AI Tools for Fertility Treatment Decision Making

Apricity is proud to have presented eight exciting posters at the annual #Fertility2023 conference which took place in Belfast last week. The topics of the posters presented this year focused on creating artificial intelligence (AI) tools that could help embryologists in their critical decision-making. These tools ensure that the patients receive the best possible care based on reliable data and improve their chances of fertility treatment success.

The Apricity Research and AI team have been presenting on the following:

• Survey of the fertility professionals’ views on AI in clinical practice

• A novel follicle classification system based on the shape of the distribution of follicle sizes

• Embryo development in 4D

• Factors to consider when deciding how many embryos to transfer

• Secure data access solution for improved research

• Blastocyst component image recognition using AI

• Inner cell mass shape classification system and aneuploidy

• Inner cell mass detection: AI vs humans

Survey of the fertility professionals’ views on AI in clinical practice

This survey, conducted by an Imperial College London Master’s student Sabahat Bajwa, focused on understanding how artificial intelligence (AI) tools are viewed by fertility professionals in clinical practice. 144 respondents including embryologists, doctors and academics from 37 countries participated in the survey.

It revealed that overall fertility professionals viewed AI positively but remained cautious about its use in fertility as a decision-making support tool. However, 53% of participants agreed that AI could help reduce the burden felt, especially when making clinical decisions. 14 out of 29 respondents that disagreed that AI reduced clinical burden had never experienced the use of AI tools in the first place.

AI could help to provide a full understanding of decision-making, which currently is a complex and non-standardised process in both the embryology lab and doctor’s consultation room. The low implementation of AI in clinical fertility environment to date can be explained by insufficient experience, knowledge and validation of AI tools. Therefore, there is a need for training programmes, smooth integration into clinics’ digital ecosystems and further testing of the AI tools, such as increased live birth validation. 

A novel follicle classification system based on the shape of the distribution of follicle sizes

Another Apricity research project led by Imperial College London Master’s student Sabahat Bajwa focuses on ovarian stimulation. Ovarian stimulation is a process during the fertility treatment cycle when a female patient’s ovaries are artificially stimulated to produce more eggs for subsequent fertilisation, instead of only one egg that is ovulated during a regular cycle. 

There are several options for ovarian stimulation used during fertility treatment which differ by certain medicines prescribed by doctors to take over several days/weeks 1. However, there is no one treatment protocol that works for each woman undergoing fertility treatment. Some patients may develop only very few eggs (and the stimulation will have to be repeated), while others may overreact with a risk of a serious complication called ovarian hyperstimulation syndrome (OHSS). 1 

Apricity has been working on AI tools to help clinicians choose the most suitable ovarian stimulation protocol 2. This project focused on classifying the growth of follicles in a transparent manner so that the clinician can understand the decision being suggested by the AI. 

Apricity proposed a novel follicle profile classification system based on the shape of the distribution of follicle sizes seen and compared the follicle distributions with patient factors, including AMH, BMI, age and number of eggs collected. This helped to identify differences in variable averages between each follicle distribution, which can be a simple method of categorising follicle tracking data and could be assessed automatically using artificial intelligence.

Embryo development in 4D

The way cells in an embryo are arranged geometrically has been shown to have an impact on further embryo development. Previous studies by Apricity researched the 3D visualisation of an embryo. This time, using a deep learning AI system, Chloe He, a PhD student at UCL and Data Scientist at Apricity, investigated how these cell arrangements impacted embryo development over time, otherwise known as 4D. 

4D embryo visualisation allows visualising the changes in cell arrangements throughout embryo development. This not only provides special insights into embryo development but can also improve the understanding of embryo quality, which is important when deciding which embryo to choose for transfer. 

Factors to consider when deciding how many embryos to transfer

Single embryo transfer is generally recommended for good prognosis patients (women under 37 on their first IVF cycle) regardless of available embryo quality 3. The alternative, a multiple embryo transfer, may lead to serious maternal and foetal complications, therefore HFEA has introduced the 10% multiple birth rate target, which means that individual fertility clinics must adjust their clinical practices to maintain their multiple live birth rate under 10%. 

