TOWARDS BREATH-BASED DISEASE DIAGNOSIS: A MACHINE LEARNING PIPELINE ON PTR-MS DERIVED VOC SIGNATURES

Abstract

Diseases such as lung cancer are often difficult to diagnose in their early stages, and by the time detection occurs, the disease may have already progressed to an aggressive phase. Noninvasive methods like exhaled breath analysis offer a promising alternative that can facilitate early detection in a painless and accessible manner. Even established techniques such as Reverse Transcription Polymerase Chain Reaction (RT-PCR ), commonly used for COVID-19 detection, fall short in delivering comprehensive diagnostic coverage. In this study, we leverage Proton Transfer Reaction - Time of Flight - Mass Spectrometry (PTR-TOF-MS) to analyze exhaled breath samples, enabling high-resolution detection of volatile organic compounds (VOCs). By examining the concentration profiles of these VOCs, we aim to identify disease-specific molecular signatures, or “fingerprints,” that distinguish between health conditions. Each disease produces a unique VOC signature, making real-time, noninvasive diagnosis possible. A key contribution of this work is the development of a cascading classification model tailored for real-time disease prediction. Despite working with a relatively limited dataset, the model achieves classification accuracies of up to 96%, highlighting the potential of breath-based diagnostics. In a separate approach, we employed Cohen’s dd effect size method to identify statistically significant VOCs that differentiate disease groups. This statistical technique serves as a powerful tool for biomarker discovery, enabling more interpretable models and guiding future research on disease-specific breath signatures. These results strongly suggest that expanding the dataset could further enhance model performance and generalizability, paving the way for scalable and rapid clinical screening tools.