Abstract / Summary
Abstract Purpose Previous knee injury increases the risk of subsequent injury, even after functional recovery, yet residual gait, running or load alterations often go undetected in routine assessment. Wearable inertial measurement units (IMUs) may enable scalable detection of such alterations outside laboratory settings. The primary aim of this study was to determine whether IMU‐derived gait and running features can distinguish physically active individuals with a history of knee injury from uninjured controls using machine learning. Methods In this cross‐sectional study, 108 sports science students completed injury‐history questionnaires and a walking‐running‐walking protocol on a 400 m track while wearing four IMUs. Of 90 participants with complete data, 23 reported prior knee injury and 67 were uninjured controls. Left‐ and right‐side sensor data were combined into one per‐participant feature vector. The model classified any prior knee injury (yes/no), not limb‐specific injury. From 220 candidate features covering baseline gait, post‐run changes, walk‐to‐run transitions, inter‐limb asymmetry, variability and composite indices—a fixed pipeline (e.g., median imputation, mutual‐information selection) was applied within leave‐one‐out cross‐validation (primary) and repeated stratified five‐fold cross‐validation (secondary). Results Leave‐one‐out cross‐validation correctly classified 76/90 participants (65/67 true negatives, 11/23 true positives), yielding an area under the receiver operating characteristic curve of 0.74, balanced accuracy of 0.72, sensitivity of 0.48, specificity of 0.97, positive predictive value (PPV) of 0.85 and negative predictive value of 0.84. Five‐fold cross‐validation produced more variable estimates, indicating substantial model instability across resampling partitions. Conclusions As a proof of concept, IMU‐derived features distinguished individuals with previous knee injury at high specificity but limited sensitivity. High specificity and PPV depend on the classification threshold and injury prevalence in this selected sample, not intrinsic model properties. Because predictors were not aligned to the injured side, findings cannot be interpreted biomechanically. Potential applications (e.g., targeted monitoring) require prospective external validation. Level of Evidence Level II.