Feasibility of Machine Learning and Logistic Regression Algorithms to Predict Outcome in Orthopaedic Trauma Surgery

Area of Science:

• Orthopedic Surgery
• Data Science
• Biostatistics

Background:

• Machine learning (ML) models offer potential for improved probability estimates of binary events compared to traditional logistic regression.
• Musculoskeletal trauma research frequently involves predicting binary outcomes, making accurate statistical modeling crucial.

Purpose of the Study:

• To compare the probability estimation performance of ML models against logistic regression models in musculoskeletal trauma.
• To identify if ML models are influenced by different variables than logistic regression models in this clinical domain.

Main Methods:

• Developed and compared ML and logistic regression models using 9 datasets from musculoskeletal trauma studies.
• Models estimated probabilities for specific fractures and adverse events, utilizing 80% training and 20% testing data splits.
• Performance evaluated using discrimination (c-statistic), calibration (slope, intercept), and overall performance (Brier score) with fivefold cross-validation.

Main Results:

• Logistic regression models showed a slightly higher mean c-statistic (0.01) compared to the best ML models.
• ML models identified fewer strongly associated variables, with many differing from those in logistic regression.
• Overall performance metrics indicated comparable predictive capabilities between the two modeling approaches.

Conclusions:

• ML models provide probability estimates for binary events in musculoskeletal trauma that are comparable to logistic regression.
• The distinct benefits of ML models may be limited in this specific clinical context due to similar performance.
• Further research could explore specific scenarios where ML might offer advantages in musculoskeletal trauma prediction.