A baseline study of interpretable machine learning using GC-MS breath VOCs for classifying asthma, bronchiectasis,...

Area of Science:

• Respiratory Medicine
• Analytical Chemistry
• Artificial Intelligence

Background:

• Accurate diagnosis of asthma, bronchiectasis, and COPD is clinically challenging due to overlapping symptoms.
• Conventional diagnostic tools have limitations in differentiating these respiratory conditions.
• Breath analysis (breathomics) offers a non-invasive approach for disease detection.

Purpose of the Study:

• To establish a reproducible method for classifying asthma, bronchiectasis, and COPD using exhaled breath volatile organic compounds (VOCs).
• To evaluate the performance of supervised machine learning classifiers for respiratory disease classification.
• To identify key VOC biomarkers associated with each disease using explainable AI.

Main Methods:

• Utilized a publicly available dataset of 121 breath samples and 76 VOCs.
• Employed gas chromatography-mass spectrometry (GC-MS) for breath sample analysis.
• Evaluated seven supervised classifiers under nested cross-validation, with XGBoost showing top performance.
• Applied Shapley Additive exPlanations (SHAP) for biomarker identification.

Main Results:

• XGBoost classifier achieved a mean accuracy of 95.83% and a macro-averaged AUC of 0.998.
• Identified influential VOCs, including 2-pentylfuran and hexadecane, as potential biomarkers.
• Demonstrated disease-specific and cross-disease relevance of identified VOCs.

Conclusions:

• Breathomics combined with explainable AI provides a scalable, non-invasive tool for respiratory disease classification.
• The study establishes a reproducible baseline for breathomics research in respiratory diseases.
• Identified VOCs serve as candidate biomarkers for future clinical validation and diagnostic development.