Promises and limitations of deep learning for predicting knee osteoarthritis progression from medical imaging: A systematic review.
TL;DR
DL models demonstrate proof-of-concept for predicting KOA progression but require substantial improvements before clinical deployment, including rigorous multi-site validation and models that effectively leverage multimodal and longitudinal data.
Abstract
Purpose
To systematically evaluate the performance, methodological quality, and translational barriers of deep learning (DL) models for predicting knee osteoarthritis (KOA) progression from medical imaging.
Methods
Following PRISMA guidelines, we searched PubMed, Scopus, and Web of Science (inception to June 2026) for peer-reviewed studies applying DL to predict KOA progression from medical imaging. Two reviewers independently screened studies, extracted data, and assessed risk of bias using PROBAST-AI. The primary outcome was area under the receiver operating characteristic curve (AUC). Narrative synthesis was performed due to heterogeneity in imaging modalities, outcome definitions, and prediction horizons.
Results
From 1116 records, 33 studies (2019-2026) met the inclusion criteria, encompassing 50 predictive models. Sample sizes ranged from 340 to 52,981 knees (median: 4298). Nine progression definitions were identified and categorised as structural deterioration (n = 14), symptomatic worsening (n = 2), surgical endpoints (n = 8) or combined outcomes (n = 9). Nearly all studies (97%) used the Osteoarthritis Initiative dataset for training, and only 27% performed external validation. Surgery prediction models showed the highest descriptive median AUC (0.87), compared with structural (0.78) and symptomatic (0.79) progression models, but reflect non-disease factors including access to care. MRI-based models showed higher internal median AUCs (0.85) than radiograph-based models (0.81), but none underwent external validation. Externally validated models showed performance degradation (median AUC: 0.75 vs. 0.81 internal). Combining imaging modalities did not consistently improve predictions. However, studies leveraging longitudinal trajectories reported higher internal performance. Most studies showed low risk of bias in participant selection and outcome assessment, but 76% had unclear or high risk in the analysis domain.
Conclusions
DL models demonstrate proof-of-concept for predicting KOA progression but require substantial improvements before clinical deployment. Overcoming translational barriers requires standardised progression definitions integrating structural and symptomatic outcomes, rigorous multi-site validation, and models that effectively leverage multimodal and longitudinal data. LEVEL OF EVIDENCE N/A.