Exhaled breath contains thousands of volatile organic compounds (VOCs) that serve as biomarkers of metabolic activity and physiological states. Electronic nose (e-nose) systems provide a non-invasive, efficient, and cost-effective platform for disease detection through VOC profiling. However, conventional ensemble learning methods for e-nose signal analysis often exhibit limited generalization and high computational cost. To overcome these limitations, we proposed the adaptive ensemble learning (AEL) framework that autonomously selects optimal feature extraction methods, base classifiers, cross-validation pattern, and ensemble strategies based on the feature distribution of the dataset. Using a self-developed e-nose equipped with 11 gas sensors and a temperature and humidity sensor, we analyzed breath samples from patients with lung cancer, chronic obstructive pulmonary disease (COPD), and healthy controls. The AEL model achieved exceptional performance in binary classification (lung cancer vs. healthy), with accuracy, recall, specificity, precision, and F1-score of 0.9642, 0.9462, 0.9873, 0.99, and 0.966, respectively. In multiclass classification (lung cancer, COPD, healthy), it achieved 0.8489 accuracy, 0.8612 recall, 0.9321 specificity, 0.8662 precision, and 0.8493 F1-score, exceeding recently reported benchmarks. This exploratory study highlights the potential of the proposed AEL coupled with a self-developed electronic nose system for reliable and noninvasive screening of pulmonary diseases. • An 11-sensor e-nose enables non-invasive lung-cancer and COPD screening via exhaled. • Adaptive Ensemble Learning (AEL) dynamically adapts to data, enhancing generalization. • Coupled with breath analysis, AEL achieves 0.964 LC and 0.849 multi-class accuracy.
Tang et al. (Thu,) studied this question.