A lightweight intelligent model for VOC mixture analysis: toward preclinical breath biomarker analysis.

For the current noninvasive lung cancer screening methods based on volatile organic compounds (VOCs) using electronic noses (e-noses), existing approaches still face limitations in modeling the long-range dependencies of sensor responses, the cross-channel global correlations, and the long-term trend features during the steady-state phase. Moreover, the associated deep learning models are often structurally complex and rely heavily on manual feature engineering, which restricts the engineering application and clinical translation of e-nose systems. To address these issues, this study proposes a lightweight global-local feature fusion framework for complex VOC sensing tasks and designs an efficient, lightweight feature extraction module (LFE) to achieve high-efficiency gas classification. For quantitative analysis of gas components, a GBDT-GRU Joint Prediction Model (JGPM) is introduced, effectively modeling the temporal evolution characteristics of sensor response signals. The above models were systematically validated using an e-nose experimental platform with synthetic gases of acetone, ethanol, isopropanol, and their mixtures at the ppm level as a proof-of-concept (PoC) study. The experimental results show that the proposed models outperform the comparative methods in both gas classification accuracy and concentration prediction performance, while maintaining low model complexity. Although current validation is at the preclinical stage, this framework provides a robust algorithmic foundation for future intelligent gas sensing and clinical breath-based disease screening.