Benchmarking longitudinal local centiles for radiographic knee osteoarthritis prognosis

Abstract

Cartilage thickness on MRI is a structural biomarker of osteoarthritis (OA), yet anatomical variation and aging can mask early disease, including changes occurring 12-36 months before incident OA. We propose longitudinal local centile maps that use normative growth charts as prior information and integrate a knee's earlier scans to identify cartilage that is thinner or thicker than expected. We built the reference charts from 5,045 scans of 957 persistently radiographically normal knees in the Osteoarthritis Initiative. Under participant-separated nested cross-validation, longitudinal centile profiles achieved higher 12- and 24-month AUCs for predicting incident radiographic OA than regional PCA, radiomics, cross-sectional centile features, and raw-map deep features across most classifiers. These conventional descriptors added little when combined with centile profiles, and cross-sectional centiles added little beyond longitudinal centiles. Deep centile features achieved the best AUCs (0.954 [95% CI 0.931-0.974] at 12 months and 0.969 [0.949-0.985] at 24 months; balanced accuracy 0.917 and 0.931) and outperformed raw-map deep features in all 12 classifiers after correction for multiple comparisons. No representation consistently improved 36-month prediction. Longitudinal centiles transform cartilage thickness maps into reference-based deviation maps: deep longitudinal centile features provide the highest predictive accuracy, while regional longitudinal centile profiles provide interpretable prognosis with the second-highest accuracy.