Abstract LB-267: Detection of lung cancer by metabolomics of exhaled breath and machine learning

Abstract

Listen

Translate article

Abstract Background: Exhaled breath-based test is an attractive option for cancer detection due to it is non-invasive and feasible nature. Here, we reported an effective approach to detect early-stage lung cancer (LC) by integration of metabolomics of exhaled breath and machine learning. Methods: Healthy controls (HC) were adults who received low-dose CT for annual physical examination with no calcified pulmonary nodules and without history of cancer within 5 years. Treatment-naïve LC patients without history of cancer were enrolled. Exhaled breath samples of LC patients were collected before surgery. Exhaled breath samples were collected and stored in Tedla bags and detected by a high-pressure photon ionization time-of-flight mass spectrometry (HPPI-TOFMS). A CO2 sensor was applied during exhaled breath collection to ensure only “alveolar air” was collected. Detection model construction and feature selection were performed by support vector machine (SVM). Results: Firstly, an exploratory cohort of 119 HC and 118 LC was set up. Exhaled breath samples were directly detected by HPPI-TOFMS and 32500 features were extracted from each sample. 22 features were selected and subjected to SVM to construct detection model in the exploratory cohort, which was randomly divided into training (n=159) and test (n=78) set. This model achieved a sensitivity of 89.74%, specificity of 92.31%, and accuracy of 89.74% in the test set, suggesting exhaled breath is accurate for LC detection. Secondly, a quantitative detection method was established by introducing mixed internal standard gas at 15 mL/min controlled by a mass flow meter, by which individual violate organic compound could be quantified. Then, a prospective cohort of 53 LC and 129 HC was set up and detected by the quantitative method. For mass spectrometry data with inner standard gas, 21 features were selected by SVM. Similarly, the prospective cohort was randomly divided into training (n=122) and test (n=60) set. The 21 features-based quantified model was trained and tested in the prospective cohort and achieved a sensitivity of 88.24%, specificity of 100%, and accuracy of 96.67% in the test dataset. Conclusions: Exhaled breath with machine learning is a highly accurate approach for early-stage LC detection and inner standard gas could improve detection performance. Citation Format: Mantang Qiu, Hang Li, Shushi Meng, Qingyun Li, Zuli Zhou, Jun Wang. Detection of lung cancer by metabolomics of exhaled breath and machine learning [abstract]. In: Proceedings of the Annual Meeting of the American Association for Cancer Research 2020; 2020 Apr 27-28 and Jun 22-24. Philadelphia (PA): AACR; Cancer Res 2020;80(16 Suppl):Abstract nr LB-267.