Clinical Diagnosis of Lung Cancer via Exhaled Breath Analysis Using an Ultrasensitive and Cross-Selective Benzene-Derivative Gas Sensor Assisted by Machine Learning

Lung cancer is often diagnosed at an advanced stage due to its subtle and imperceptible early symptoms. Therefore, the development of a convenient and reliable method or device for early screening and diagnosis of lung cancer is urgently needed. Benzene derivatives and alkanes have been identified as key breath biomarkers for lung cancer. In this study, we fabricated an ultrasensitive and cross-selective gas sensor based on Pd/PdO co-doped SnO2, specifically designed to detect benzene derivatives. The sensor demonstrates detection limits at the ppb level for several key lung cancer breath biomarkers, including toluene (40 ppb), 1-methyl-4-(1-methylethyl)-benzene (100 ppb), o-xylene (90 ppb), styrene (500 ppb), ethylbenzene (100 ppb), 2-methylhexane (500 ppb), and ethyl alcohol (150 ppb). Clinically, exhaled breath samples from 50 lung cancer patients and 60 healthy control subjects were analyzed using the sensor. Two machine learning approaches were employed to distinguish between the two groups: (1) manual feature extraction, followed by principal component analysis (PCA), and (2) a deep learning framework integrating convolutional neural networks with a multilayer perceptron. Diagnostic models based on these approaches achieved overall accuracies, sensitivities, and specificities of 0.95, 1.00, and 0.89 (PCA-based) and 0.86, 0.91, and 0.83 (deep learning-based), respectively. Receiver operating characteristic curve analyses yielded area under the curve values of 0.98 and 0.94 for the PCA and deep learning models, respectively. These findings suggest that the sensor has a significant potential for clinical lung cancer diagnosis.