Predicting Blastocyst Viability

Abstract

Objective: To develop and evaluate a machine learning approach for predicting blastocyst viability using automated expansion measurements, clinical variables and image features.
Methods: A convolutional neural network was developed to automatically segment and measure blastocyst cross-sectional area from time-lapse images. We generated expansion curves and extracted features for 315 blastocysts. Various machine learning models were trained to predict biochemical and ongoing pregnancy using expansion, clinical and image-derived features. Model performance was evaluated using cross-validation and an unseen test set.
Results: The segmentation model achieved a Jaccard index of 97.6% on the validation set. Support vector machines using clinical and expansion features achieved the highest performance, with AUCs of 0.71 and 0.70 for predicting biochemical and ongoing pregnancy, respectively, on the test set. Blastocysts resulting in pregnancy expanded significantly faster and reached larger final cross-sectional areas compared to those that did not implant. Key predictive features included expansion rate and maternal age.
Conclusions: Automated quantification of blastocyst expansion dynamics combined with clinical variables enables prediction of implantation potential. Incorporating objective expansion metrics into embryo selection may enhance IVF success rates beyond traditional morphological grading systems.