A fast, non-invasive auxiliary screening algorithm for lung cancer based on electronic nose system

Abstract

Lung cancer is the second most common cancer worldwide, with conventional screening methods, including imaging and pathological diagnostics, facing challenges such as high misdiagnosis rates, radiation exposure, and invasiveness. In contrast, human breath analysis based on electronic nose technology offers a promising non-invasive, rapid, and cost-effective approach for large-scale lung cancer screening. To improve the accuracy of breath-based screening in the pattern recognition module of Electronic Nose technology, a Transformer-based architecture, termed Gasformer, has been specifically developed. This model is designed to detect low-concentration volatile organic compounds (VOCs) in complex human breath and is optimized to perform effectively even on small-sample datasets. Channel and self-attention mechanisms are incorporated into a restructured encoder to enhance the allocation of sensor signals and temporal positions. The feature representations of the encoder are decoded using additive attention, enabling accurate mapping to label space for exhalation recognition. Evaluation through three-fold cross-validation on a dataset of actual breath samples demonstrates the superior performance of Gasformer, achieving average classification accuracy, sensitivity, and specificity of 0.974 ± 0.036, 1.000 ± 0.000, and 0.944 ± 0.079, respectively. These results surpass those of traditional machine learning algorithms such as K-nearest Neighbors (KNN), Support Vector Machines (SVM), and Decision Trees (DT), as well as advanced deep learning models, including Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN), with or without attention mechanisms. The findings highlight the potential of Gasformer as a reliable and effective tool for preliminary screening of lung cancer and other diseases through breath analysis.