Tuberculosis severity classification from exhaled breath using an electronic nose system with Inception-1D and ResNet-1D

Abstract

Exhaled breath contains volatile organic compounds (VOCs) that can be analyzed for tuberculosis (TB) detection. Electronic nose systems have demonstrated promise for this application; however, accurately distinguishing between healthy individuals and TB patients with different severity levels, namely, low and high TB, remains challenging and requires advanced deep-learning methods. Unlike previous studies that focused on binary TB detection, this study proposes an electronic nose system combined with one-dimensional deep learning models, including residual network (ResNet), visual geometry group (VGG), EfficientNet, and Inception architectures, for multiclass classification of healthy individuals and TB severity levels using exhaled-breath analysis. The results show that Inception-1D and ResNet-1D achieve the best performance for healthy and TB classification, each attaining an F1-score of 94.99%. For TB severity classification, ResNet-1D outperforms other models with an F1-score of 68.50%. Meanwhile, Inception-1D yields the highest performance on the healthy and TB severity classification dataset, with an F1-score of 74.07%. Moreover, dimensionality reduction using principal component analysis (PCA) reduces the healthy and TB dataset to ten principal components and improves the F1-score to 95.64% with Inception-1D. Overall, the proposed framework successfully captures distinctive gas sensor-response patterns associated with TB presence and severity and may support clinical decision-making in resource-limited healthcare settings.