JMIR Publications

ABSTRACT

Background:

Machine learning (ML) has been investigated for its predictive value in knee osteoarthritis (KOA) progression. However, systematic evidence on the effectiveness of ML is still lacking, posing a challenge to precision prevention.

Objective:

This meta-analysis intends to systematically review the application status and accuracy of ML in predicting KOA progression, and compare the predictive performance of ML, traditional methods, and deep learning (DL) under different datasets, model types, modeling variables, and definitions of KOA progression. The findings are expected to provide an evidence-based basis for developing clinical prediction tools.

Methods:

Following the PRISMA statement, a systematic search was conducted in Embase, Web of Science, PubMed, and Cochrane Library up to October 10, 2025. Observational studies that predicted KOA progression by ML models and reported metrics for assessing model accuracy were included. Two investigators were independently responsible for study screening, data extraction, and assessment of risk of bias in included studies using the Prediction Model Risk of Bias Assessment Tool (PROBAST). Meta-analyses were conducted on the concordance index (C-index) and diagnostic fourfold table using a random-effects model, with prediction intervals (PIs) reported. In addition, subgroup analyses were performed by the model type, modeling variable, and definition of KOA progression.

Results:

Thirty-two studies were finally included. For predicting all progression, the pooled C-index was 0.773 (95% CI: 0.727-0.821, 95% prediction interval(PI): 0.567-1.000) for the clinical feature-based model, 0.798 (95% CI: 0.755-0.843, 95% PI: 0.646-0.984) for the MRI-based model, 0.712 (95% CI: 0.657-0.772, 95% PI: 0.526-0.965) for the X-ray-based model, 0.806 (95% CI: 0.765-0.849, 95% PI: 0.639-1.000) for the MRI + clinical feature-based model, 0.772 (95% CI: 0.731-0.815, 95% PI: 0.610-0.976) for the X-ray + clinical feature-based model, and 0.731 (95% CI: 0.669-0.798, 95% PI: 0.518-1.000) for the clinical feature + X-ray + MRI-based model. The clinical feature-based model was established mainly by logistic regression, and exhibited accuracy comparable to other ML models. Among image-based models, traditional ML or DL possessed higher accuracy.

Conclusions:

ML demonstrates high accuracy in predicting KOA progression, but current evidence should be interpreted with caution due to the broad PI and risk of bias in included studies. Meanwhile, its clinical translation is restricted due to the heterogeneity in the definition of KOA progression, variations in validation strategies, and lack of external validation. Future research should standardize outcome metrics, optimize multimodal data fusion, and conduct external validation to improve model generalizability, thereby contributing to personalized risk stratification and early intervention in KOA.