Clinically aligned COPD severity prediction using ordinal neural networks.

Area of Science:

• Computational medicine and artificial intelligence in respiratory disease research.
• Development of advanced machine learning models for clinical outcome prediction.

Background:

• Accurate staging of Chronic Obstructive Pulmonary Disease (COPD) is crucial for effective treatment and prognosis.
• Existing machine learning models often treat COPD severity stages as independent categories, neglecting their inherent order.

Purpose of the Study:

• To develop and evaluate an ordinal neural network framework for COPD severity staging that explicitly models the ordered structure of Global Initiative for Chronic Obstructive Lung Disease (GOLD) stages.
• To integrate heterogeneous clinical datasets with differing feature sets using shared and private encoders.

Main Methods:

• An ordinal neural network framework was designed, incorporating shared encoders for common features and private encoders for dataset-specific features.
• Value-mask encoding was utilized to handle missing data.
• The model was trained and validated on two publicly available COPD datasets.

Main Results:

• The full shared-private ordinal model achieved 76.9% accuracy, a mean absolute error of 0.234 stages, and a quadratic weighted kappa (QWK) of 0.894 on the validation set.
• Ablation studies confirmed the necessity of both shared and private encoders for optimal performance.
• The ordinal model significantly outperformed standard multiclass classification and logistic regression baselines.

Conclusions:

• Explicit ordinal modeling, combined with heterogeneous data integration, offers a powerful approach for accurate COPD severity staging.
• The developed framework demonstrates strong predictive performance, with most misclassifications occurring within a single stage.
• Further validation on larger external cohorts is recommended to confirm generalizability.