Machine learning assisted breathomic approach for early-stage thoracic cancer detection.

Objective: This study explores the feasibility of using breathomic biomarkers analyzed by machine learning as a non-invasive diagnostic tool to differentiate between benign and malignant thoracic lesions, aiming to enhance early detection of thoracic cancers and inform clinical decision-making. Methods: This study enrolled 132 participants with confirmed diagnosis of lung cancer, esophageal cancer, thymoma, and benign diseases. Exhaled breath samples were analyzed by thermal desorption-gas chromatography-mass spectrometry. A logistic regression algorithm was employed to construct a classification model for benign and malignant thoracic lesions. This model was trained on a subset of 80 cases and subsequently validated in a separate set comprising 52 samples. Results: A logistic regression model based on thirteen exhaled volatile organic compounds (VOCs) was developed to differentiate benign and malignant thoracic lesions. The 13-VOC model achieved an AUC of 0.85 (0.72, 0.96), accuracy of 0.79 (0.66, 0.88), sensitivity of 0.82 (0.67, 0.91), and a specificity of 0.71 (0.45, 0.88). It correctly classified 80% of lung cancer, 80% of thymoma, and 100% of esophageal cancer cases, distinguishing 71.4% of benign lesions. For lung cancer, the model achieved an AUC of 0.79 (0.57, 0.98), sensitivity of 0.80 (0.63, 0.91), and specificity of 0.63 (0.31, 0.86), with 81.8% accuracy in detecting early-stage (Stage 0 + I + II) disease. The model outperformed a 4-serum tumor marker panel in sensitivity (0.90 vs. 0.39, Conclusion: This study demonstrates the feasibility of utilizing breathomics biomarkers for developing a non-invasive machine learning model for the early diagnosis of thoracic malignancies. These findings provide a foundation for breath analysis as a promising tool for early cancer detection, potentially facilitating improved clinical decision-making and enhancing patient outcomes.