Development and validation of a screening model for lung cancer using machine learning: A large-scale, multi-center study of biomarkers in breath

How well does your e-nose detect cancer? Application of artificial breath analysis for performance assessment

Comparing electronic nose (e-nose) performance is a challenging task because of a lack of standardised method. This paper proposes a method for defining and quantifying an indicator of the effectiveness of multi-sensor systems in detecting cancers by artificial breath analysis. To build this method, an evaluation of the performances of an array of metal oxide sensors built for use as a lung cancer screening tool was conducted. Breath from 20 healthy volunteers has been sampled in fluorinated ethylene propylene sampling bags. These healthy samples were analysed with and without the addition of nine volatile organic compound (VOC) cancer biomarkers, chosen from literature. The concentration of the VOC added was done in increasing amounts. The more VOC were added, the better the discrimination between 'healthy' samples (breath without additives) and 'cancer' samples (breath with additives) was. By determining at which level of concentration the e-nose fails to reliably discriminate between the two groups, we estimate its ability to well predict the presence of the disease or not in a realistic situation. In this work, a home-made e-nose is put to the test. The results underline that the biomarkers need to be about 5.3 times higher in concentration than in real breath for the home-made nose to tell the difference between groups with a sufficient confidence.

A combined screening study for evaluating the potential of exhaled acetone, isoprene, and nitric oxide as biomarkers of lung cancer

Background: the early lung cancer (LC) screening strategy significantly reduces LC mortality. According to previous studies, lung cancer can be effectively diagnosed by analyzing the concentration of volatile organic compounds (VOCs) in human exhaled breath and establishing a diagnosis model based on the different VOCs. This method, called breath analysis, has the advantage of being rapid and non-invasive. To develop a non-invasive, portable breath detection instrument based on cavity ring-down spectroscopy (CRDS), we explored the feasibility of establishing a model with acetone, isoprene, and nitric oxide (NO) exhaled through human breath, which can be detected on the CRDS instrument. Methods: a total of 511 participants were recruited from the Cancer Institute and Hospital, Tianjin Medical University as the discovery set and randomly split (2 : 1) into training set and internal validation set with stratification. For external validation, 51 participants were recruited from the General Hospital, Tianjin Medical University. Acetone and isoprene from exhaled breath were detected by proton-transfer-reaction time-of-flight mass spectrometry (PTR-TOF-MS), and NO was measured using CRDS. The model was constructed using the ensemble learning method that set eXtreme gradient boosting and logistic regression as the basis model and logistic regression as the senior model. The model was evaluated based on accuracy, sensitivity, and specificity. Results: the model achieved an accuracy of 78.8%, sensitivity of 81.0%, specificity of 70.0%, and area under the receiver operating curve (ROC, AUC) of 0.8341 (95% CI from 0.8055 to 0.8852) in the internal validation set. Furthermore, it attained an accuracy of 66.7%, sensitivity of 68.2%, specificity of 65.5%, and AUC of 0.6834 (95% CI from 0.5259 to 0.7956) in the external validation set. Conclusion: the model, established with acetone, isoprene, and NO as predictors, possesses the ability to identify LC patients from healthy control (HC) participants. The CRDS instrument, which simultaneously detects acetone, isoprene, and NO, is expected to be a non-invasive, rapid, portable, and accurate device for early screening of LC.