The study investigated exhaled volatile organic compounds (VOCs) to distinguish lung cancer (180 patients), pneumonia (228 patients) and healthy controls (180 subjects) using micro-GC-MSD and machine learning. Lung cancer and pneumonia often look similar on X-ray images, leading to diagnostic errors, while current methods are invasive and expensive. Patients with lung cancer had lower levels of heptane, propane, 1-(methylthio)-styrene and higher levels of 2-hexanone, 6-hydroxy-o-xylene compared to healthy people. In pneumonia, 1,4-pentadiene, toluene, butyl acetate, p-xylene, D-limonene and isobutylnonylcarbonate were increased, heptane 2,2,4,6,6-pentamethyl-. Between lung cancer and pneumonia, seven VOCs had lower concentrations in cancer. Machine learning models achieved AUCs of 0.980 (cancer vs. health), 0.956 (inflammation vs. health), and 0.983 (cancer vs. inflammation). The results provide preliminary evidence for non-invasive diagnostics, but require further validations.