Background and Aims

Assisted hatching (AH) aims to enhance embryo implantation, but its impact remains uncertain. This study assessed whether artificial intelligence (AI) models can differentiate between natural hatching (NH) and assisted hatching (AH) embryos using static blastocyst images and maternal age, without AH labels, and examined their influence on pregnancy prediction.

Methods

The data consisted of 5,217 static images of day-5 blastocysts, retrospectively obtained from three IVF clinics in Korea and three in the United States. All embryos were captured using optical microscopy prior to cryopreservation, and AH status was manually labeled. Fetal heart tone (FHT) outcomes were used to define pregnancy success. A convolutional neural network (CNN) model was trained to predict pregnancy using embryo images and maternal age. The dataset was split 60:20:20 for training, validation, and testing, and performance was evaluated by accuracy and AUC, comparing NH and AH embryo

Results

The model achieved an accuracy of 0.6781 ± 0.1043 and an AUROC of 0.6569 ± 0.1153. Natural hatching (NH) embryos showed higher pregnancy rates and AI-predicted scores than assisted hatching (AH) embryos. Including AH status as an input resulted in similar performance (accuracy: 0.7079 ± 0.0762; AUROC: 0.6342 ± 0.1327), indicating that the model could recognize differences without explicit AH labels.

Conclusions

This study demonstrates that AI models can distinguish natural from assisted hatching embryos using static images and maternal age, without assisted hatching labels. The model assigned higher pregnancy prediction scores to natural hatching embryos, which were consistent with clinical outcomes. These findings highlight the potential for prediction models independent of hatching status, reducing unnecessary procedures.