Identifying who are unlikely to benefit from total knee arthroplasty using machine learning models

Area of Science:

• Orthopedics
• Medical Artificial Intelligence
• Health Economics

Background:

• Total knee arthroplasty (TKA) is a common procedure for knee osteoarthritis (OA).
• Predicting long-term patient outcomes and satisfaction after TKA is crucial for healthcare expenditure and patient well-being.
• Identifying patients unlikely to benefit from TKA can optimize resource allocation.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for predicting 2-year postoperative dissatisfaction following TKA.
• To compare the predictive performance of image-only, clinical-data-only, and multimodal ML models.
• To assess the potential of ML in identifying patients who may not benefit from TKA.

Main Methods:

• Trained ML models (image-only, clinical-data only, multimodal) on a dataset of 5720 knee OA patients.
• Defined postoperative dissatisfaction based on minimal clinically important differences in Knee Society Scores (KSS), SF-36 (PCS, MCS), and Oxford Knee Score (OKS).
• Evaluated model performance using Area Under the Curve (AUC) metrics.

Main Results:

• Clinical-data-only and multimodal ML models significantly outperformed image-only models in predicting TKA dissatisfaction.
• The clinical-data-only model achieved AUCs ranging from 0.806 (OKS) to 0.888 (KSS).
• The multimodal model demonstrated comparable performance, with AUCs ranging from 0.816 (OKS) to 0.891 (KSS).

Conclusions:

• ML models incorporating clinical data are effective in predicting postoperative dissatisfaction after TKA.
• These models can aid in identifying patients unlikely to achieve long-term benefit from TKA.
• Predictive modeling holds promise for improving patient selection and reducing healthcare costs associated with TKA.