Integrating deep learning-based radiographic features with demographic variables for hip osteoarthritis prediction:...

To evaluate whether hybrid models integrating deep learning (DL) features from baseline radiographs and demographic information can predict incident radiographic hip osteoarthritis (RHOA) and to compare their performance to models using radiographic or demographic information in isolation.

Data were pooled from seven prospective cohorts within the Worldwide Collaboration on OsteoArthritis prediction for the Hip (World COACH) consortium. Incident RHOA was defined as hips without RHOA at baseline that developed RHOA within four to eight years. From baseline radiographs, radiographic features were extracted with a DL-based feature extractor that was either unoptimized, optimized with baseline RHOA grades, or optimized with follow-up RHOA grades. The baseline RHOA grade acted as conventional reference. Baseline age, body mass index, and birth-assigned sex were considered as demographic predictors. Using internal and leave-one-cohort-out cross-validation, we assessed the predictive performance of models using only demographic predictors, only radiographic predictors, or both modalities combined (hybrid).

In the pooled dataset of 25,090 hips, 5.1% developed incident RHOA in four to eight years. Both the radiographic-only and hybrid models including follow-up-optimized radiographic features showed the best discriminative performance (AUC: 0.81 ± 0.01), whereas the demographic-only model showed an AUC of 0.71 (0.01) during internal cross-validation. The performance of the hybrid model including follow-up-optimized features dropped to an average AUC of 0.64 ± 0.09 during leave-one-cohort-out cross-validation.

Hybrid models combining DL-based features from baseline radiographs with demographic predictors show promise for RHOA risk identification, though demographics provided limited complementary information and performance declined in leave-one-cohort-out cross-validation.