Development and validation of DNA methylation-based prognostic models for cervical cancer via multiple machine learning algorithms

Cervical cancer is a leading cause of morbidity and mortality among reproductive cancers in women. Current treatment options for patients with persistent, recurrent, or metastatic disease often fall short of achieving curative outcomes. This study aimed to develop an exploratory DNA methylation-based model for survival-status risk stratification in human papillomavirus (HPV)-positive individuals. We obtained shared DNA methylation sites via differential methylation probe (DMP) analysis, log-rank analysis, and Cox analysis, followed by LASSO regression for signature feature selection. We applied nine machine learning algorithms, including logistic regression, linear discriminant analysis, naive Bayes, K nearest neighbors, decision tree, random forest, support vector machine, GLM with elastic net regularization and eXtreme gradient boosting, to construct a prognostic model for cervical cancer via the TCGA-CESC cohort and validated the model via the CGCI-HTMCP-CC cohort. Model performance was evaluated via receiver operating characteristic (ROC) curve analysis. Our results demonstrated the successful construction of a prognostic model for cervical cancer via 15 gene-associated DNA methylation sites (MUL1, HDAC2, HSPA12A, MAP3K14, RPH3A, SH3TC1, USP2, CD7, CD22, ARNT2, OR52W1, AP3B2, WWTR1, CEP120, and ECHDC2) and 3 uncharacterized methylation sites (cg23217512, cg12162450, and cg05337137). Three learners, K nearest neighbors, support vector machine, and GLM with elastic net regularization (glmnet), exhibited the best performance in the risk stratification of cervical cancer. We also developed a small prognostic model based on glmnet, consisting of ARNT2, USP2, HDAC2, HSPA12A, and cg23217512, with a similar performance to that of the 18-methylation site model. Our machine learning models used a panel of DNA methylation sites to stratify survival outcome in HPV-positive individuals, providing preliminary evidence for survival-status risk stratification that requires further validation in larger cohorts.