We recruited 20 participants with chronic ankle instability and trained a temporal convolutional network to classify gait cycles labeled GOOD or BAD based on ankle angle thresholds. Model performance was assessed with participant‑held‑out leave‑one‑subject‑out (LOSO) cross‑validation; across the 20 folds the mean AUROC was 0.948, sensitivity for BAD cycles 0.941, and specificity for GOOD cycles 0.366.
In the adaptive‑intervention arm, seven participants completed nine sessions over three weeks, each session providing a single motion‑capture recording. After a failed classification, the model was updated offline and evaluated on the same‑session validation subset and on the first subsequent adaptive session. Candidate models improved BAD‑class F1 by 0.187 on the same‑session subset and by 0.118 on the first subsequent recording.
Frontal‑plane ankle angle was compared between the adaptive group and ten sequentially enrolled controls. After baseline adjustment, the adaptive group showed a 5.168° lower angle at Post (95% CI: 1.766–8.569°); the contrast at the 7‑day Retention point remained uncertain.
These findings characterize the discrimination ability of a population‑level model and offline participant‑specific updating during repeated biofeedback use, but they do not establish independent clinical gait classification nor a causal benefit of model updating.
Review