Enhancing breath-based diagnostics through eXplainable Artificial Intelligence

Breath analysis is emerging as a non-invasive and promising diagnostic approach capable of assessing a patient's metabolic state by detecting volatile organic compounds in exhaled breath. This study investigates the potential of breath analysis for the early detection of lung cancer, respiratory and gastrointestinal diseases using open-access data from three distinct datasets. An artificial intelligence methodology is implemented to predict diagnostic labels while addressing class imbalance, an inherent challenge in medical datasets. After evaluating model performance and stability, the most relevant volatile organic compounds identified by the best-performing model for each dataset are analyzed. Using eXplainable Artificial Intelligence, the influence of volatile organic compound abundances on predictions is examined, enabling the identification of key variables and improving model interpretability. The proposed methodology provides a robust framework for breath-based diagnostics, emphasizing the potential of integrating breath analysis with machine learning to advance clinical decision-making despite ongoing challenges related to sampling variability, detection sensitivity, and standardization across studies.