Fertility and Sterility On Air - TOC: February 2026

Welcome to Fertility and Sterility On Air, the podcast where you can stay current on the latest global research in the field of reproductive medicine. This podcast brings you an overview of this month's journal, in-depth discussions with authors and other special features. FNS On Air is brought to you by the Fertility and Sterility family of journals, in conjunction with the American Society for Reproductive Medicine, and is hosted by Dr. Kurt Barnhart, Editor-in-Chief, Dr. Eve Feinberg, Editorial Editor, Dr. Micah Hill, Media Editor, Dr. Pietro Bortoletto, Interactive Associate-in-Chief, and Associate Editor, Dr. Kate Devine.

Hello, and welcome back to another episode of FNS On Air, where we are discussing the February 2026, Volume 125, Issue No. 2 for Fertility and Sterility. I'm your co-host, Pietro Bortoletto, and I'm joined by three out of the four legs of the tripod, Eve Feinberg, Kate Devine, and Kurt Barnhart.

Kate, Eve, Kurt, how are you guys? I'm great, Pietro. Well-rested Monday morning, ready to go. I'm good.

I'm staying warm in this negative 10-degree weather of Chicago. Happy New Year, everybody. The listeners should know that I am in sunny, warm Miami, Florida, and I have a University of Miami background on our Zoom recording.

You may or may not know that they're playing the National Championship game tonight for the first time in 23 years, so it's a big day in South Florida. You may be noticing the absence of Micah Hill. Micah's out on a short leave, but will be back with us, hopefully, next month for the rest of the podcast for the year.

This month's Fertility and Sterility has a bunch of really great content. If you've been listening to the podcast for the last several months, you notice that we have a third podcast episode each month, FNS Roundtable, where they spend time going over the views and section of the journal, as well as the Fertile Battle. This month actually has two great pieces of content, one talking about fertility decline globally, as well as the economic consequences of depopulation, and this is led by Dr. Dom DeZiegler, and then we have a Fertile Battle led by ESRM President Liz Ginsburg, entitled to Ultrasound and Trigger or Not in IOI Cycles? That is the question.

So if you're looking for a little bit more in-depth picking apart of those articles, please check out the FNS Roundtable, hosted by Dr. Emily Barnard and Dr. Ben Teippert. In this month's Fertility and Sterility, we have a really wonderful seminal contribution by our very own Dr. Eubanks. Eve, tell us a little bit more about Beyond the Numbers.

I love it when an article starts with a catchy title and a colon. Yeah, so the title of this is Beyond the Numbers, AMH and AFC are not independent predictors of embryo euploidy in IVF with PGTA, with Allison Eubanks as the first author, who's one of the REI fellows from the NIH. Allison's also our incredible FNS editorial intern, and co-authors are our very own Kate Devine and senior author Micah Hill.

So Kate, I'm looking forward to having you jump in to talk about this as well. So the objective of this study was to study the association of AMH and antral follicle count in patients undergoing IVF with PGTA, and to evaluate whether these are independently associated with embryo euploid rates in patients undergoing IVF after adjustment for maternal age and other confounders. The study design, this was a large retrospective cohort study using centralized electronic medical records obtained from U.S. fertility centers, encompassing more than 90 fertility clinics across the United States, although I think that this was probably restricted to a few of those centers and not the entire clinic, just based on the numbers.

So the study population was 11,473 women who underwent 13,451 autologous IVF cycles with PGTA. Data were collected between October 1, 2016 and August 1, 2024, and all of these cycles used NGS for PGTA. They included only autologous cycles, and then this was good.

There was AMH and or AFC measured six months within the date of oocyte retrieval, and all patients had available blastocyst biopsy results. So as I alluded to earlier, donor cycles were excluded. Cycles with missing or outdated ovarian reserve markers were excluded.

Patients who didn't have blasts available for biopsy were excluded. Those who didn't have PGT results reported were not included, and embryo-only, embryo banking cycles were excluded, and mosaic embryos were excluded from the euploid group. So patients could contribute multiple cycles to the data, and this allowed for assessment of interpatient variability and minimizing cycle-level selection bias.

