants, the mean age was 38.4 ± 7.8 years and the mean BMI was 24.5 ± 3.7 kg/m². An efficient convolutional network for online video understanding (ECO) was used to process the video data while considering both spatial and temporal characteristics. Among the predefined sampling settings, the 16-frame ECO model showed the highest mean area under the receiver operating characteristic curve (AUROC) for classifying FMS/D3 scores (AUROC: 0.808; 95% confidence interval [CI]: 0.778-0.837) and reported fall-incident history (AUROC: 0.730; 95% CI: 0.688-0.773), although the performance differences among frame settings were modest. Class-agnostic activation maps revealed that the model frequently focused on the trunk and pelvic regions during deep squat movement for both FMS groups, whereas distinct activation patterns were observed in the fall-history classification model. This study demonstrates the potential of ECO-based spatiotemporal video analysis as a proof-of-concept approach for automated functional movement assessment in occupational health settings.