Camera-Based AI-Assisted Scoring for the Balance Error Scoring System: A Cross-sectional Validity and Agreement Study

BACKGROUND: The Balance Error Scoring System (BESS) is widely used to assess static postural control in sports medicine and rehabilitation. However, conventional real-time scoring is observer dependent and may be influenced by rater experience, viewing angle, attention, and the interpretation of transient or overlapping errors. This study evaluated the criterion-related validity and agreement of a camera-based artificial intelligence (AI)-assisted BESS relative to consensus video-review manual scoring and compared its error magnitude with on-site manual scoring. METHODS: This single-center cross-sectional study included 80 of 84 recruited participants: 40 healthy controls, 20 with chronic ankle instability (CAI), and 20 after anterior cruciate ligament reconstruction (ACLR). All participants completed the standard six-condition BESS protocol. Scores were obtained using AI-assisted, on-site manual, and consensus video-review manual scoring; the latter was prespecified as the pragmatic manual reference standard. Primary total-score analyses included Pearson correlation, the intraclass correlation coefficient (ICC) with 95% confidence intervals (CIs), Bland-Altman mean difference and 95% limits of agreement (LoA), mean absolute error (MAE), and root mean square error (RMSE). Condition- and group-specific analyses were exploratory. RESULTS: For the total BESS score, AI-assisted scoring versus the reference showed r = 0.989, ICC = 0.989 (95% CI 0.982-0.993), a mean difference of 0.188 errors, 95% LoA from -2.063 to 2.438 errors, an MAE of 0.787 errors, and an RMSE of 1.157 errors. On-site manual scoring versus the reference showed r = 0.922, ICC = 0.920 (95% CI 0.877-0.948), a mean difference of -0.612 errors, 95% LoA from -6.680 to 5.455 errors, an MAE of 2.212 errors, and an RMSE of 3.136 errors. Condition-specific discrepancies were larger in the single-leg and tandem-stance conditions. Exploratory group comparisons showed higher total BESS scores in individuals with CAI than in healthy controls. CONCLUSIONS: Camera-based AI-assisted BESS scoring showed strong total-score agreement with the pragmatic manual reference standard and smaller discrepancies than on-site manual scoring. This approach may support more standardized BESS assessment; however, external validation and longitudinal reliability testing are required before interchangeability with manual scoring can be considered.