Clinically aligned COPD severity prediction using ordinal neural networks

Chronic Obstructive Pulmonary Disease (COPD) requires accurate severity staging for treatment planning and prognosis. Machine learning models for COPD severity prediction typically treat Global Initiative for Chronic Obstructive Lung Disease (GOLD) stages as nominal categories, ignoring their natural ordering. We developed an ordinal neural network framework that explicitly models the ordered structure of GOLD stages (1-4) while learning from heterogeneous clinical datasets with differing feature sets. The model employs shared encoders for common clinical variables and private encoders for dataset-specific features, with value-mask encoding for missing data. Training used two publicly available COPD datasets (N = 224 source, N = 101 target) with stratified validation splits. The full shared-private ordinal model achieved 76.9% accuracy, mean absolute error 0.234 stages, and quadratic weighted kappa 0.894 on the validation set (N = 20). Ablation studies showed both encoder types are essential (removing either reduced accuracy to <30% with complete loss of ordinal agreement). Baseline comparisons demonstrated improvements over standard multiclass classification (37.3% accuracy, QWK 0.093) and logistic regression (57.8% accuracy, QWK 0.766). Over 95% of misclassifications occurred within ±1 GOLD stage. This work demonstrates that explicit ordinal modeling combined with heterogeneous data integration can achieve strong predictive performance for COPD severity staging, though validation on larger external cohorts is needed.