In vitro fertilization (IVF) involves creating different embryos from eggs and sperm. Embryologists then select the embryo almost definitely to lead to a successful pregnancy and transfer it to the patient.
Embryologists make this decision by utilizing their expertise to use a set of generally accepted principles based on the looks of the embryo. In recent years there was great interest within the use various artificial intelligence (AI) techniques on this process.
We developed such an AI system and tested it in a study with greater than 1,000 IVF patients. Our system chosen the identical embryo as a human expert in about two-thirds of cases and had only a rather lower overall success rate. The results are published in Nature Medicine.
Can deep learning help with IVF?
Over the past few years, we’ve been working with colleagues in Sweden to develop software to discover which embryos have the perfect likelihood of success in IVF. Our system uses deep learning, an AI method for identifying patterns in large amounts of knowledge.
While we were developing our system, we conducted retrospective studies comparing the system's decisions to previous real-world decisions made by embryologists. These initial results suggested that the deep learning system might even have the opportunity to do a greater job than a human expert. So the following step was to thoroughly test the system in a randomized trial.
Our study involved 1,066 patients at 14 fertility clinics in Australia and Europe (Denmark, Sweden and the UK). For each patient, each the deep learning system and a human expert chosen an embryo to be implanted after which randomly decided which of the 2 to make use of.
This study is the primary randomized controlled trial ever conducted using a deep learning system in embryo selection. Deep learning could many medical applicationsbut that is certainly one of the few prospective randomized trials of this technology in any healthcare setting.
Our findings
What we present in the study was that there was virtually no difference between the 2 approaches. The clinical pregnancy rate (the probability of seeing a fetal heart after transferring the primary embryo) was 46.5% when the deep learning system chosen the embryo and 48.2% when the embryologist chosen the embryo.
In other words, there was hardly any difference. In fact, the deep learning system selected the identical embryo because the embryologist 65.8% of the time. However, we also found that the synthetic intelligence system accomplished the embryo selection task ten times faster than the embryologist.
One goal of our study was to prove the “non-inferiority” of our deep learning system. This is common in medical research, as we at all times wish to be certain that that a proposed latest technique doesn’t result in worse results than the present standard.
Despite the proven fact that the deep learning system produced very similar results to human experts, our study couldn’t quite overcome the hurdle of proving “non-inferiority.”
In fact, the general success rates of the study were much higher than expected. This modified the statistical situation and meant that we’d have needed a much larger study – with almost 8,000 patients – to prove that the brand new method was non-inferior.
No significant differences
A variety of ethical concerns have been raised before in regards to the use of deep learning in embryo selection. One of those concerns is a possible change within the sex ratio – that’s, the final result with more male or female embryos – attributable to biased selection by the deep learning model.
However, we didn’t detect any change within the sex ratio in consequence of embryo selection using deep learning.
From our study, we concluded that there isn’t any significant difference in pregnancy rate whether the embryo is chosen by a deep learning system or by an experienced embryologist.
This suggests that using a deep learning tool for embryo selection is not going to transform the final result for a patient undergoing IVF (since it’s going to mostly select the identical embryo), but using a reliable automated tool of this type could make embryology labs more efficient and consistent.
Another conclusion from this study is that randomized trials, which take years to conduct, might not be the optimal approach for studying rapidly evolving technologies reminiscent of this one. Our future work evaluating this technology must explore alternative, yet clinically valid, approaches to this topic.
_The writer would love to thank Peter Illingworth for his work and the research described therein._
