Enhancing the success of IVF with artificial intelligence

For decades, in-vitro fertilization (IVF) has offered those unable to conceive naturally the chance to have biological children; however, global success rates are notoriously low. Artificial intelligence (AI) is being used to improve the success rates of IVF, but it is not without its challenges. Does AI have the potential to be a clinically useful tool to increase the success of IVF?

Additional tools to embryonic screening technologies such as preimplantation genetic testing for aneuploid status (PGT-A) are highly sought, as they can be costly and invasive, and manual embryonic morphological assessment is subject to operator bias. In this issue of The Lancet Digital Health, Josue Barnes and colleagues describe how their non-invasive AI-driven method, STORK-A, can be used to predict ploidy status of an embryo, with the aim to select healthy blastocysts for implantation. The authors emphasise their approach is not designed to replace more traditional methods like PGT-A, but instead recommend it as a supplemental tool to aid clinicians. As such, this proof-of-concept approach in one of the largest studies of its kind, along with external validation, showcases a potential integrative role for AI in clinical decision making in the context of embryo selection ahead of implantation.

This study is one of several published since 2020 that show the potential of AI tools for embryo selection; for example, Sonya Diakiw and colleagues showed their convolutional neural network (CNN) tool resulted in a 12% reduction in time-to-pregnancy, and Charles Bormann and colleagues' CNN outperformed 15 embryologists in predicting embryos with the greatest implantation potential. However, there are numerous barriers that currently prevent clinical usage of novel AI tools in this field. The AI-based approaches used in some studies suffer from the black box effect, due to their complexity and in some cases, the use of proprietary algorithms. This is a common problem for AI tools used in medicine, but it can be alleviated by choosing interpretable, transparent models, which can aid in reducing hidden biases.

Another criticism of the use of digital tools in IVF is the scarcity of concrete medical evidence for their efficacy, meaning technologies approved by regulatory bodies (such as the Human Fertilisation and Embryology Authority in the UK) are scarce. It is imperative, therefore, that AI tools for embryo selection are rigorously tested in high-quality randomised controlled trials to evaluate their efficacy before regulatory approval and clinical implementation. There are currently no published RCTs that evaluate AI-driven tools for use in IVF, possibly due to the scarcity of robust externally validated approaches. As such, this external validation must be a priority ahead of conducting RCTs. In addition, follow-up studies assessing the health of children conceived with AI-assisted IVF are crucial in understanding the long-term health outcomes that might be linked to the use of these methods, given assisted reproductive technology might be associated with adverse childhood and adolescent outcomes such as cardiovascular dysfunction.

As with all scientific investigations into the use of novel approaches, it is also important to enrol participants from diverse backgrounds in any potential trials into AI-assisted embryo selection. A 2022 systematic review and meta-analysis suggested that following IVF, Black women had more spontaneous abortions and fewer live births than White women. Hajirasouliha and colleagues' tool is independently validated in two populations (one in the US and one in Spain), which gives an indication of generalisability across these cohorts; however, race or ethnicity data for participants in their study was not available.

It is promising to see a growing number of AI methods in development for selection of viable embryos, but the field faces ongoing challenges before such tools can be routinely used in clinical decision making. To bridge this gap, large and comprehensive RCTs need to be carried out to confirm the clinical value of externally-validated AI approaches ahead of regulatory approval for use. Most importantly, diverse cohorts must be recruited for future studies to assess generalizability of models and ensure equitable benefit. This will represent a substantial step forward towards clinical implementation of these tools, and increase the chances of conception for hopeful parents of all races and ethnicities.