Effectiveness of lung cancer early diagnosis by analysing exhaled breath composition using neural network and multimodal approach

Background. The five-year survival rate of lung cancer patients remains extremely low, with the average rate of 22 %. Early detection of this disease can improve survival rates and reduce mortality. Lung cancer development is influenced by various risk factors, with smoking being the most significant, followed by other factors like occupational exposure, infections, genetic predisposition, presence of chronic diseases, etc. Considering risk factors is crucial for improving the efficacy of automated gas analysis complexes for lung cancer diagnosis. These complexes are promising for the application of scalable neural network data processing algorithms for noninvasive diagnosis of early stage lung cancer. Purpose of the study: to evaluate the effectiveness of the multimodal lung cancer detection method by analyzing the exhaled breath composition from 100 volunteers using simultaneous assessment of the exhaled breath composition and risk factors. Material and Methods. Along with exhaled breath samples, the study database also recorded all volunteers’ medical history data. Neural networks with a variety of architectures were used for data processing. The dataset for training neural network classifiers included exhaled breath samples from 100 volunteers, including 47 from healthy subjects and 53 from patients with morphologically confirmed lung cancer. Data determining the age group, smoking status and the presence of chronic lung diseases were analyzed as risk factors for lung cancer. Results. The integration of data, including exhaled breath composition and lung cancer risk factors, into a single neural network results in a 3 % increase in its original monomodal architecture. Taking into account a relatively small number of risk factors, such as age, gender, smoking status and COPD, increases the classifier’s sensitivity by 1.89 % and specificity by 6.39 %. The best generalization performance of the neural network classifier is achieved with a two-stream hybrid model with normalization. Conclusion. By incorporating diverse patient history data, such as age, smoking history, and chronic diseases, in addition to exhaled air composition data, into a unified neural network classifier, the accuracy of the exhaled air method for lung cancer diagnosis can be increased by an average of 4 %. This improvement, coupled with the expanded range of risk factors considered in the future application of the exhaled air method for lung cancer diagnosis in medical practice, will improve the reliability of population screening results.