Artificial intelligence is being introduced into many aspects of our lives. From speech and handwriting recognition, to gaming and healthcare. It has been demonstrated that Artificial Intelligence, specifically ‘Machine Learning’, can outperform humans in many tasks and in decision making.

Humans make decisions based on their experience and often very simple rules. In IVF, scientists and medics make many decisions. One of the most critical ones, to treatment success, is which embryo to choose.

Care Fertility scientists have undertaken extensive research for over a decade to improve the accuracy of embryo selection. This resulted in models, or algorithms, called Caremaps® to select embryos with the highest chance of becoming a baby. Care’s work in this area has been widely published. Our latest development uses Machine Learning to analyse embryo development from fertilisation right up to the point of embryo transfer to the patient.

Because we have a large group of clinics performing the same high-tech embryology, we collated almost half a billion images, from 63,000 embryos, from our time lapse incubators and trained a machine learning model to assess them accurately and automatically.

Embryo assessment using Machine Learning, when trained on large high-quality data, is more accurate and consistent than the assessments made by humans. Machine Learning assessment is consistent. Humans change their minds. It is estimated that an embryologist with 10 years’ training and experience may have experience from assessing approximately 10,000 embryos over a decade.

Caremaps-AI was trained on 63,000 embryos and almost half a billion images of them.

Accuracy was assessed using previously unseen embryos by comparing the results of the Machine Learning model with human assessment of embryo development. This research was recently selected and presented at the European Society for Human Reproduction (ESHRE) conference by Care’s Chief Scientific Officer.

The Machine Learning model (Caremaps-AI) automatically and rapidly generated assessments of the embryos. Efficacy and comparability of the model to automate reliable, utilisable information was demonstrated by comparison with manual assessment data and the model’s ability to generate information which could be used to predict birth. Birth-predictive capability was measured, and benchmarked against manual methods and known outcomes.