Optimal inputs for machine learning models in predicting total joint arthroplasty outcomes: a systematic review.

Area of Science:

• Orthopedic Surgery
• Medical Informatics
• Data Science in Healthcare

Background:

• Machine learning (ML) models offer potential for improving outcomes and reducing complications in total joint arthroplasty (TJA).
• Implementing ML models is challenging due to the variety of models and potential input variables.
• A systematic review was conducted to identify optimal ML model inputs for predicting postoperative outcomes in TJA.

Purpose of the Study:

• To systematically review and assess the most effective inputs for machine learning models.
• To predict postoperative medical outcomes, orthopedic outcomes, and patient-reported outcome measures (PROMs) after TJA.
• To identify key predictive factors for TJA complications.

Main Methods:

• A systematic literature search was performed across PubMed, MEDLINE, EBSCOhost, and Google Scholar (2000-2023).
• 25 relevant studies evaluating over 20 ML models and 1,555,300 surgeries were included.
• Key metrics analyzed included Area Under the Curve (AUC), accuracy, and input variable importance.

Main Results:

• ML models for medical complications showed AUCs from 0.57-0.997; key predictors included age, hyper-coagulopathy, and malnutrition.
• ML models for orthopedic complications had AUCs from 0.49-0.93; age, BMI, and CCI were significant predictors.
• For PROMs, ML models achieved AUCs from 0.453-0.97, with preoperative PROMs being a key input. Age was the most consistent predictor across outcomes.

Conclusions:

• ML models effectively predict complications and outcomes in TJA, identifying both known and novel risk factors.
• Age emerged as the most significant predictor of postoperative complications.
• Future research should focus on standardized model development, validated inputs, and external validation for ML in TJA.