A study conducted by Jia Lu and Apricity's research team investigated which factors should be considered when deciding on whether to use elective single embryo transfer (eSET) or multiple embryo transfer (MET). The factors studied included:

• Maternal age

• Number of eggs collected

• Day of embryo transfer

• Whether it was a fresh or frozen embryo transfer (FET)

• Whether PGT-A was used

The study was conducted based on data from five different datasets in four different countries. Different clinics and countries have different fertility treatment practices and guidelines, which is why it is important to reflect diverse contexts in research and control for them. 

The findings demonstrated that elective single embryo transfer was overall a more successful strategy in all age groups, as it:

• Increased success rates in patients under 40

• Led to equivalent success rates in patients over 40

• Reduced the risk of multiple live birth in all patients thus avoiding severe complications for mother and baby

Secure data access solution for improved research

If you ever wondered how Apricity uses its AI tools to innovate, improve patient care and maintain data security, you’re in the right place. AI can be an invaluable tool in our everyday lives, including healthcare and fertility. However, to develop these tools and make them as much useful and reliable as possible, one needs to access a lot of data so the AI can learn from it. However, obtaining data for research is a long and usually complicated process due to data protection and security laws.

To ensure that patient data is completely secure, Apricity data scientist Timothy Ferrand and the rest of Apricity AI team developed three machine learning algorithms using a novel data access solution. The novel platform enabled the data scientist team to send computer code to run securely on the collaborators’ centre’s on-premise server, thus, achieving a high level of data security as the data never left the centre at any point during the study.

The performance of each model was really encouraging. It demonstrated the promise of using a secure third-party data-sharing solution allowing external researchers to provide and use clinically-relevant insights on sensitive fertility data in a secure and trustworthy manner.

Blastocyst component image recognition using AI

Another important factor when developing AI tools is whether they will be able to work on a variety of data. The main problem when developing machine learning algorithms is that they very often don’t work when tested on different data than they were trained on. We want it to learn and help humans make important clinical decisions, which requires a broader understanding and perceptibility, usually unavailable for humans as easily. 

For this reason, a master’s student at UCL, Ivan Popov, led a research project focused on creating a machine learning algorithm that can work well on different data. He developed a reliable tool that’s able to recognize different parts of a blastocyst (a day 5 embryo) including:

• Zona pellucida

• Trophectoderm

• Inner cell mass

• Blastocoel

It was comparable to the existing state-of-the-art models and was capable of performing high-quality segmentation for unseen data from different equipment. The tool has the potential to aid embryo evaluation by embryologists which is a critical decision when selecting which embryo to transfer. 

Inner cell mass shape classification system and aneuploidy

By classifying the inner cell mass (ICM) of a blastocyst, an MSc student at Imperial College London, Elaina Lausic, led a research project on predicting embryo quality and outcome, including aneuploidy, or chromosomal imbalance in embryos, implantation and live birth. 

She found that, compared to other shapes, blastocysts with spherical ICM led to aneuploid embryos, which generally result in failure of implantation or miscarriage. However, ICM shape proved to not be a predictor of implantation or live birth.

These results are highly essential in the work towards non-invasive tools that can help predict embryo aneuploidy. Aneuploidy is the most common obstetrical complication that can explain single and recurrent pregnancy loss 4. Currently, the only way to test it is by performing pre-implantation genetic testing for aneuploidy (PGT-A), which is a highly invasive procedure for the embryo involving a biopsy. Because embryos are fragile, a non-invasive method of finding out whether it is chromosomally balanced (euploid) is much valued. 

Inner cell mass detection: AI vs humans

The main goal of using AI tools in embryology is to set a high-quality standardisation. A high degree of variability, both between and within, embryologists has been found. Since embryologists select embryos based on the grades of such evaluations, there should be no inconsistency, and AI would be really helpful to eliminate this human error. 

A project led by Dr Celine Jacques, a data analyst at Apricity, investigated how ICM grading compares between and within AI and embryologists based on different factors:

• The focus of the ICM image

• ICM position within an embryo

• ICM compaction level

• ICM cell number

She found that the variation between AI and humans was impacted by the factors in the same way as the variation between humans. The largest variation was seen when ICM was out of focus.

When training embryologists or AI models, all four factors should be considered but especially having ICM in focus. Using AI becomes essential, although it should be checked and corrected by experts if necessary.