So what did they really look at and what did they find? The primary outcome was the percentage of biopsied embryos measured as euploid for each cycle. The secondary outcomes per euploid embryo transfer were things like live birth, clinical pregnancy, implantation, and spontaneous abortion. The AMH categories were divided into three.

So less than 1.1 nanograms per milligram was the lower DOR, and that's consistent with the bologna criteria. 1.1 to 4 was the normal ovarian reserve or normal AMH, and greater than 4 was considered to be high. In the AFC, less than 7 follicles was considered to be low.

7 to 16 follicles was considered to be normal, and 16 follicles was considered to be high or more than 16. What I really liked is you guys looked at it categorically, and then you also looked at it as continuous variables to assess nonlinear associations. You did GEE to account for a clustering of multiple IVF cycles within patients, and then the models were adjusted for maternal age and PCOS diagnosis.

Sensitivity analyses tested the interactions between AMH and age and AFC and age, and then analysis of clinical outcomes were restricted to patients who had one euploid embryo. So I think getting to the meat of what they found, first, euploid rates were similar across all AMH and AFC categories, and second, neither AMH nor AFC predicted live birth implantation or clinical pregnancy rates once a euploid embryo was transferred. Finally, as I always say, I feel like I say this time and time again in every podcast and every lecture I give, age is queen.

Maternal age was the dominant predictor of embryo, euploid, and clinical outcomes. And so I think overall, I think it was a really great study that demonstrates that AMH and AFC are not independent or clinically meaningful predictors of embryo, euploid, or pregnancy rates in patients who are undergoing IVF with PGTA. And so I think that's great, and we can reassure our patients who have low AMH levels, who are young, that their age is much more important.

What I'm having difficulty, in my mind, reconciling, and I think the study design just doesn't address this, because patients who didn't have embryos for biopsy were excluded. Like, I do think that AMH is a predictor of who will have a live birth insofar as that AMH levels predict the ability to get multiple eggs, and multiple eggs lead to multiple embryos. And ultimately, if you get that euploid embryo, you are going to have a higher likelihood of having a euploid if you have multiple embryos to select for.

So this study design, AMH does not predict live birth, but I would still argue that when you see an older patient who has a higher AMH, that patient is going to have a higher likelihood of having a live birth than an older patient with low AMH, because you're going to be better able to take that patient through an IVF cycle and find that euploid embryo, but that study does not address this. And so I think within the confines of this study, what we can say is that they're not independent predictors for getting a euploid embryo. Kate, I'm curious to hear your take on it and my interpretation of your study.

Yeah, you're spot on as usual, Ivitz. The question really being asked here is, does ovarian reserve correlate with egg quality as manifested by all of the outcomes that you described? Not, you know, does more eggs make more babies or does MHREFC result in getting more eggs from in ART cycles or in controlled ovarian hyperstimulation, which, you know, I think, yes, we all on this call, I'm sure agree that the latter is true. And this is one of many analyses that have shown little to no correlation between markers of ovarian reserve and here specifically AMH and AFC and egg quality as manifested by the rate of euploidy, which is also, of course, as you say, age is queen associated with probability of success and age is a much better predictor for that.

I remain surprised day after day when I even hear some of our colleagues who are REIs, who don't, aren't convinced yet of this. To me, I think it's fairly clear from an abundance of data and I think it's just helpful for patient counseling when they come in terrified about their markers of ovarian reserve. Kate and Kurt, you guys were involved with this article.

I want to pick apart something that you guys wrote about in the discussion section where you talk about, is there anything intrinsically different about the population of women who is choosing to undergo PGTA versus those that aren't at all different AFC and AMH time points? Do you think that there really is a introduction of bias here by nature of the population that you're studying? I don't think so, but curious to see what you guys think. I mean, I think that it's a different population without question. Nothing comes to mind for me in terms of a way in which it would be associated both with the exposure and the outcome.

And so I'm not particularly concerned about how we should interpret these data based on that. That said, in our population, and I can tell you this firsthand, there are massive differences in terms of who chooses to do PGT. That is obviously socioeconomically determined as well as geographically, and some of our practices do it 100%.

