Background
Estimating the risk of revision after arthroplasty could inform patient and surgeon decision-making. However, there is a lack of well-performing prediction models assisting in this task, which may be due to current conventional modeling approaches such as traditional survivorship estimators (such as Kaplan-Meier) or competing risk estimators. Recent advances in machine learning survival analysis might improve decision support tools in this setting. Therefore, this study aimed to assess the performance of machine learning compared with that of conventional modeling to predict revision after arthroplasty.

Question/purpose
Does machine learning perform better than traditional regression models for estimating the risk of revision for patients undergoing hip or knee arthroplasty?

Methods
Eleven datasets from published studies from the Dutch Arthroplasty Register reporting on factors associated with revision or survival after partial or total knee and hip arthroplasty between 2018 and 2022 were included in our study. The 11 datasets were observational registry studies, with a sample size ranging from 3038 to 218,214 procedures. We developed a set of time-to-event models for each dataset, leading to 11 comparisons. A set of predictors (factors associated with revision surgery) was identified based on the variables that were selected in the included studies. We assessed the predictive performance of two state-of-the-art statistical time-to-event models for 1-, 2-, and 3-year follow-up: a Fine and Gray model (which models the cumulative incidence of revision) and a cause-specific Cox model (which models the hazard of revision). These were compared with a machine-learning approach (a random survival forest model, which is a decision tree–based machine-learning algorithm for time-to-event analysis). Performance was assessed according to discriminative ability (time-dependent area under the receiver operating curve), calibration (slope and intercept), and overall prediction error (scaled Brier score). Discrimination, known as the area under the receiver operating characteristic curve, measures the model’s ability to distinguish patients who achieved the outcomes from those who did not and ranges from 0.5 to 1.0, with 1.0 indicating the highest discrimination score and 0.50 the lowest. Calibration plots the predicted versus the observed probabilities; a perfect plot has an intercept of 0 and a slope of 1. The Brier score calculates a composite of discrimination and calibration, with 0 indicating perfect prediction and 1 the poorest. A scaled version of the Brier score, 1 – (model Brier score/null model Brier score), can be interpreted as the amount of overall prediction error.

Results
Using machine learning survivorship analysis, we found no differences between the competing risks estimator and traditional regression models for patients undergoing arthroplasty in terms of discriminative ability (patients who received a revision compared with those who did not). We found no consistent differences between the validated performance (time-dependent area under the receiver operating characteristic curve) of different modeling approaches because these values ranged between -0.04 and 0.03 across the 11 datasets (the time-dependent area under the receiver operating characteristic curve of the models across 11 datasets ranged between 0.52 to 0.68). In addition, the calibration metrics and scaled Brier scores produced comparable estimates, showing no advantage of machine learning over traditional regression models.

Conclusion
Machine learning did not outperform traditional regression models.

Clinical Relevance
Neither machine learning modeling nor traditional regression methods were sufficiently accurate in order to offer prognostic information when predicting revision arthroplasty. The benefit of these modeling approaches may be limited in this context.

@en

Background:
Statistical models using machine learning (ML) have the potential for more accurate estimates of the probability of binary events than logistic regression. The present study used existing data sets from large musculoskeletal trauma trials to address the following study questions: (1) Do ML models produce better probability estimates than logistic regression models? (2) Are ML models influenced by different variables than logistic regression models?

Methods:
We created ML and logistic regression models that estimated the probability of a specific fracture (posterior malleolar involvement in distal spiral tibial shaft and ankle fractures, scaphoid fracture, and distal radial fracture) or adverse event (subsequent surgery [after distal biceps repair or tibial shaft fracture], surgical site infection, and postoperative delirium) using 9 data sets from published musculoskeletal trauma studies. Each data set was split into training (80%) and test (20%) subsets. Fivefold cross-validation of the training set was used to develop the ML models. The best-performing model was then assessed in the independent testing data. Performance was assessed by (1) discrimination (c-statistic), (2) calibration (slope and intercept), and (3) overall performance (Brier score).

Results:
The mean c-statistic was 0.01 higher for the logistic regression models compared with the best ML models for each data set (range, −0.01 to 0.06). There were fewer variables strongly associated with variation in the ML models, and many were dissimilar from those in the logistic regression models.

Conclusions:
The observation that ML models produce probability estimates comparable with logistic regression models for binary events in musculoskeletal trauma suggests that their benefit may be limited in this context.@en