Salivary Glycoprofiling via a Machine Learning-Augmented Lectin Microarray for Noninvasive Risk Stratification of Lung Cancer

Abstract Although aberrant glycosylation is closely associated with lung cancer (LC), salivary glycopatterns and their discriminatory value across healthy individuals, benign pulmonary disease, and LC remain insufficiently characterized. We profiled saliva from 307 participants using a lectin microarray, comprising a development cohort of 227 participants and an independent test cohort of 80 participants. Five task-specific LASSO-logistic regression models were constructed to distinguish pulmonary disease, LC, and LC subtypes. A separate Nom-LC model combined consensus lectin features selected by LASSO and support vector machine recursive feature elimination with clinical variables. Salivary glycopatterns differed among healthy volunteers, benign pulmonary disease, and LC groups, and the five task-specific models showed discriminatory performance across pulmonary disease, LC, and exploratory subtype classification tasks. The resulting Nom-LC model incorporated smoking history and four lectin signals (HHL, PNA, RCA120, and PWM), yielding AUCs of 0.908 and 0.907 in the training and internal validation sets. In the independent test cohort, the AUC was 0.903 (95% CI, 0.831–0.976); at the prespecified cutoff, sensitivity, specificity, and accuracy were 86.36%, 94.44%, and 90.00%, respectively. These findings support further evaluation of salivary glycopatterns as an auxiliary tool for LC risk stratification, but prospective multicenter validation is required.