Abstract

Anterior cruciate ligament (ACL) injury can be missed during the initial evaluation because compensatory strategies may conceal subtle gait abnormalities. We evaluated the accuracy with which healthcare professionals recognize ACL deficient gait from brief visual observation and the classification performance of depth camera-based artificial intelligence (AI) for the same task. In this retrospective diagnostic study, 128 participants were selected from 688 gait registry records, including 65 patients with preoperative ACL rupture and 63 healthy controls. Ten-meter walks were recorded using a depth camera, and 2,700 whole-body spatial features were derived from the mean, minimum, and maximum distances among 25 skeletal landmarks during one gait cycle. After feature selection, 23 supervised learning algorithms generated 22,124 candidate models. In an independent test set of 30 participants, the final model achieved 80.0% sensitivity, 66.7% specificity, 73.3% accuracy, and an area under the receiver operating characteristic curve of 0.776. Pooled assessments by 60 healthcare professionals showed numerically smaller sensitivity, specificity, accuracy, and agreement. Most selected features captured intersegmental relationships beyond the knee, with trunk-related features accounting for 60%. These findings indicate that depth camera-based AI can quantify distributed whole-body gait adaptations and may support more standardized recognition of subtle ACL related gait abnormalities in clinical practice.

Credit: Homan et al., Scientific Reports (2026), https://doi.org/10.1038/s41598-026-69626-9. © The Author(s) 2026. Licensed under CC BY-NC-ND 4.0.