Helping you categorize AI algorithms for embryo evaluation

by Alexandra Boussommier-Calleja

 

If you’ve attended ESHRE, chances are you do not need to be convinced that AI for embryo evaluation is not just a trend bound to disappear in a couple of years. The interest has now persisted long enough that we all know AI for embryo evaluation is here to stay. All labs will leverage AI for embryo evaluation sooner or later. But how do we navigate the new technology that is out there, and make sense of it? Let us help you do exactly that. 

Let’s put things in boxes

We all love putting things in boxes to help us make sense of the complex world we live in. Let’s do the same to break down AI algorithms applied to embryo evaluations, with these 5 suggested categories: 

1-  Objective vs. subjective labels

Algorithms that are trained on subjective labels follow human intelligence. They have been trained to replicate how embryologists evaluate embryos today, and have been asked to focus only on criteria deemed to be predictive of success by embryologists. Therefore, while they save the embryologists time by automating their usual way of evaluating embryos, they cannot go beyond the embryologist’s current limitation (e.g. when you are faced with poor quality embryos which you deem to not be worth transferring, when we know some can lead to a pregnancy).

Algorithms trained on objective labels are only told to focus on the objective end result (e.g. clinical pregnancy) and are given free rein to focus on any information it identifies as being predictive of such result. They are not biased into only looking at what embryologists have deemed to be important. These algorithms can not only help embryologists save time in helping them reach a decision faster, they can – unlike algorithms trained on subjective labels – complement embryologists and help in scenarios where they lack the biomarkers to make a decision (e.g. when faced with only poor quality embryos).

2-  Static vs kinetic algorithms

Static algorithms focus on a single image of the embryo, generally acquired at the end of the embryonic development on Day 5, with any kind of microscope.

Kinetic algorithms leverage the rich kinetic information from the entire embryonic development by being trained on a series of images recorded by time-lapse systems. 

Today, it is widely accepted that a single image of an embryo is far from representative of its viability, giving an obvious advantage to kinetic algorithms.

3-  Visual, tabular or hybrid algorithms

Visual algorithms are only trained with images of embryos.

Tabular algorithms take any categorical or numerical data as inputs, without ever being exposed to an image of an embryo.

Hybrid algorithms are trained both with visual data and other sources of tabular data to personalize its output and maximize accuracy.

While morphokinetics is crucial to determining the viability of an embryo, it is also accepted that the clinical context can hugely impact the embryo quality and the uterus receptivity. Not all of this information might be reflected or captured in the embryo morphokinetics, and it can therefore increase the accuracy of predictions.

Spoiler: this is where ImVitro stands out from the rest of existing products.

4-  Biological or clinical endpoints

Biological algorithms are trained to measure the chances of a biological event occurring i.e. an endpoint that happens in the lab: ranging from fertilization, to cleavage, blastulation or PGT results.

Clinical algorithms go beyond what happens in the lab, and are trained to measure the chances of a clinical event occurring following its transfer into the uterus (e.g. implantation, pregnancy, birth).

Biological endpoints are generally used as measures of the quality of the embryology work carried out in the lab, following the oocyte retrieval. In contrast, clinical endpoints will also depend on the quality of embryo transfer and on uterus receptivity, and are therefore harder to predict if you only consider the morphokinetics.

5-  Prediction vs. recognition

All algorithms measure the chance of an event occurring.

Predictive algorithms measure the chances an event has to occur in the future i.e. it predicts an endpoint that is not already visible in the image.

Other algorithms recognize information that is already visible in the image.

Predictive algorithms carry more inertia in their clinical validation, as the actual result cannot be immediately checked in the images. But, they help you feel more confident about the potential outcomes!

Diagram showing existing AI methods used to evaluate embryos

Which box wins?

The simple answer is: there is no winner. Putting algorithms in boxes helps, but it comes at the risk of missing subtleties and changes, as technology evolves quickly in this field. 

EMBRYOLY’s core algorithm predictive of clinical pregnancy stands out today because it is trained on:

  • an objective label to complement the embryologist’s way of thinking
  • kinetic data to make the most of all of the information in the embryonic development
  • clinical and predictive endpoints to help embryologist see beyond what happens in the lab
  • hybrid data to personalize predictions and maximize their accuracy

Nonetheless, we are constantly working on integrating the latest advances in machine learning to stay ahead of the curve. Regardless, since a single algorithm cannot answer all your questions, we have developed a series of such algorithms to provide you with multiple data points to truly augment your final decision-making, and add transparency to our recommendations.

Final thoughts

I believe that the key to success over time will be to keep blending human intelligence into existing algorithms, which already today are quite human: remember that, even if the label is objective, it is the embryologist’s eyes that help us prepare the data. We still hold our algorithm’s hands very tight. But more to it, it is because AI should not be here just to automate what senior embryologists already know how to do, even if this already is of huge help to junior ones. AI should be here to also assist them in situations where they have nowhere to turn to. Think about the help algorithms can be when you are faced with poor-quality embryos: do you really need to be told again that they are in fact as hopeless and should be discarded, when we know that some of them can lead to a pregnancy?

Don’t settle only for AI algorithms that have exclusively been trained the exact same way you have: go for those which can complement you. With AI by your side and with you at the wheel, the final decision can only be stronger.   

 

References 

Kragh & Karstoft, 2021 Journal of Assisted Reproduction and Genetics 

References from the diagram 

Barnes et al., 2023 Lancet Digit Health

Berntsen, 2022 PLOS ONE

Chavez-Badiola, 2020 Reprod Biomed Online

Chen et al,. 2019 Fertility and Reproduction

Danardono et al., 2022 J Reprod Infertil.

Dirvanauskas 2019., Comput Methods Programs Biomed

Duval et al., 2023 Human Reproduction

Fukunaga et al., 2020 Reproductive Medicine & Biology

Khosravi et al., 2019 NPJ Medicine

Kragh et al., 2019 Computers in Biology & Medicine 

Lassen et al., 2023 Scientific Reports

Liao et al., 2021 Communications Biology

Myiagi et al., 2019 Reproductive Medicine and Biology

Myiagi et al., 2020 Artificial Intelligence in Medical Imaging

Petersen et al., 2016 Human Reproduction

Tran et al., 2019 Human Reproduction

VerMilyea et al., 2020 Human Reproduction

Zabari et al., 2023 Journal of Assisted Reproduction and Genetics