So, you know, you could in theory do a sub-analysis only of those practices where we're doing it 100%, and maybe we'd get a little bit of a clearer answer. But then, you know, is that cohort also different from the other geographic areas in the United States? Yeah, that's a really astute question, because statistically, as Kate said, it's hard to imagine you meet the criteria of a confounder by being, or by associated, you know, both with the outcome and the exposure. But bias is different, right? Bias means there's just something inherently different about a population that sometimes you can't measure.

And, you know, and clinicians and patients often come in with premonitions and are different, and some, you know, so it's really hard to statistically understand this. So I think the cleanest answer is it's just a different population. I think I completely understand, other than money, what leads a person to decide they want their embryos biopsies versus those that don't.

So anyway, we can just kind of leave it there, but without an answer, unfortunately. Yeah, I was going to say, I do often hear people saying, like, you have a low AMH, I'm worried about your egg quality, therefore, you should do PGT. And I think that's the wrong message.

And in fact, sometimes I have patients who have low AMHs, secondary infertility, and they want to jump right into IVF because they have a low AMH. And I'm not sure that's the right answer either. Like, I would take a young patient who had a low AMH straight to IVF, if she wanted multiple children, really, because we don't know the rate of decline.

But in that population of somebody who just wants one more child, like, AMH really just predicts your response to stimulation, nothing else. And I don't know how many different studies that we have to drill that in, but it really tells us how somebody's going to respond to stimulation, not how well they're going to do from a live birth perspective. So, Pietro, to drag this out further is a shameless plug for publicity.

You know, FNS wrote a special article on PGTA a couple of months ago that covered really soup to nuts, you know, the development of PGTA, but also at our discretion included an article of how patients misunderstand, I want everything done for my embryo. Patients that come in and hear bad news, and low AMH is bad news, all of a sudden shift to a mode where I want everything done. And PGTA is not necessarily the answer of everything that should be done in this case.

So that goes into the bias of why someone is selecting that, because they have no other choice, because it's not in their control, because it's offered to them, because it's offered to them with the wrong evidence. So all I can say is, it's a really complicated decision, but this discussion is really important. This paper, it shows that that's the wrong clinical advice.

Just because you have a low AMH doesn't mean you need PGTA. In fact, you could make the other argument, if you've got precious few eggs and a young woman, why would you want to bother biopsying embryos if there's any possibility of harm or getting a false answer. So I agree with whoever said it.

It's ingrained in our society, and I'm not sure why. And I think it's one of these things that I'd really love to flip the switch back to what the evidence shows. Amen.

Well, that becomes a 40-minute follow-up conversation instead of a 15-minute follow-up conversation. I understand the pressures from the field to just, PGTA is fine. Go for it.

Doesn't mean it's right. Well, yeah, but fine is not correct. That's the wrong answer.

Fine is not correct. It's our job to spend the extra, I would say, five minutes explaining the value of the test, both AMH and PGTA. I'm going to take that one step further.

It's our ethical responsibility, and it's our duty to give true informed consent on what a test shows, what are the limitations of the test, and what are the alternatives. So I don't know that it's fine. I think that in a young patient, PGT may do more harm than good, especially in a low patient with low reserve.

You may be discarding embryos that otherwise may have led to a baby. And that's not incontroversial. There's a controversy about the value of PGTA, and we're usually arguing about its value, especially in older women, but we forget the idea that it's really been shown not to be helpful in young women.

So anyway, a discussion for another day. Well, we're going to stick with the theme of controversy in this article. We've talked a little bit about how AFC and AMH may not be great predictors of ploidy.

Kurt, you have an article on artificial intelligence and the reliability of these bottles for embryo selection. I know this is a topic that's been near and dear to your heart as editor, where you've come online to the Fertility and Serenity Editorial Board in the era of AI papers. Tell the listener a little bit more about some of the pitfalls here.

Sure. It sounds like you guys just set me up, because I'm going to be the naysayer again about the early adoption of new technology. I had a feeling.

I had a feeling. But that's not my only mantra, believe me. But this is a good paper, and this is another advantage of let the story flow over you and don't get stuck up in the data.

