ANNUAL MEETING 2022

This conference debate on artificial intelligence in the lab did reach a consensus – that, although there are benefits in the application of AI in ART, there is as yet insufficient evidence to support its widespread use.

First to speak was Dean Morbeck, chief scientific officer for Kindbody, US, who made the case for machine learning and AI-guided decision-making on the basis that it enhances embryology and fertility treatment. He outlined key benefits, such as improved and standardised embryo grading, more efficient and accurate data entry, a better ‘patient experience’, greater efficiency in the laboratory, and automated quality control.

The human brain, he said, can’t possibly deal with everything that’s involved in embryo development. Even assessing chance of implantation, embryologists must deal with countless variables and ‘there’s no way the brain can do this’.

Ranking embryos correctly is the most important decision an embryologist will make, said Morbeck, but humans lack the ability to be objective on grading and have ‘inherent subjectivity’; this is also a problem which affects decisions on whether to freeze blastocysts or biopsy them. Staff may be overworked, making decisions alone because it’s a weekend, or just simply tired.

This then is an area where AI can help the chance of live birth, argued Morbeck. Clinics fail more than they succeed in IVF, clearly shown by SART data in which 14% of cycles produce no embryos and 62% no live birth.

Patients who have unsuccessful treatment are going away with ‘nothing to show’. To make the situation better, Morbeck explained that machine learning tools could standardise the ‘minimal usable threshold’ applied to reduce the rejection of viable but low-grade embryos. AI could learn how embryos develop and then use that information to select the best embryo.

The case against AI was made by Peter Tennant, associate professor of health data science at the University of Leeds, UK. His argument was that much of the hype around what AI can do is based on ‘overpromise’ and the technology has a long way to go to prove its worth in embryology and ART. By way of illustration, he said Google’s medical AI – such as its algorithm to detect diabetic retinopathy – has proved accurate in the lab but not in real life.

In embryology specifically, Tennant outlined the promise and problems of AI. Among the benefits are enhancement and automation (eg, follicle counting), prediction and classification (eg, live birth), smarter and more personalised care, and targeted treatment regimens. However, the problems are significant - machine learning is developed in small and specialist samples and contexts, performance is variable, and external validation is lacking. AI-based tools also rarely focus on a true outcome of interest, said Tennant. The few studies that have focused on live birth outcomes have found that AI performs poorly.

Although it seems a good prediction and forecasting tool, AI has significant limits, which include a lack of common understanding: algorithms merely see patterns in data so can be fooled by a doctor’s purple pen mark on a cancer biopsy slide. Tennant said this could occur in AI analysis of blastocyst images – an algorithm may ‘see patterns we don’t want it to pick up’.

In response to a question from the audience, Tennant acknowledged that humans make mistakes too but he emphasised that a sceptical approach is needed around AI until the technology is fully understood, and ended the session by saying it’s not a case of being afraid of intelligent machines, but of stupid ones.