MOVING BEYOND THE KELLGREN-LAWRENCE GRADE: A DEEP-LEARNING-BASED RADIOMICS SCORE FOR PREDICTING KNEE REPLACEMENT RISK.

INTRODUCTION: KL grading is widely used for radiographic knee osteoarthritis assessment but remains categorical, reader-dependent, and limited in prognostic granularity. An automated continuous imaging biomarker from standard radiographs may improve knee replacement (KR) risk prediction. OBJECTIVE: To develop a deep-learning-based radiomics score (DR Score) from posteroanterior knee radiographs and compare its prognostic performance with KL grade for KR risk prediction. METHODS: We analyzed baseline standing posteroanterior knee radiographs from MOST and OAI for model development and internal testing, including 6,578 participants and 12,905 knees. External testing used MenTOR and KICK, including 135 participants. An automated deep-learning pipeline localized the knee joint, extracted patch-level radiographic features, aggregated informative regions using attention-based multiple-instance learning, and generated a continuous 0-4 DR Score using DeepSurv survival modelling. The outcome was time to KR in the US cohorts and virtual KR in the UK cohorts. Predictive performance was assessed using C-index and integrated time-dependent AUC, with bootstrap confidence intervals and permutation testing. RESULTS: In internal testing, the DR Score achieved a C-index of 0.849 (95% CI: 0.826-0.871), higher than KL grade alone (0.821, 95% CI: 0.795-0.851; p = 0.037). At 108 months, the DR Score also showed higher iAUC than KL grade (0.883, 95% CI: 0.859-0.904 vs 0.856, 95% CI: 0.830-0.880; p = 0.019). Combining DR Score with KL grade further improved performance, with a C-index of 0.863 (95% CI: 0.838-0.889) and 108-month iAUC of 0.903 (95% CI: 0.878-0.923). In external testing, the DR Score showed similar discrimination to KL grade, with C-indices of 0.755 and 0.751, respectively. The combined model achieved the highest external C-index of 0.804. The DR Score was moderately correlated with KL grade (ρ = 0.696, p < 0.001) and remained independently associated with KR after adjustment for age, sex, BMI, and KL grade. CONCLUSION: The DR Score is an automated, continuous radiographic biomarker for KR risk prediction. It provides prognostic information complementary to KL grade and may support scalable risk stratification in knee osteoarthritis.