It's quite a story, and I think that's the reason that it got in an FNS. So again, Stability and Reliability of Artificial Intelligence Models and Embryo Selection for In-vitro Fertilization is the title of the article. The goal is to evaluate the stability and reliability of AI when it approaches embryo selection and therefore ranking your embryos.

And just up front, we're talking about single instance-based embryo morphology feature selection. I know that's a mouthful, and I'm purposely using the terms in the paper, and then I'll translate it to English in just a second. So the model performance was assessed through something called embryo rank order, critical error rates, intermodal variability, gradient weight class, activation mapping, key distributed, stochastic neighborhood embedding were conducted to explore decision-making discrepancies among replicates.

You with me? Now you're just yanking our chain. Barely. Let's just say there's a reason I didn't pick this one for myself.

So this is why it gets really complicated, and people tend to not read this article, but hopefully I'm going to give you a good idea. Now, let me read the conclusions and we'll revisit them, but I'm going to read, again, what I think are the over-technical conclusions of the authors directly from the paper. So it concludes that single instance learning AI models for IVF embryo selection exhibited substantial instability and inconsistency under mining their clinical reliability.

High intramodal variability and clinical error rate raises concerns about the suitability for real-world development. Again, that sounds terrible, but I don't think any of us can translate what that means. Now, when I go back and look at the methods and collisions, they're not wrong.

They just can't say it in English. So let me see if I can tell you a bit about what this paper is all about. So I love the first line of this paper, because this is really what it's about to make.

Trust is a fundamental driver in the adoption of breakthrough technology, including solutions based on artificial intelligence. You have to trust the technology. You have to understand that it's going to give you a reliable, valuable, trusted answer.

So several AI tools for IVF are being tested in clinical trials. However, there are many studies, and one found that eight commercial available AI systems often disagree with experienced embryologists, that's number one, and disagree with themselves, number two, when ranking embryos. In fact, two systems perform no better than random chance alone.

So why is this concerning? It's because AI is influencing decisions and ultimately patient outcomes. Although these differences might reflect differences applied across AI at different clinics, which is the biggest thing, my AI is for my clinic, your AI is for your clinic, each model in this study was adjusted for clinical-specific data to try to remove that variable. So these findings therefore raise important questions about how reliable current AIs are.

And it's rather than saying, how good is my model, this paper is saying, why do models disagree? What is it about AI that's making different decisions, and should that be of importance to us? So this study doesn't introduce a new model. It doesn't say my model is better than yours, or my model has got better features. It's just evaluating why the models are not giving you the same answer, and does that actually matter? So it was done in two large sets.

It basically used embryos, and most of the technology team was from Massachusetts General Hospital, and then they used some validity embryos and embryo selections from Cornell. So the first thing they tried to do was they ranked variability evaluation. So what does that mean? So this study, they tested how stable AI models were by training it multiple times using the same data, by starting each training with slightly different variables or different circumstances.

This is something called a seed. I learned a new vocabulary. So in other words, they seeded the model to say, learn on this, and they can change it slightly to say, learn on something slightly different, or change the variables a little bit.

So if the model is truly reliable, these small differences should not really result in big recommended changes in any meaningful way. So if, in other words, the model is picking the right features, it shouldn't really matter what the seed was. So what they did was 50 different seeds, and some specifics here.

This is AI for only good prognosis patients, because they only use patients that had four embryos to choose from. The rationale was you can't give a ranking if you only have a couple of embryos. So ultimately, these were good prognosis patients with lots of blastocysts.

So when they made 50 different models with different seeds, and then made a predict on these really good prognosis patients, they ended up testing 92 patients from MDH and 49 from Cornell. So how do they assess it? So I'm not going to try to get too specific here with the statistics, because intuitively, you should understand this, and then you can learn the statistics later. So they use something called Kendall's W to measure how consistently different versions of the model rank the embryos for each patient.

It's not different from when I learned kappa as a residence. You know, if the kappa is really good, it's a one. If the kappa is really lousy, it's a zero.

It's somewhere in between that shows you the reliability between the predictions based on the different things. So the average W summarizes the consistency. So to address how the model performed in clinical practice, they looked at something called transfer rates and live birth rates.

