Aim
Knee strength is a key modifiable rehabilitation target after anterior cruciate ligament reconstruction (ACLR), yet evidence reports only weak associations between strength and patient-reported outcomes, limited by cross-sectional, linear analyses that ignore the hierarchical structure of clinical data1. We aimed to identify which isokinetic knee strength metrics best predict psychological return-to-play readiness (ACL-RSI) after ACLR using a longitudinal, mixed-effects machine learning framework, and to test whether predictive performance differs between new and follow-up patients.
Methods
180 patients (82 females, 98 males; age 23.8 years) were tested between 9 and 12 months following primary unilateral ACLR with ipsilateral autograft (hamstring 70%, quadriceps 28%, patellar tendon 2%; lateral extra-articular tenodesis in 57%), yielding 1,097 isokinetic trials, with 61% of patients tested at both timepoints. ACL-RSI was completed at each session. Over 150 biomechanical features were derived, including peak torque, angle-specific torque (10 to 30 degrees of flexion), angle of peak torque, rate of force development, wavelet-based neuromuscular control measures (steadiness and tremor ratio), limb symmetry indices and extension-to-flexion ratios. Boruta and recursive feature elimination reduced this to the top 25 predictors, combined with three clinical characteristics (sex, graft type, lateral extra-articular tenodesis status). A generalised mixed-effects extreme gradient boosting model (GMEXGBoost)2, augmented with a linear mixed model on patient-specific random intercepts, was compared with standard XGBoost on an 80/20 split with 10-fold cross-validation, tested (A) on new patients and (B) on follow-up patients from an earlier visit. SHAP (SHapley Additive exPlanations) analyses ranked predictor contributions.
Results
For new patients (single visit), both models performed similarly poorly (GMEXGBoost: R² 0.10, RMSE ±20.36; XGBoost: R² 0.12, RMSE ±20.01; 0–100 ACL-RSI scale, cohort mean 67.1). For follow-up patients, GMEXGBoost substantially outperformed XGBoost (GMEXGBoost: R² 0.73, RMSE ±9.33; XGBoost: R² 0.02, RMSE ±20.43), a 2-fold reduction in error. SHAP revealed three patterns. First, graft type emerged as the strongest population-level predictor once individual baselines were isolated. Second, among biomechanical features, angle-specific torque (10 to 30 degrees of flexion) and non-operated limb steadiness and tremor ratios ranked above peak torque and limb symmetry. Third, healthy peak extension torque occurs at 60 to 70 degrees of knee flexion, and patients whose peak migrated toward full extension (30 to 50 degrees) reported lower ACL-RSI.
Conclusion
A single-visit snapshot of isokinetic strength did not predict psychological return-to-play readiness. Given an earlier visit from the same patient, the mixed-effects model predicted that patient's next ACL-RSI within roughly nine points, being twice as accurate as standard machine learning. At a population level, graft type was the strongest predictor of ACL-RSI once individual variation was accounted for. Among biomechanical features, the strongest predictors were neuromuscular quality and the angle of peak torque, not peak strength or limb symmetry. Time-course, within-patient strength monitoring is more informative than one-off testing, and reliance on peak torque and limb symmetry as return-to-sport clearance assessments may miss the neuromuscular qualities most associated with psychological readiness. SHAP rankings reflect what the model finds useful rather than causal effects, and these findings warrant confirmation in stratified analyses.
