AI embryo selection: introduction and limits
On this page you will learn
- What an artificial intelligence (AI) embryo tool actually does
- How these tools are trained and why the training data matters
- What a score or rank means, and what it does not mean
- The professional and ethical limits on using AI in embryo selection
What the tools do
An AI embryo tool is a computer model that looks at images or time-lapse films of embryos (pages 38 and 39) and produces a number, a rank or a category. Most use a type of machine learning that finds patterns in thousands of labelled images. The label is usually what happened after transfer, for example whether a pregnancy with a heartbeat followed. The model then learns which image features tended to go with that outcome.
Some tools also try to predict other things from the image, such as the chromosome status of the embryo. These predictions are less reliable than a laboratory test, and they should never be described to a patient as a substitute for one (see page 58 for PGT-A).
Why the training data matters
A model is only as good as the embryos it learned from. If it was trained in laboratories with different culture media, incubators, camera settings, patient ages or transfer policies, it may not rank your embryos in the same way. Images of transferred embryos are also a selected group: the poorer embryos were rarely transferred, so the model has seen few of them with a known outcome. This is called selection bias, and it is built into almost every data set in this field.
What a score means
- A score ranks embryos within one cohort; it is a relative ordering, not a forecast for that couple.
- A score from one tool cannot be compared with a score from another, or across versions of the same tool.
- Most tools cannot explain which features drove the score. This "black box" problem makes errors hard to spot.
- A high score does not make a poor-quality embryo good, and a low score does not make a good one worthless.
What the evidence says
Many studies report that AI ranks embryos at least as consistently as human graders, because a model gives the same answer for the same image every time. Fewer well-designed randomised trials compare outcomes with and without AI selection, and the results so far do not show a clear advantage over careful grading by trained embryologists. The field is moving quickly, so you should expect new evidence and read it critically: who funded the study, which patients were included, and whether the result was tested in a separate laboratory.
Professional and ethical limits
- The embryologist and the clinician remain responsible for the decision. "The software chose it" is never an acceptable explanation.
- A tool must be validated in your own laboratory before its output is used, and checked again after any change of equipment or method.
- Patients must not be told that AI improves their chances unless good evidence supports it for their situation; that would be a misleading claim.
- Images are patient data. They must be stored and shared only as the ART (Regulation) Act 2021 and data protection rules allow, and never used to train outside models without consent.
- No tool may be used to choose or report an embryo by its sex. This is forbidden by the PCPNDT Act.
At the bench
In the Genesis course you do not operate an AI tool on patient embryos. You will look at a set of library images that have been ranked by a model and grade them yourself first, blind to the score. Then you compare. Where you and the model disagree, write down why. This exercise teaches the most important skill in this area: being able to look at an embryo and form your own view before a number influences you.
Common errors
- Looking at the score before grading, which biases your grade.
- Treating a rank as a probability for the patient.
- Assuming a tool tested elsewhere works the same in your laboratory.
- Forgetting that an image file carries patient identity and must be handled as a record.
Check yourself
- What is selection bias in an embryo image data set?
- Why can scores from two different tools not be compared?
- Who is responsible for the final choice of embryo?
- Which law forbids using any tool to choose an embryo by sex?
Questions on this page
Can AI tell which embryo will become a baby?
No tool can. AI tools rank embryos within one cohort from image features. They do not give a forecast for an individual couple, and the final choice stays with the clinician and embryologist.
Should I grade an embryo before seeing the AI score?
Yes. Grade first, blind to the score, and then compare. Seeing the number first biases your own judgement.
Can an AI tool replace PGT-A?
No. Predictions of chromosome status from images are less reliable than a laboratory test and must never be presented to a patient as a substitute (see page 58).
The Genesis ART Textbook is teaching text for trainees and doctors of the Genesis ART Academy. It is not treatment advice and not a laboratory standard operating procedure; every technique is learned under supervision. Page numbers never change.
