Background
Measurement of patient reported outcome measures following anterior cruciate ligament (ACL) reconstruction enable the recognition of patients who have experienced an inferior outcome but who have not undergone subsequent revision surgery. While we can identify these patients based on their post-operative scores, the ability to accurately predict the risk of a poor subjective outcome pre-operatively remains challenging. Our hypothesis was that machine learning analysis has the potential to improve our predictive capability. We applied machine learning analysis of primary ACL reconstructions to the Norwegian Knee Ligament Register (NKLR) with the purpose of: (1) identifying the most important risk factors associated with experiencing a subjective failure of ACL reconstruction and (2) developing a clinically meaningful calculator for predicting the risk of an inferior outcome.
Methods
Machine learning analysis was performed on the NKLR dataset. The primary outcome measure was subjective failure of ACL reconstruction, defined as a Knee Injury and Osteoarthritis Outcome Score (KOOS) Quality of Life (QoL) subscale score less than 44 at two-years post-operatively. Data was split randomly into training (75%) and test (25%) sets. Four machine learning models were tested: Lasso logistic regression, random forest (RF), generalized additive model, and gradient boosted regression (GBM). Calibration and area under the curve (AUC) were calculated for all four models.
Results
In total, 20,818 patients met the inclusion criteria and complete follow-up KOOS data was available for 11,630 patients at two-years (56%). Of these, subjective failure was reported by 2,556 patients (22%). All models except the RF were well calibrated and the GAM exhibited the best performance (AUC 0.68; 95% CI 0.64-0.71). Inverse-probability weighted analysis suggested there were no significant differences between those with complete KOOS data and those with missing data. In total, six pre-operative variables were required for outcome prediction and an in-clinic calculator was developed which can estimate the risk of experiencing a subjective failure of primary ACL reconstruction (https://swastvedt.shinyapps.io/calculator_koosqol/). While the overall risk of subjective failure was 22%, this calculator can quantify risk at a patient-specific level.
Conclusions
Machine learning analysis of a national knee ligament register can predict the risk of a patient experiencing an inferior outcome and subjective failure after primary ACL reconstruction with moderate accuracy. This algorithm supports the creation of an in-clinic calculator for point-of-care risk stratification prior to surgery based on the input of six readily available factors. Interestingly, all predictors of subjective failure were patient-related and non-modifiable by the surgeon nor affected by surgical technique.