TWO-STAGE XGBOOST FRAMEWORK FOR SPIROMETRY BASED CLASSIFICATION OF ASTHMA AND COPD AND GOLD SEVERITY PREDICTION

Abstract

This study describes an explainable two-step XGBoost machine-learning pipeline for classifying spirometry reports as Normal, Asthma or Chronic Obstructive Pulmonary Disease followed by COPD staging according to the Global Initiative for Chronic Obstructive Lung Disease system. The study analysed 1600 spirometry reports, including 851 Normal, 320 Asthma and 429 COPD cases. Authors used demographic data, pre- and post bronchodilator measurements and bronchodilator response features as inputs. After data quality checks and leakage audits, we split the dataset in an 80:20 ratio to train and test the model. Authors applied randomized hyperparameter optimization and 5-fold stratified cross-validation during the training stage. To compare the performance, we trained two models: one used a full-report set of 22 features, while the other employed a pre bronchodilator feature set of 13 features. The full report model demonstrated an accuracy of 99.4% and a macro F1 score of 99.1%, outperforming the conventional bronchodilator reversibility rule. The pre-bronchodilator model achieved an accuracy of 91.6%, a macro-averaged ROC-AUC of 0.988 and a recall of 98.8% for COPD, with no COPD cases incorrectly classified as Normal in the test set. Finally, the second-step model correctly classified all 86 COPD cases in the test set, while the entire pipeline correctly classified six diagnostic and severity categories with an accuracy of 99.4%. The study revealed that post-bronchodilator FEV1/FVC, bronchodilator response and smoking history were the most relevant features for diagnosing Asthma, COPD or Normal spirometry, while post-bronchodilator FEV1 percentage predicted was a critical determinant for COPD staging. In conclusion, the study proved the efficacy of the explainable machine learning model for classifying and staging COPD.