Rapid, Non-Invasive Breath Analysis for Enhancing Detection of Silicosis Using Mass Spectrometry and Interpretable Machine Learning.

Abstract

Occupational lung diseases, such as silicosis, are a significant global health concern, especially with increasing exposure to engineered stone dust. Early detection of silicosis is helpful for preventing disease progression, but existing diagnostic methods, including X-rays, computed tomography scans, and spirometry, often detect the disease only at late stages. This study investigates a rapid, non-invasive diagnostic approach using atmospheric pressure chemical ionization-mass spectrometry (APCI-MS) to analyse volatile organic compounds (VOCs) in exhaled breath from 31 silicosis patients and 60 healthy controls. Six different interpretable machine learning (ML) models with Shapley additive explanations (SHAP) were applied to classify these samples and determine VOC features that contribute the most significantly to model accuracy. The extreme gradient boosting classifier demonstrated the best performance, achieving an area under the receiver-operator characteristic curve of 0.933 with the top ten SHAP features. The m/z 442 feature, potentially corresponding to leukotriene-E3, emerged as a significant predictor for silicosis. The VOC sampling and measurement process takes less than five minutes per sample, highlighting its potential suitability for large-scale population screening. Moreover, the ML models are interpretable through SHAP, providing insights into the features contributing to the model's predictions. This study suggests that APCI-MS breath analysis could enable early and non-invasive diagnosis of silicosis, helping to improve disease outcomes.