Application of machine learning for the diagnosis of some socially significant diseases from an exhaled human air by the infrared laser spectroscopy

The infrared spectra of the air exhaled by several groups of volunteers were studied: those suffering from type 1 diabetes, bronchial asthma, and pneumonia. To record infrared spectra, a tunable quantum-cascade laser (QCL) was used. QCL emits in the wavelength range from 5.3 to 12.8 $\mu$m in a pulsed mode with a pulse width of 50 ns, a power of up to 150 mW, and a tuning step of 1 cm$^{-1}$. The laser is optically coupled to an astigmatic Herriot gas cell with an optical path length of 76 m. A difference was found in the intensity of selective lines of biomarker molecules in the spectra of exhaled air of healthy volunteers compared to similar indicators of volunteers suffering from a certain disease. For an example of methods such as the support vector machine (SVM), the $k$-nearest neighbors ($k$-NN) and the random forest algorithm (Random Forest), the possibility of classifying volunteers by the infrared spectra of their exhaled air is shown. In terms of the accuracy metric, the accuracy of disease classification improved to 98% by the use of dimensionality reduction techniques (PCA and $t$-SNE).