AI Turns an Ordinary Tablet Camera Into a Precise Balance Test Scorer

A simple balance test that has been a staple of sports medicine for decades may be on the verge of its biggest upgrade yet. Researchers in Shanghai have shown that an artificial intelligence system running on an ordinary tablet computer can score the Balance Error Scoring System, or BESS, more consistently than a trained human examiner watching the same test in real time. The findings, published in Sports Medicine – Open, suggest that the familiar ritual of a clinician squinting at an athlete’s wobbles could soon be replaced, or at least reinforced, by computer vision that never blinks, gets distracted, or misjudges a viewing angle.

The BESS is deceptively simple. Participants hold six different stances, each for twenty seconds: double-leg, single-leg, and tandem positions on both a firm floor and a foam pad, barefoot, hands on the hips, eyes closed. Every stumble counts. Opening the eyes, lifting a foot, stepping, stumbling, excessive hip movement, or removing the hands from the iliac crests each add an error, with a maximum of ten per condition and sixty overall. Higher scores indicate poorer postural control. The test is cheap, fast, and portable, which explains its popularity in concussion assessment and rehabilitation, but its Achilles heel has always been the human rater. Scores depend on experience, attention, viewing angle, and how one interprets fleeting or overlapping errors.

The research team, led by investigators at Huashan Hospital of Fudan University, set out to determine whether a camera-based AI system could remove that subjectivity. They recruited 84 people and analyzed 80: forty healthy controls, twenty individuals with chronic ankle instability, and twenty who had undergone anterior cruciate ligament reconstruction six to twelve months earlier. Every participant completed the full six-condition BESS protocol while a tablet mounted on a fixed stand, roughly 1.5 meters away, recorded a full-body frontal view. The same testing session then generated three independent scores: real-time scoring by an experienced sports physiotherapist, AI-assisted scoring from the tablet video, and a consensus reference score produced by two blinded physiotherapists who could pause, rewind, and replay the recordings at will.

The technical approach is notable for what it does not require. Rather than training a bespoke neural network on study data, the system uses the full-complexity variant of MediaPipe BlazePose to estimate 33 body landmarks in real time from the tablet’s video frames, supplemented by MediaPipe Face Mesh to track facial landmarks for eye-state estimation. Both are pretrained models, used without any fine-tuning on the study cohort. A fixed, deterministic, rule-based scoring layer then translates landmark trajectories into conventional BESS error counts. Eye opening is detected through an eye-aspect-ratio signal with participant-specific calibration and temporal filtering so that brief blinks are not penalized. Hand removal is judged from wrist, finger, and hip landmark geometry; excessive hip motion from the hip-knee segment angle relative to vertical; stepping and stumbling from foot, ankle, and trunk displacement; and forefoot or heel lifting from small-amplitude landmark shifts over short time windows. All processing runs locally in the tablet’s browser, and the scoring rules were locked before any data were collected.

The results were striking. Against the consensus video-review reference, the AI system achieved a Pearson correlation of 0.989 and an intraclass correlation coefficient of 0.989 for the total BESS score, with a mean difference of just 0.188 errors and 95 percent limits of agreement spanning roughly minus two to plus two and a half errors. The mean absolute error was 0.787 errors. The on-site human examiner, by contrast, correlated at 0.922 with the reference, with a mean absolute error of 2.212 errors and limits of agreement stretching from minus 6.680 to plus 5.455 errors. In plain terms, the machine disagreed with the meticulous video-review standard by less than one error on average, while the trained clinician’s real-time score could plausibly drift by several errors in either direction.

Those limits of agreement take on extra meaning when set against the published reliability literature for the test itself. A frequently cited intertester minimal detectable change for the total BESS score is 9.4 points, meaning that human raters’ scores can vary by that much before the difference is distinguishable from measurement noise. The AI system’s entire 95 percent agreement band with the reference standard, spanning about 4.5 errors, fits comfortably inside that human-rater error threshold. The authors are careful to note that this comparison is indirect, since the earlier figure came from a human reliability study rather than an AI-versus-reference design, and no validated minimal clinically important difference exists for the BESS. Still, the machine’s disagreement was small by any conventional yardstick of the test’s own measurement error.

Perhaps the most counterintuitive finding emerged when the researchers broke the results down by stance condition. One might expect the AI to struggle most in the hardest conditions, where movements are rapid and errors overlap. The opposite pattern appeared. The AI remained close to the consensus reference across all six conditions, with condition-specific mean absolute errors ranging from 0.013 to 0.287 errors, while the largest discrepancies clustered in comparisons involving the on-site human rater, particularly in firm-surface single-leg stance, foam-surface single-leg stance, and foam-surface tandem stance. These conditions invite brief corrective movements and rapid transitions in and out of position, precisely the events that are hardest to catch during a single real-time observation. The authors suggest that some of the disagreement traditionally attributed to the test’s difficulty may actually reflect the inherent limits of live human observation rather than any failure of automated scoring.

The study also provided a sanity check on whether the AI captures clinically meaningful patterns rather than random noise. Individuals with chronic ankle instability scored significantly higher than healthy controls across all three scoring methods, with large standardized differences, and the gap was concentrated in single-leg and tandem conditions, exactly what the existing CAI literature predicts. Participants after ACL reconstruction, tested six to twelve months after surgery and cleared for return-to-sport functional assessment, did not differ significantly from controls on total scores, a result the authors attribute to their relatively recovered status and the limited sensitivity of static balance testing for subtle residual deficits. The clinical groups were small and the between-group analyses were explicitly exploratory, so these findings support the system’s plausibility without establishing diagnostic performance.

The researchers are candid about the caveats. This was a single-center study under highly controlled recording conditions: a fixed front-facing camera, consistent lighting, a plain white wall, and clothing chosen to minimize occlusion of key landmarks. All four participants who could not complete the testing positions came from the clinical groups, potentially underrepresenting the most impaired patients. Only one session was analyzed, so the study says nothing about test-retest reliability or whether the AI can distinguish true within-patient change from ordinary variability over time. The reference standard itself, though rigorous, remains human judgment. And because the on-site comparison involved a single highly experienced rater, the results do not prove that AI would outperform every clinician, particularly novices or experts alike across varied settings.

What the study does establish is a proof of concept with unusually strong numbers: an ordinary tablet camera, running pretrained pose-estimation models and transparent rule-based logic entirely on-device, can preserve conventional BESS error counts while tracking a consensus reference more closely than real-time human scoring does. Before the system reaches clinics and sidelines, the authors call for external validation across independent cohorts, testing under messier real-world conditions with different devices, distances, lighting, and body types, and dedicated longitudinal reliability work. If those steps succeed, the humble balance test could become one of the clearest examples of AI quietly standardizing a clinical measurement that has depended on the fallible human eye for more than two decades.

Subject of Research: Validation of camera-based AI-assisted scoring for the Balance Error Scoring System in sports medicine populations

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

Article References: Camera-Based AI-Assisted Scoring for the Balance Error Scoring System: A Cross-sectional Validity and Agreement Study. (n.d.). https://doi.org/10.1186/s40798-026-01122-w

Image Credits: AI Generated

DOI: 10.1186/s40798-026-01122-w

Keywords: Balance Error Scoring System, artificial intelligence, computer vision, pose estimation, postural control, sports medicine, chronic ankle instability, ACL reconstruction, validity and agreement, mobile health, rehabilitation, concussion assessment