A smaller gap than on-site scoring
A camera-based artificial intelligence system produced total Balance Error Scoring System (BESS) scores that were very close to consensus video-review manual scores in a study of 80 participants. The AI's mean absolute error, the average size of its gap from the reference, was 0.787 errors. The corresponding figure for real-time on-site manual scoring was 2.212 errors. Root mean square error, a measure that gives more weight to larger misses, was 1.157 for AI and 3.136 for on-site scoring.
BESS is a six-condition balance test. Each condition receives 0 to 10 errors, producing a total from 0 to 60, and higher totals indicate poorer postural control. In this study, the performance was scored by the camera-based system, by an on-site scorer in real time, and by manual reviewers who watched tablet video and formed a consensus. The comparison asked whether the automated method could track that manual benchmark more closely than on-site scoring.
Total scores tracked closely
On total scores, the AI and consensus methods had a Pearson correlation of 0.989. Correlation measures whether two sets of scores move together; here, participants with higher or lower manual totals generally had higher or lower AI totals. The intraclass correlation coefficient, reported as ICC(A,1), was also 0.989, with a 95% confidence interval from 0.982 to 0.993. The reported correlation had p < 0.001.
Agreement was also close when the researchers looked at the difference between scores. On average, AI totals were 0.188 errors higher than the manual reference, with a 95% confidence interval from -0.068 to 0.443. The 95% limits of agreement, which show how widely individual differences spread, ran from -2.063 to 2.438 errors for AI. For on-site scoring, they ran from -6.680 to 5.455. The narrower range in the AI comparison is consistent with its smaller average error in this sample.
How the benchmark was built
Consensus video-review manual scoring was the prespecified pragmatic reference standard. Before that consensus was reached, the two independent video reviewers had an ICC(A,1) of 0.956. They gave identical ratings for 346 of 480 condition-level ratings, differed by no more than two errors on 462 ratings, and needed a third-reviewer adjudication for only two.
The final analysis included 40 healthy controls, 20 people with chronic ankle instability, and 20 people assessed after anterior cruciate ligament reconstruction. Four of the 84 recruited participants were excluded because they could not complete the required BESS positions. The investigators had targeted at least 73 participants, based on a calculation that required 65 analyzable participants and allowed for approximately 10% exclusions; the final analysis included 80.
The AI processed tablet video in real time with pretrained BlazePose and Face Mesh landmark models. It estimated 33 body landmarks and converted their movement paths into scores through a fixed, deterministic rule-based layer. The algorithm was fixed before statistical analysis and was not changed after the consensus results were examined.
A result with a narrow reach
The design places a limit on what can be concluded. This was a single-center, cross-sectional validity and agreement study using convenience sampling, and its manual comparator was a pragmatic reference standard. The findings therefore describe agreement with that benchmark in this sample, not evidence that the automated score is interchangeable with manual scoring in every setting.
The authors present the system as a potential adjunct for standardizing conventional BESS assessment. They say external validation, longitudinal reliability testing, and testing under less controlled conditions are needed before interchangeability with manual scoring or individual-level clinical use can be assumed.
Paper data and sources
Original title: Camera-Based AI-Assisted Scoring for the Balance Error Scoring System: A Cross-sectional Validity and Agreement Study
Authors: Chen Y, Cheng J, Wang T et al.
Journal/Repository: Sports Medicine - Open
Status: Peer-reviewed
First online: 2026-10-12
DOI: Not available
Original paper · Full text