Assessing the Value of Imaging Data in Machine Learning Models to Predict Patient-Reported Outcome Measures in Knee...

Area of Science:

• Orthopedics
• Artificial Intelligence
• Medical Imaging

Background:

• Knee osteoarthritis (OA) impacts over 650 million globally, with total knee replacement for end-stage disease.
• The utility of imaging in tracking symptomatic knee OA progression requires further clarification.

Purpose of the Study:

• To compare machine learning (ML) models with and without imaging features for predicting two-year Western Ontario and McMaster Universities Arthritis Index (WOMAC) scores in knee OA patients.
• To evaluate the performance of ML models in predicting clinically meaningful physical impairment.

Main Methods:

• Utilized data from 2408 patients (Osteoarthritis Initiative) and 629 patients (Multicenter Osteoarthritis Study).
• Developed ML models using 18 clinical features and an expanded dataset with 10 additional imaging features.
• Employed Gradient Boosting Machine (GBM) for external validation, setting Minimal Clinically Important Difference (MCID) at 24.

Main Results:

• Clinical and imaging ML models demonstrated comparable predictive performance (AUC < 0.025).
• GBM models achieved the best performance in external validation for both datasets, with AUCs of 0.734 and 0.747, respectively.
• Key predictive features included education, family history of OA, comorbidities, osteoporosis medication use, and prior knee procedures.

Conclusions:

• ML models achieve comparable performance in predicting knee OA progression whether or not imaging features are included.
• This study is the first to demonstrate that imaging data does not significantly enhance ML model performance for predicting WOMAC scores in knee OA.
• Clinical factors alone are sufficient for accurate prediction of knee OA symptom progression using ML.