Transfer rate was defined as the proportion of patients in whom the clinician selected the embryo ranked highest by the model, meaning that the model and the clinician picked the same embryo. The metric for live birth is slightly different, because you could only then say, did the embryo predicted result in a live birth if the embryo was transferred and you had data whether it resulted in a live birth? So therefore, that data set is smaller, because you don't have the outcome of every embryo in the data set. Now, I'm going to pause for a second, because that always bothers me, and it's never clear in any of the papers I read, which is that if you're predicting the live birth, and you don't have the live birth potential from every embryo in your pool, how do you know whether you're predicting it properly and how many times you made a mistake? And this paper isn't any better in telling me how many embryos of this pool they tested it on.

Did they have live birth results rates? Did they know live birth results on 50 of them? Did they only know it on 25% of them? But anyway, that's a rabbit hole, but I still think that the clinical prediction rate is hard if you don't have a denominator on all of the patients. All right, you're with me so far. I can see you guys nodding, and I can tell you to stay on the road, hang on with me.

So the next important thing, which I thought was a really neat twist in the paper, was they evaluated something called clinical error rate evaluations. So what does that mean? So let's go back to embryology. They basically graded all the embryologies.

They were all, by the way, day five. They didn't restrict it to day three embryos, or they didn't go out to day seven embryos. And they basically ranked them according to the Gardner stage, and they basically said anything below a three was not a very good embryo, and a grade three or up to five was the embryos that we're used to working with that we would actually transfer.

So a critical error was defined by ranking a grade one embryo or a grade two embryo when a grade three or better was available. So basically saying you're going to transfer an embryo that might not even be a full nice-looking blastocyst when there was another embryo that was a nice-looking blastocyst available to transfer. So how often did that happen? Surprisingly high.

So the answer, and again, the answer is there was about 15 to 17 percent in the different data sets. So, you know, well, when a one or two out of 10 times you're transferring an embryo that the embryologist would have deemed as probably a very, very low-grade embryo. Curt, just to interject, like not even just the embryologist would have graded as a very low embryo, but a lot of those had actually arrested prior to blastocyst stage.

So like not even a transferable embryo that this machine model was picking. Right. And let's hold that thought because I'm going to go back to it because people listening to me in the car are going to say, well, I wouldn't have transferred that embryo, but let's get back to that.

If you wouldn't have transferred the embryo, why are you using AI? But let's continue. So then they looked at the variability in single instance ranking order. So most AI papers have metrics that we understand, like they compare the prediction to what it actually was.

And you get metrics like error under the curve, sensitivity, specificity, accuracy, there's something called an F value. And those are misleading is what this paper is saying. First of all, when you look at them, they're never as good as I think they should be.

Their sensitivity is like 60% and the accuracy is 50%. But if you just looked at those metrics, you would be misled is what these authors are saying. Because what they're saying is that even though the 50 models had similar test characteristics, they actually had very low agreement using the Kendall score.

And the Kendall score is around 0.3. I think you could use it just like Kappa, 0.1 to 0.3 is bad, 0.4 to 0.6 is moderate, 0.7 to 0.8 is good. So this agreement was actually bad between the models. And they couldn't find out that this was generalizable to centers or even to the cohort.

So these failures of picking the wrong embryos were not predicted by the standard metrics, including accuracy, area under the curve, et cetera, highlighting the robustness between clinics and between AI models. So finally, they came up with something called intermodal interpretability evaluations. And this shows that when models were with similar performance relied on markedly different features and embryo regions indicating no shared decision logic.

Did you hear what I just said? 50 different models trained slightly differently, all picked different variables, had really very little overlap between what they're using to predict, and they didn't really show clinical logic. So in other words, they all got to about the same stage by forcing a prediction onto the learning, but they were all picking different things. So we don't even know what AI is picking, so to speak, and whether it's picking the right thing.

So even the highest performance models interpreted the same embryo very differently and inconsistently. And this was across different cohorts. So again, I'm using their words now to quote, such variables raises clinical concern as model recommendations may be driven by unstable or non-overlapping features rather than reproducible biological signals.

So again, it's the black box problem. You don't know what's predicting it, and you might be predicting it based on a spec in the microscope or the magnification of something rather than cleavage size or something you think is important like intercell mass things. So that was a mouthful for them to say that they didn't like the way AI models were predicting, and they were very, very unstable.

