<![CDATA[Purpose – This study examines whether explicitly modeling ordinal severity relationships improves automated Knee Osteoarthritis (KOA) grading compared to conventional multi-class classification. Most existing studies treat KL grade prediction as a categorical problem, overlooking the ordered structure inherent to disease severity progression. Methods – A comparative experimental study was conducted using the publicly available Knee Osteoarthritis (KOA) dataset obtained from Kaggle (Tiwari, 2021). Although the original dataset contains 9,786 knee X-ray images categorized into five Kellgren–Lawrence (KL) severity grades, only the predefined training, validation, and testing subsets were utilized, resulting in a total of 8,260 images for model development and evaluation. Four classification approaches were evaluated under identical experimental conditions: conventional multi-class classification (baseline), Cumulative Link Model (CLM), Binary Cross-Entropy (BCE), and Consistent Rank Logits (CORAL), all using a pretrained ResNet-18 backbone. Performance was assessed using Accuracy, AUC, MAE, Quadratic Weighted Kappa (QWK), and F1-score. Findings – BCE achieved the highest observed performance among the evaluated approaches under the current predefined experimental setting, obtaining 67.81% ACC, 89.72% AUC, 0.3702 MAE, 0.8393 QWK, and 0.6750 F1-score. The baseline model remained competitive with 64.01% ACC and 0.7948 QWK, while CLM produced the lowest QWK (0.7713) and CORAL recorded the lowest F1-score (0.3769). Across all methods, Grade 1 remained the most challenging severity category, indicating persistent difficulty in distinguishing early-stage KOA cases. Research implications – The findings suggest that threshold-based ordinal decomposition may provide a practical balance between ordinal awareness and predictive flexibility for severity grading tasks under the current experimental setting. The study is limited to a single dataset and backbone architecture, and external validation is needed before broader clinical deployment. Originality – This study provides a controlled comparative evaluation of multiple ordinal regression formulations for KOA severity grading, demonstrating that ordinal learning effectiveness depends heavily on how rank relationships are modeled rather than on ordinal awareness alone.]]>
