Commentary Getting a young, pivoting athlete with an anterior cruciate ligament (ACL) tear back to sport within 1 year postoperatively, with sustained high-level performance and without physical or psychological limitations for many years afterward, continues to be the primary challenge for clinicians treating such patients. For the most part, we have succeeded at this with the use of modern surgical and rehabilitation techniques, but there remains a larger-than-desired subset of patients for whom this goal is difficult to achieve. With too many failures still occurring due to poor surgical technique, a well-done ACL reconstruction (ACLR) still reigns as the most important factor for clinical success. Assuming that a well-done ACLR can be performed, a perfect machine-learning (ML) prediction model would be able to quantify failure risk prior to surgery to help us prioritize modifiable risk factors that can be acted upon early in the treatment process and improve the course of recovery. Current ML models are nice proofs of concept but are far from being the "holy grail" prediction tools that they strive to be. ML requires big data sets, and, naturally, our best sources of such data sets are national registries. The study by Anderson and colleagues used the Danish Knee Ligament Reconstruction Registry to develop an enhanced ML-Cox regression model for the prediction of ACLR revision surgery. Using only 4 variables (patient age at the time of primary ACLR and 3 Knee injury and Osteoarthritis Outcome Score KOOS items obtained 12 months postoperatively), the model demonstrated good prediction accuracy at 2, 3, and 6 years following primary ACLR. However, the intelligence that we gain is only as reliable as the inputted data. The data used to train the model were pulled from an expansive timeframe (2005 to 2023) during which practice patterns likely changed. Furthermore, the study cohort consisted largely of patients with hamstring autografts (80%) and patients without meniscal treatment (60%), and the median time from injury to surgery was 7 months. Whether this model would be accurate for other populations such as the U.S. population, in whom patellar tendon and quadriceps autografts predominate, is unknown—and, certainly, the model requires external validation. As is the case with many of the current ACL registries, important anatomic risk factors, such as coronal and sagittal deformity and joint hyperlaxity, as well as concomitant lateral extra-articular procedures, were not available as variables that could be inputted during the construction of the ML model. Three of the required variables were 12-month postoperative KOOS items, thereby restricting the clinical use of the model to after 1 year, when many patients may have already been cleared to return to sport on the basis of objective strength and dynamic movement criteria1. A deeper dive into these 12-month KOOS items demonstrates that they may merely represent the psychological readiness of the patient to return to activities2. The question, then, is whether a patient's increased knee pain, modification of lifestyle to avoid damaging activities to the knee, and lack of confidence in the knee (corresponding to the 3 KOOS items) can be modified at all after 1 year. Lastly, the study defined primary failure as ACLR revision, thereby underestimating the true number of clinical failures and overlooking patients who were dissatisfied with their outcome, who lacked confidence in the knee during sport activities, or who sustained a rerupture but opted not to undergo revision surgery3. Clearly, current ML models to predict ACLR failure are still limited by the robustness of the data from which they were built. Accumulating more data and increasing the number of variables analyzed will likely improve these prediction algorithms, as well as their clinical utility. However, in the setting of ever-evolving surgical and rehabilitation strategies, such improvements will be challenging and take a lot of time. Until then, clinical skill and intuition, combined with a thorough assessment of an individual patient's risk factors (and the know-how to modify them), will still be essential in reducing ACLR failure risk in our young, pivoting athletes.
Dean Wang (Wed,) studied this question.