What a 1,066-patient IVF trial found about AI embryo selection

An AI system can rank embryos quickly. Whether its choice helps an IVF patient achieve a pregnancy requires a different test. A trial published in August 2024 put iDAScore into that decision at 14 clinics in Australia, Denmark, Sweden and the UK, comparing it with conventional assessment by experienced embryologists.[3][1]

The underlying technology starts inside the incubator. Time-lapse imaging records embryos as they develop, allowing a laboratory to inspect a sequence of changes while leaving them in place. An embryologist can interpret those images, or software can turn them into scores used when choosing an embryo for transfer.[5]

Learning from the embryo’s development

The development study for iDAScore version 1 used 115,832 embryos from 18 clinics. Only 14,644 had been transferred with known implantation outcomes. Many others had been discarded by embryologists and were assigned negative labels during training. That distinction reveals what the model learns from: observed outcomes where available, together with earlier human selection decisions.[2]

Its architecture combines a neural network for image sequences with another component that follows information through time. The input includes images covering development across several days. It learns relationships between those patterns and a later fetal heartbeat, then converts its output into a ranking score. The score helps order available embryos; it is not a direct measurement of the patient’s chance of taking a baby home.[2]

The transfer made the comparison real

Participants were women under 42 with at least two early blastocysts available on day five, so there was a choice to make. One embryo was selected for the first transfer, either fresh or after freezing. The study’s sponsor, incubator and software manufacturer Vitrolife, is identified in the public research record.[3]

The embryologist recorded a conventional choice before randomisation. In the AI arm, version 1.2 of iDAScore then ranked the embryos and its highest-scoring embryo was selected. Patients and treating clinicians remained unaware of the assignment until after transfer; embryologists carrying out selection could see it.[1]

Why close percentages did not settle the main question

The trial asked whether AI was no more than five percentage points worse on clinical pregnancy. The observed difference was −1.7 points, with a 95% confidence interval from −7.7 to +4.3. Because the lower limit crossed −5, the study did not demonstrate noninferiority. A difference that is not statistically significant does not, by itself, establish equivalence.[1]

Researcher Christos Venetis explained that the conventional group performed better than expected when the trial was planned. Under the revised assumptions, a much larger study would be needed to resolve a difference of this size. The result therefore leaves a precise uncertainty about pregnancy outcomes, alongside a clear difference in the measured assessment time.[4]

In the AI arm’s agreement analysis, the two methods chose the same embryo about two-thirds of the time. This helps explain the technology’s practical role: it often reaches a familiar decision through a faster route. A laboratory could value that consistency and speed even if changing the ranking method rarely changes which embryo is transferred.[4]

Laboratory efficiency and birth outcomes need their own evidence

As checked in October 2026, the UK’s HFEA rates time-lapse incubation with automated embryo analysis black for improving the chance of a baby for most fertility patients. In its rating system, that means the balance of moderate- or high-quality evidence shows no effect on that treatment outcome. The authority separately recognises the convenience of continuous observation without disturbing the embryos.[5]

The classification concerns the patient outcome being claimed. It does not erase a measured reduction in assessment time. Equally, faster ranking does not establish more births. The IVF trial gives developers two concrete paths to investigate: selecting an embryo that improves the outcome, and making a dependable selection process more efficient. Each needs to be measured at the point where its benefit is supposed to appear.

Report an error or suggest a correction