Promises and limitations of deep learning for predicting knee osteoarthritis progression from medical imaging: A systematic review.
Amna Gillani, Rüdiger von Eisenhart-Rothe, Klaus Woertler et al.
✦ AI-curated · Sources linked
In 30 seconds
This PRISMA-guided systematic review assessed deep-learning models that predict knee osteoarthritis progression from medical imaging. It included 33 studies published from 2019 to 2026, covering 50 models and varied progression definitions and prediction horizons. Surgery prediction had the highest descriptive median AUC (0.87), although these outcomes may reflect factors beyond disease progression. Only 27% of studies performed external validation, and externally validated models had lower median AUC than internally assessed models (0.75 vs. 0.81). The authors judged the evidence proof-of-concept, with substantial methodological and translational barriers to clinical deployment.
Key findings
- Across 50 models, surgery prediction had a descriptive median AUC of 0.87, versus 0.78 for structural and 0.79 for symptomatic progression.
- Only 27% of studies performed external validation; median AUC was 0.75 in externally validated models versus 0.81 internally.
- MRI-based models had higher internal median AUCs than radiograph-based models (0.85 vs. 0.81), but none underwent external validation.
- Analysis-domain risk of bias was unclear or high in 76% of studies.
Why it matters
Imaging-based prediction could help characterise future knee osteoarthritis progression, but reported performance is difficult to interpret across inconsistent outcome definitions and prediction horizons. The limited external validation and frequent analysis-domain concerns mean that strong internal results do not yet establish reliable performance in other settings. Distinguishing structural change, symptoms, and surgery is also important because these outcomes capture different aspects of progression.
What to keep in mind
The review found substantial heterogeneity in imaging modalities, progression definitions, and prediction horizons, precluding quantitative synthesis. Nearly all studies used the Osteoarthritis Initiative dataset for training, and external validation was limited.