Abstract
Objectives:
Previous machine learning (ML) analysis has shown that a novel ML model built using the Multicenter Anterior Cruciate Ligament (ACL) Revision Study (MARS) cohort data can predict graft failure at 6 years postoperatively as a binary outcome with moderate ability. However, binary classification does not consider the time-to-event component of graft failure as an outcome. In contrast, survival models can factor the time-to-event component of an outcome, incorporate time-dependent variables, and appropriately process censored data to deliver more clinically meaningful outcome predictions. While the previous binary classifier can provide information about whether the primary outcome graft failure will occur at 6 years, survival analysis models are able to provide information on when graft failure occurs (for example, within 2, 4, or 6 years). Therefore, this study sought to apply ML survival analysis methodology to the MARS cohort data to produce a model that can most accurately estimate the probability of revision ACL reconstruction (rACLR) graft failure within 2, 4, and 6 years postoperatively, and identify factors that impact probability estimates at each of these time horizons.
Methods:
rACLR patients were prospectively recruited by the MARS group. Preoperative radiographs, surgeon-reported intraoperative findings, and 2 and 6-year follow-up data on patient-reported outcomes (PROs), subsequent knee surgeries, and graft failure were obtained. ML models including Cox Proportional Hazards (CoxPH), XGBoost Survival Embeddings (XGBoost), Random Survival Forest (RSF), Extra Survival Tree (EST), and CoxBoost, were built to predict the probability of graft failure within 2, 4, and 6 years postoperatively. Validated performance metrics and feature importance measures were used to evaluate model performance.
Results:
The cohort included 1142 patients, of which 4.8% (n = 55) experienced documented graft failure during the 6-year postoperative period. CoxBoost demonstrated the highest discriminative power across all studied time points (Timepoint: Time Dependent C-index | 2-year: 0.725 ± 0.127 | 4-year: 0.669 ± 0.057 | 6-year: 0.645 ± 0.054), with well-calibrated scores (Timepoint: Time dependent Brier Score | 2-year: 0.023 ± 0.012 | 4-year: 0.035 ± 0.010 | 6-year: 0.046 ± 0.012) as listed in Table 1 and Table 2 respectively. Based on the Kaplan Meier survival plot, the survival probability (probability of not tearing the ACL graft) at 2, 4, 6 years was 0.977, 0.965, and 0.954 respectively. Within the context of the predictive models, features deemed important for CoxBoost differed from those deemed important for CoxPH predictive ability.
Conclusions:
Survival machine learning models can predict the risk of rACLR graft failure up to 6 years postoperatively, with the most robust predictions at the 2-year postoperative timepoint. While external validation with further combined registry datasets is required, this present study builds on prior ML analyses towards the development of a bedside calculator for rACLR outcome prediction and risk stratification.
