Fully automated AI-based system developed for IVF embryo quality assessment

Developed by researchers at Weill Cornell Medicine, ‘BELA’ is the team’s latest AI-based platform for assessing whether an embryo has a normal (euploid) or abnormal (aneuploid) number of chromosomes, which is a key determinant of IVF success.

The researchers said that unlike prior AI-based approaches, BELA does not need to  consider embryologists' subjective assessments of embryos. Consequently it offers an objective, generalisable measure and, if its utility is confirmed in clinical trials, could be used in embryology clinics to improve the efficiency of the IVF process.

Embryologists typically assess an IVF embryo's quality by examining it under a microscope. If it looks relatively normal but there are reasons to suspect possible problems, such as in cases of advanced maternal age, they may test its chromosomal status more directly.

According to the researchers, the ‘gold standard’ test is a somewhat risky biopsy-like procedure called preimplantation genetic testing for aneuploidy (PGT-A) – and, as a result, embryologists have been collaborating with AI experts to find ways to automate workflow and improve outcomes.

In a 2022 study, Dr. Hajirasouliha – associate professor of physiology and biophysics and a member of the Englander Institute for Precision Medicine at Weill Cornell Medicine – and colleagues developed an AI-based system called STORK-A, which uses a single microscopic image of an embryo, plus maternal age and embryologists’ scoring, to predict the embryo’s ploidy status (the number of sets of chromosomes in a cell or organism) with around 70 per cent accuracy.

Now, the researchers have developed BELA to generate accurate ploidy prediction independently of embryologists’ assessments. The heart of the system is a machine-learning model that analyses nine time-lapse video images of an embryo under a microscope in a key interval, around five days after fertilisation, to generate an embryo quality score. The system then uses this score and maternal age to predict euploidy or aneuploidy.

The researchers trained the model on a Weill Cornell Medicine CRM deidentified dataset with image sequences of nearly 2,000 embryos and their PGT-A-tested ploidy status. They then tested the model on new Weill Cornell Medicine CRM datasets and those from separate, large IVF clinics in Florida and Spain.

The study found that the model predicted ploidy status with ‘moderately higher accuracy’ than previous versions and worked well for the external and internal datasets.

Looking ahead, the researchers said that the next step is to test BELA’s predictive power prospectively in a randomised, controlled clinical trial, which they are currently planning.

“BELA and AI models like it could expand the availability of IVF to areas that don’t have access to high-end IVF technology and PGT testing, improving equity in IVF care across the world,” Dr Nikica Zaninovic, associate professor of embryology in clinical obstetrics and gynaecology at Weill Cornell Medicine said in a statement.

The research, supported in part by the National Institute of General Medical Sciences, was published in Nature Communications and can be read in full here.