PR002211 (Project)

Description:The majority of lung cancer (LC) patients are diagnosed at advanced stages due to ineffective screening methods. Therefore, there is a pressing need for a non-invasive LC screening and early detection method suitable for large-scale clinical application. In this project, we collected 1043 saliva samples from 334 LC patients and 709 non-LC volunteers across six hospitals, obtaining their metabolomics data through mass spectrometry Lab-on-a-Chip (MS LOC). This technique ensures rapid and high-throughput acquisition of stable salivary metabolic fingerprints (96 samples per batch). Using machine learning-based feature selection, we identified 35 metabolic markers associated with LC, highlighting metabolic disruptions in the saliva of LC patients. We then developed a classification model, SalivaMLD, leveraging an ensemble voting strategy incorporating multiple machine learning algorithms. This approach improved classification accuracy and robustness by aggregating predictions from various models. In the validation set, SalivaMLD achieved an area under the curve (AUC) of 0.850, with a sensitivity of 83.33% and specificity of 74.39%. Similarly, in the test set, it attained an AUC of 0.849, with sensitivity and specificity of 81.69% and 74.23%, respectively, outperforming traditional tumor markers like carcinoembryonic antigen (CEA) and carbohydrate antigen 125 (CA125). Notably, SalivaMLD accurately identified early-stage LC with an accuracy range of 77.42%-81.97% and effectively differentiated LC across various pathologies in both validation and test sets. Thus, this method combining machine learning and MS LOC-based salivary metabolic fingerprints offers a promising tool for rapid, non-invasive LC screening in clinical settings.