Machine learning-based cervical secretions DNA methylation profiling of implantation window genes enhances live birth prediction in frozen embryo transfer

Identifying non-invasive biomarkers of endometrial receptivity may improve outcome prediction in frozen–thawed embryo transfer (FET) cycles. In this prospective case–control study, cervical secretions were collected during the late proliferative phase prior to progesterone administration from 141 women undergoing FET, followed by DNA methylation profiling and machine-learning modeling to predict clinical pregnancy, ongoing pregnancy, and live birth. Late proliferative–phase methylation signatures showed strong concordance with mid-secretory–phase signatures (R 2 = 0.91–0.99), providing preliminary support for temporal concordance relevant to endometrial receptivity. Feature selection using the Boruta algorithm identified a six-gene panel— ANK3, HIVEP2, IL15, SERPINE1, SERPINE2, and TAGLN2 —with predictive value for pregnancy outcomes. In the test set, the machine learning–based DNA methylation model achieved AUCs of 0.903 for clinical pregnancy, 0.853 for ongoing pregnancy, and 0.833 for live birth. For live birth prediction, the random forest model achieved a positive predictive value of 76.5% and a negative predictive value of 88.9% in the test set, supporting the potential utility of this approach for non-invasive pre-transfer risk stratification. These findings suggest that late proliferative–phase cervical secretion methylation profiling provides a cycle-synchronized, non-invasive approach for assessing endometrial receptivity and predicting pregnancy outcomes in FET cycles; further external and prospective interventional validation is warranted.