Artificial Intelligence-Assisted Embryo Selection in In Vitro Fertilisation: A Systematic Review of Model Performance and Clinical Evidence
This systematic review concludes that while AI-assisted embryo selection in IVF has advanced significantly in model development and technical performance, current evidence of clinical effectiveness remains limited and inconsistent, necessitating further independent validation and prospective studies before its routine clinical adoption can be established.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
For millions of couples around the world, the path to starting a family is blocked by infertility, a condition that affects roughly one in six people of reproductive age. When natural conception does not happen, doctors often turn to a laboratory procedure called in vitro fertilisation, where an egg and sperm are combined outside the body to create an embryo. The most critical and difficult step in this process is choosing which single embryo to place back into the woman's uterus. Traditionally, highly trained scientists called embryologists have made this choice by looking at the embryo under a microscope, judging its health based on how it looks and how fast it grows. However, this visual assessment can vary from one expert to another, and even the best human eye cannot always predict which embryo will successfully take root and grow into a baby. In recent years, scientists have begun testing whether artificial intelligence, the same kind of computer technology that can recognise faces or drive cars, can help make this selection more accurate.
A new systematic review, a type of study that gathers and analyses all the existing research on a specific topic, has now taken a hard look at how well these computer systems are actually performing in the real world. The researchers behind this review, Emmanuel Emmanuel and Ekwueme Precious, did not run a single new experiment. Instead, they collected and examined dozens of previous studies that tested artificial intelligence tools designed to rank and select embryos. Their goal was to separate the technical success of these computer models from their actual ability to improve the chances of a healthy pregnancy. They wanted to know if the software, which can process thousands of images and find patterns invisible to humans, truly leads to more babies being born, or if its success is limited to the computer screen.
The review found that the technology has advanced rapidly. Early attempts used simple computer programs, but the most recent studies employ deep learning, a sophisticated method where the computer learns by analysing vast libraries of embryo images, including time-lapse videos that show the embryo growing hour by hour. These systems can now score embryos, predict their genetic health, and rank them in a way that often matches or even exceeds the consistency of human experts. In terms of pure prediction, the computer models have shown they can distinguish between embryos that are likely to implant and those that are not. For instance, one large study involving nearly two hundred thousand embryos from twenty-two different clinics found that the AI system could correctly identify viable embryos with a level of accuracy that ranged from sixty percent to seventy-five percent, depending on the specific testing conditions.
However, the review highlights a significant gap between what the computer can predict and what it can achieve in a clinic. Just because a model can accurately rank embryos in a database does not mean that using it will result in more successful pregnancies for patients. The most rigorous test of this technology came from a major randomised trial involving over one thousand patients across fourteen clinics. In this study, one group of patients had their embryos selected by a deep learning system, while another group had embryos selected by the traditional method of human visual assessment. The results were clear: the group using the artificial intelligence did not perform better than the group using the human experts. In fact, the clinical pregnancy rate was slightly lower for the AI group at forty-six point five percent, compared to forty-eight point two percent for the conventional group. The study was designed to see if the AI was at least as good as the human method, and it failed to prove that point.
This finding suggests that while artificial intelligence is a powerful tool for analysing data, it has not yet proven itself as a superior replacement for human judgment in the clinic. The review points out that many earlier studies showing great promise were retrospective, meaning they looked back at data from embryos that had already been transferred and studied. These studies often showed strong links between high AI scores and successful pregnancies, but they could not prove that the AI caused the success, as the embryos chosen by the AI were not compared against a control group in a fair, randomised way. Furthermore, the review noted that different clinics use different microscopes, different incubators, and different ways of caring for embryos, which means a computer model trained in one laboratory might not work as well in another. The technology also showed signs of instability, where the same embryo could receive a different rank if the computer model was run multiple times, raising questions about its reliability.
The authors conclude that the field is at a turning point. The computer models have demonstrated that they can learn the visual language of embryo development and provide consistent rankings, which is a remarkable technical achievement. Yet, the evidence that this technology improves the ultimate goal of in vitro fertilisation—having a live birth—is still weak. The review does not dismiss the potential of artificial intelligence, but it insists that the technology must be treated as a decision-support tool rather than an automatic solution. Before clinics can confidently rely on these systems to choose embryos, more independent testing is needed. Future studies must focus on real-world outcomes, specifically the number of babies born, and must ensure that the software works reliably across different hospitals and patient groups. Until such proof is available, the human embryologist remains the essential guide in the journey to parenthood, with artificial intelligence serving as a helpful, but not yet fully trusted, assistant.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.