Machine Learning Analyses For Determination Of Self-reported Knee Function From Biomechanical Variables

Self-reported sub-optimal knee function may be associated with altered biomechanics and subsequent risk for knee injury. Recent evidence indicates that reductionistic approaches to biomechanical analyses are lacking and do not fully capture one’s global function and injury risk profile, but feasible comprehensive biomechanical analyses are unknown. PURPOSE: To elucidate multi-joint and multivariate biomechanical patterns associated with self-reported sub-optimal knee function and to compare the performance of manual logistic regression, automated logistic regression and a machine learning classification and regression tree (CART) analysis. METHODS: We recruited 90 physically active participants (45 females, 45 males) aged 18 and 25, who performed five single-leg forward landings onto an embedded force plate. Initial and peak triplanar trunk, hip and knee joint angles were used as independent variables. State variables of optimal and suboptimal knee function were a KOS-SAS score of 100% and a score of ≤99%, respectively. RESULTS: 50 and 40 participants reported optimal and sub-optimal knee function, respectively. Manual Logistic Regression. A 3-factor model included initial trunk flexion ≥30°, initial ipsilateral trunk lean ≥8°, and ≤ 2° of initial hip external rotation and an overall accuracy of .80. Automated Logistic Regression. All features were initially included, followed by recursive feature elimination and 10-fold cross validation, resulting in a 3-factor model including increased initial trunk flexion, increased peak trunk flexion, and increased initial ipsilateral trunk lean and an accuracy of .52. Decision Tree. CART analysis resulted in a three-factor model including initial external hip rotation ≥8.03°, peak knee adduction ≥6.66° and peak knee abduction ≥1.33° and an accuracy of .71. CONCLUSION: The combination of increased trunk flexion, increased ipsilateral trunk lean, frontal plane knee motion, and an externally rotated hip was associated with self-reported sub-optimal knee function. These patterns are consistent with risky trunk and knee movement. Manually determining biomechanical cutoff scores resulted in the most accurate model, which may be beneficial to clinicians seeking to extrapolate movement tendencies from self-reported knee function.