Deep Learning Enabling Analysis of Exhaled Breath Using Fourier Transform Spectroscopy in the Mid-Infrared

Exhaled breath analysis is a promising noninvasive method for rapid diagnosis of diseases by detecting different types of volatile organic compounds (VOCs) that are used as biomarkers for early detection of various diseases such as lung cancer, diabetes, anemias, etc... and more recently COVID-19. Infrared spectroscopy seems to be a promising method for VOCs detection due to its ease of use, selectivity, and existence of compact low-cost devices. In this work, the use of Fourier transforms infrared (FTIR) spectrometer to analyze breath samples contained in a gas cell is investigated using deep learning and taking into account the practical performance limits of the spectrometer. Synthetic spectra are generated using infrared gas spectra databases to emulate real spectra resulted from a breath sample and train the neural network model (NNM). The dataset is generated in the spectral range of 2000 cm
^-1
to 6500 cm
^-1
and assuming a light-gas interaction length of 5 meters. The FTIR device performance is assumed with a signal-to-noise ratio (SNR) of 20,000:1 and a spectral resolution of 40 cm
^-1
. The proposed NNM contains a locally connected and 4 fully connected layers. The concentrations of 9 biomarker gases in the exhaled breath are predicted with r
²
score higher than 0.93, including carbon dioxide, water vapor, acetone, ethene, ammonia, methane, carbonyl sulfide, carbon monoxide and acetaldehyde demonstrating the possibility of detection.