A deep learning-based system for fast and automated detection of volatile organic compounds in raw gas chromatography-mass spectrometry breath data

Through their breath, humans exhale hundreds of volatile organic compounds (VOCs) that can reveal a large spectrum of health conditions, including many types of cancer in the early stages. Therefore, it is thought that breath analysis can provide fast, accurate and non-invasive diagnostics. Gas chromatography-mass spectrometry (GC-MS) is an analytical method commonly used to measure VOCs in breath samples. The subsequent VOC detection is based on the recognition of their specific ion patterns in GC-MS data, which requires expert-driven data pre-processing that is time-consuming, subjective and may introduce errors. This thesis explores the original idea of applying deep learning pattern recognition abilities, and in particular convolutional neural networks (CNNs), to learn ion patterns in the complex data from GC-MS processing. The proposed approach provides a comprehensive system for fast and automated detection of any set of VOCs directly from raw GC-MS data, thus bypassing expert-led analysis. The developed system showed to outperform the state-of-the-art method by detecting a significantly higher number of VOCs in just a fraction of time while maintaining high specificity. Furthermore, automated quantification of the detected VOCs is introduced. The comprehensive system, combining compound detection and quantification, returns robust results that allow for accurate samples classification, proved on the clinical breath data obtained from participants before and after a radiation dose. The achieved results suggest a practical use of the proposed approach to analyse GC-MS data efficiently. This novel method can drastically reduce cost and time and improve the reliability of breath analysis, significantly supporting a game-changing opportunity for advanced breath-based medical diagnostics.