DEVELOPING RISK PREDICTION MODELS FOR HIP OSTEOARTHRITIS BASED ON AUTOMATED HIP MORPHOLOGY MEASUREMENTS AND EVALUATING ON UNSEEN POPULATIONS: DATA OF THE WORLD COACH CONSORTIUM

INTRODUCTION Early identification of hip OA is crucial in enhancing our understanding of HOA development and treatment options. Hip morphology could be a modifiable risk factor for the development of radiographic hip osteoarthritis (RHOA), but the exact risk contribution of hip morphology in the general population remains unclear. By combining individual participant data (IPD) of various studies while considering study heterogeneity, novel modeling techniques could be explored to work towards individualized prediction models. OBJECTIVE To develop hip morphology based RHOA risk prediction models on multi-cohort datasets and assessed their generalizability to similar and unseen populations. METHODS We combined IPD from nine prospective cohort studies collected within the Worldwide Collaboration on OsteoArthritis prediCtion for the Hip (World COACH consortium). These studies all had standardized anteroposterior (AP) pelvic, long-limb, and/or hip radiographs taken and graded for RHOA at baseline and 4-8 years follow-up. Risk of incident RHOA was defined as hips with no signs of RHOA at baseline (any RHOA grade <2) which developed RHOA within this follow-up period (any RHOA grade ≥ 2). The lateral center edge angle (LCEA) and alpha angle (AA) were calculated automatically and relied on automated landmark placements on the outline of the hip (Figure 1). Included subjects had a mean age of 66.4 years (SD= 8.5), 71.3% was female, and mean body mass index (BMI) was 27.4 kg/m2 (SD=4.6).Risk prediction models were built with generalized linear mixed effects models (GLMM) and random forest models (RF). The discriminative performance (AUC) of models including the LCEA and AA measurements was compared to models based on hip side, sex, age, BMI and baseline RHOA grade alone. Stratified 5-fold cross-validation was performed to investigate the effect of a cohort specific intercept on predicted risk by a GLMM model. With leave-one-cohort-out cross-validation, the generalizability to a new population was evaluated for both GLMM and RF models. The mean AUC over the resulting test sets was compared in both settings. RESULTS In total, 35,922 hips without definite RHOA at baseline were included of which 4.7% developed RHOA within 4-8 years . Performance differences between the model configurations and between GLMM and RF models were small (Table 1). Using a marginal intercept instead of a cohort-specific intercept in the GLMM on caused a decrease (∼0.1 in AUC) in performance in the stratified 5-fold cross-validation. The leave-one-cohort-out cross-validation showed mean AUC values between 0.70-0.73. CONCLUSION In hips free of definite RHOA, we could fairly predict incident RHOA in both similar and unseen populations. However, the added value of hip morphology measurements on the discriminative performance is small.