Spatially resolved longitudinal changes in tibial subchondral trabecular bone texture for exploratory prediction of...

Although prediction of radiographic knee osteoarthritis (rKOA) using baseline clinical and radiological data has been widely studied, the impact of temporal changes in imaging and clinical biomarkers remains less investigated. This study examined whether longitudinal changes in spatially resolved trabecular bone texture (TBT) improve prediction of incident rKOA. Participants with longitudinal data were selected from the Osteoarthritis Initiative. Incident rKOA was defined as Kellgren-Lawrence grade (KL) 0 at baseline and KL ≥ 2 at 48 months. One knee per participant was included (n=481). Fractal-based TBT descriptors were extracted from 16 automatically defined regions of interest across the tibial subchondral bone. Five logistic-regression models combined baseline clinical and radiological covariates, baseline TBT, and 24-month longitudinal changes (ΔTBT). Bidirectional stepwise selection based on the Akaike Information Criterion was applied once to TBT descriptors, whereas clinical and radiological covariates were incorporated a priori. The resulting fixed model specifications were evaluated using stratified 10-fold cross-validation repeated 300 times. Performance was assessed using area under the ROC curve (AUC), F1 score, balanced accuracy, positive and negative predictive values, and DeLong comparisons with the reference model. Adding baseline TBT to the clinical and radiographic reference model increased AUC, although not significantly (0.64 versus 0.53; p=0.055). Models incorporating ΔTBT yielded the highest apparent discrimination (AUC 0.76, 95% CI 0.70-0.82; p<0.001 versus the reference model), but threshold-dependent performance remained modest, with maximum balanced accuracy of 0.56, PPV of 0.22, and F1 score of 0.21. Selected baseline and longitudinal TBT descriptors were distributed across medial and lateral compartments, including regions near the central axis and subchondral cartilage. In this exploratory analysis, models incorporating spatially resolved 24-month TBT changes showed higher apparent ranking discrimination than the clinical and radiographic reference model. However, classification at the 0.5 threshold remained limited, and AUC improvement may be optimistic because feature selection preceded cross-validation. These findings support further investigation of ΔTBT as a candidate imaging marker but do not establish clinical utility. Leakage-free internal and external validation across independent cohorts and imaging systems are required.