Preoperative performance data mapped to 1-year ACL reconstruction recovery
A retrospective cohort study from Korea University tested whether machine learning could predict two clinically meaningful 1-year outcomes after primary anterior cruciate ligament reconstruction: return to pivoting sports and achievement of the patient acceptable symptom state. The appeal is straightforward. Surgeons and rehabilitation teams routinely gather demographic data, treatment details, isokinetic strength measures, and patient-reported outcomes before surgery, yet much of that information remains underused when it comes to forecasting who will regain high-level sport function and who will still feel limited a year later.
Seventy-five patients who underwent ACL reconstruction between 2016 and 2023 formed the retrospective dataset. Preoperative features included demographic characteristics, treatment-related variables, isokinetic muscle performance, and patient-reported outcome measures. One-year outcomes were defined as return to pivoting sports, using a Tegner activity level of at least 6, and achievement of the patient acceptable symptom state, defined here by an International Knee Documentation Committee score of at least 75.9. Variables identified in univariable logistic regression were then used to build several predictive models, which were trained and evaluated with 3-fold cross-validation.
The best-performing model was a random forest algorithm. It reached an area under the receiver operating characteristic curve of 0.832 for return to sport and 0.815 for patient acceptable symptom state. In practical terms, those values suggest the model separated patients who met the target outcome from those who did not with respectable discrimination for an exploratory dataset of this size. Calibration plots and Brier scores were also used, which matters because discrimination alone can make a model look stronger than it truly is if predicted probabilities are poorly aligned with real-world outcomes.
One finding stood out with clinical consistency: younger age and greater preoperative flexor strength of the uninvolved limb emerged as the most important predictors in SHapley Additive exPlanations analysis. Age is an intuitive signal in ACL recovery because younger patients are often more likely to pursue pivoting sports and to tolerate the demands of a higher-risk athletic return. The role of contralateral flexor strength is more interesting at the bedside. It may reflect a broader phenotype of conditioning, neuromuscular reserve, and lower-limb symmetry before surgery, all of which can influence postoperative rehabilitation capacity and eventual sport readiness. Even though the uninvolved limb is not the injured knee, it can function as a marker of the patient’s overall physical platform heading into reconstruction.
The results also reinforce a familiar orthopedic reality: one-year recovery after ACL reconstruction is not determined by the operation alone. Preoperative function, athletic goals, and baseline strength appear to carry meaningful prognostic weight, and a model that integrates those elements may help identify patients who need closer counseling or a more intensive rehabilitation trajectory. The study does not suggest replacing clinical judgment with an algorithm. It points toward a more structured use of information already available before surgery.
A narrow dataset can still generate useful signals, but it cannot settle the question of how well those signals travel to other hospitals, other surgeons, and other rehabilitation programs. With only 75 patients, the model is vulnerable to overfitting even with cross-validation, especially when the number of candidate predictors is limited and the outcome definitions are tightly tied to sport participation and symptom reporting. The absence of external validation is the most important restraint on interpretation. Without testing in an independent cohort, the performance estimates remain local rather than generalizable.
The paper’s framing is appropriately cautious. These models are exploratory and hypothesis-generating, not ready for clinical deployment. That distinction matters because predictive tools in sports medicine can easily drift from helpful to misleading if they are introduced before independent validation, calibration testing across settings, and demonstration that they improve patient counseling or rehabilitation decisions. A model that looks accurate in a retrospective single-center sample may fail when used in a different population with different surgical techniques, therapy access, or expectations for sport return.
Even so, the study reflects a direction that feels clinically credible. Preoperative physical performance is not just a snapshot of the injured knee; it is a measure of reserve, compensation, and the body’s readiness to absorb a major reconstructive intervention. Younger patients with stronger contralateral flexor performance may be entering surgery with a more favorable neuromuscular baseline, and machine learning may be able to translate that baseline into a more individualized prognosis. The next step is not a broader claim. It is a larger, independent cohort that can test whether the same predictors retain their value beyond this single retrospective experience.
For now, the most actionable message remains bounded. Preoperative age and strength data may help stratify expectations after ACL reconstruction, but no clinician should use this model as a stand-alone decision tool. The study ends where careful orthopedic prediction research should end: with a promising signal, a useful analytic framework, and a clear need for external validation before routine use.