Finding the needle in the haystack of isokinetic knee data: Random Forest modelling improves return-to-sport information

Abstract Background : The difficulties of rehabilitation after anterior cruciate ligament (ACL) injuries, subsequent return-to-sport (RTS) let alone achieving pre-injury performance are well known. Isokinetic testing is often used to assess strength capacities during that process. The aim of the present applied machine learning (ML) approach was to examine which isokinetic data differentiates athletes post ACL reconstruction and healthy controls. Data from unilateral concentric and eccentric knee flexor and extensor tests (30°/s, 150°/s) was used to train Random Forest models from 366 male (63 post ACL reconstruction) and 183 female (72 post ACL reconstruction) athletes. Via a cross validation predictive performance was evaluated and accumulated local effects plots analysed the features of the models. Results : Random Forest showed outstanding predictive performance for male (AUC=0.90, sensitivity=0.76, specificity=0.88) and female (AUC=0.92, sensitivity=0.85, specificity=0.89) athletes. For both male and female athletes, the ten most impactful features on the predictive likelihood of the model either referred to the disadvantageous (injured, non-dominant in control group) leg or to lateral differences. The eccentric hamstring work at 150°/s was identified as the most impactful single parameter. Conclusion : A ML model trained with parameters from isokinetic knee data discriminated between athletes 6 to 24 months post ACL reconstruction and healthy athletes with high accuracy. We see potential for improving RTS decision making by incorporating and combining measures, which focus on hamstring strength, leg symmetry and contractional work.