Enhancing breath-based diagnostics through eXplainable Artificial Intelligence

Area of Science:

• Biomedical Engineering
• Computational Biology
• Medical Diagnostics

Background:

• Breath analysis offers a non-invasive method to assess metabolic states via volatile organic compounds (VOCs).
• Early disease detection, particularly for lung cancer, remains a critical challenge in healthcare.

Purpose of the Study:

• To investigate the efficacy of breath analysis for early detection of lung cancer, respiratory, and gastrointestinal diseases.
• To implement and evaluate an artificial intelligence (AI) methodology for disease prediction using breath VOCs.
• To identify key VOCs and enhance model interpretability using eXplainable AI (XAI).

Main Methods:

• Utilized open-access datasets comprising breath analysis data.
• Applied AI models to predict diagnostic labels, incorporating strategies to handle class imbalance.
• Employed XAI techniques to analyze the influence of VOC abundances on model predictions.

Main Results:

• Successfully developed and evaluated AI models for disease prediction from breath samples.
• Identified specific VOCs as relevant biomarkers for different diseases across datasets.
• Demonstrated the utility of XAI in pinpointing influential VOCs and improving model transparency.

Conclusions:

• Breath analysis integrated with AI presents a robust framework for non-invasive disease diagnostics.
• The methodology highlights the potential for advancing clinical decision-making through personalized breath-based biomarkers.
• Further research is needed to address challenges in standardization, sensitivity, and sampling variability for clinical translation.