Machine Learning Did Not Outperform Conventional Competing Risk Modeling to Predict Revision Arthroplasty

Area of Science:

• Orthopedic Surgery
• Medical Data Science
• Predictive Analytics

Background:

• Accurate prediction of revision risk after arthroplasty is crucial for patient and surgeon decision-making.
• Current prediction models often lack sufficient performance, potentially due to conventional statistical approaches.
• Machine learning (ML) survival analysis offers a potential advancement for improving decision support tools.

Purpose of the Study:

• To compare the performance of ML survival analysis with traditional regression models in predicting revision risk after hip and knee arthroplasty.
• To determine if ML models offer superior predictive accuracy compared to conventional methods.

Main Methods:

• Eleven observational registry datasets from the Dutch Arthroplasty Register (2018-2022) were analyzed.
• Time-to-event models, including Fine and Gray, cause-specific Cox, and random survival forest (ML), were developed and compared.
• Model performance was evaluated using discriminative ability (time-dependent AUC), calibration (slope/intercept), and prediction error (scaled Brier score).

Main Results:

• No significant differences in discriminative ability were found between ML and traditional regression models (time-dependent AUC range: 0.52-0.68).
• Calibration metrics and scaled Brier scores also showed comparable performance across all modeling approaches.
• Machine learning did not demonstrate an advantage over traditional regression models in predicting arthroplasty revision.

Conclusions:

• Machine learning models did not outperform traditional regression models for predicting revision after arthroplasty.
• Current ML and traditional regression methods lack sufficient accuracy for providing prognostic information in this context.
• The clinical utility of these modeling approaches for predicting revision arthroplasty may be limited.