So let me go through a little bit more. Are you just saying that the AI models are exactly like us, fallible, prone to error, inconsistent? Is that what I'm hearing? I think they're actually worse because they're picking embryos that have stopped growing. And so I think what this shows, and Kurt, I don't want to steal your thunder, but it's worse than humans, Pietro.

Well, let's get to that because it is it, right? So let's go through what I think are the real important discussion points here. So the models have tools to support embryo selection. We all agree with that, but the question is they have to be consistent and reliable for them to work.

So currently AI models typically optimize metrics such as, again, area under the curve and things like that. However, in practice, these models are used to rank embryos, and if the ranking is not going to be consistent, then it doesn't matter how good your area under the curve is or your sensitivity. So lesson number one is, what is the correct outcome for AI models to be published? And this paper is basically saying it's not sensitivity, specificity, area under the curve.

If you publish another paper like that, you're just going to add to the noise and it's not going to tell you what you want to know. So that's lesson number one in interpreting a paper. Number two, the findings show similar predictive performance do not ensure you're picking the reliable embryos.

So again, the models are picking on different images and different features, and they're not going to give you the same clinical answer. So lesson number two, we're not sure what AI should be picking. At least in embryology, we thought the inner cell mass has this predictive value or how much it expands.

It's based on some thought process of embryology, whereas this AI can pick anything it wants, and sometimes it might not pick the correct variables, and that might lead to instability to the model, as I said. And again, so consequently, clinicians may unknowingly rely on AI recommendations that are inconsistent, non-reproducible, and sensitive to local conditions. So what do I mean by a local condition? I guess it's years ago now, we talked about AI, and I gave the analogy of the AI asked to differentiate a husky from a wolf, because that was, hey, they're pretty close.

And it found out that AI was picking snow, that if it saw snow in the picture, it must have been a husky. So it didn't even care what the wolf looked like or the husky looked like. So it can be analogous to where you're just picking on some nonsensical variable, because it has predictive value in your particular images in your particular clinic.

So lesson number three, do we know what the right outcome is? And most importantly, if you're then going to predict live birth, I guess lesson number 3a is how do we predict live birth when we don't know what the live birth is on all of the embryos? And are you making up for the live birth by doing a second transfer? And would it be the same predictability in the second transfer when you're changing so many things for your second transfer? So it's really hard to get these metrics. So let me end, and I want your comments on this. Does it matter, right? If AI does not work as, quote, as advertised, as they're saying right now, it may simply be that we're just choosing between good embryos anyway.

You could argue that if it picked that non-viable embryo, your embryologists are going to pick it up, that we're not going to select it, they weren't going to transfer it. That's a mathematical exercise. But then again, why are they using AI? Why are you publishing that you can do it better than the clinician? Why are you charging extra money for it? Why are you spending all this going on? So it could be that it's just a very, very expensive placebo, and that really the outcome is it's not going to be wrong because you're going to correct any error you had by actually having a good prognosis patient number one, having a follow-up embryo transfer saying, you know, still it was only 60 percent, you were going to get pregnant.

Sorry, you know, it was your fault, not the AI's fault, you didn't get the 60 percent, and we'll make up for it on the second embryo, which you might have put in first in the first place. So there's almost no way for AI to fail. Now, can it really make the decision process worse? I don't know.

It's not like, pardon me again, PGTA. PGTA has a real risk of harm. You're biopsying an embryo.

The papers that say that the biopsy of the embryo is equivalent to not biopsy are just false. There is a harm to biopsy the embryo, but AI isn't doing that. So what's the harm of picking the wrong embryo first from a pool of three or four really good embryos? So probably very, very little.

So the difficulty then again goes back to AI is a black box. This is telling us we don't know how it's developed, we don't know what variables it's picking on, and ultimately all AI models look good, probably because you're taking advantage of good prognosis patients that have multiple embryos to choose from, and it probably wouldn't matter which embryo you picked first anyway. So that was a very long story to say we have a long way to go with AI in picking what's the best embryo.

It's intuitive, makes sense to our patients, makes us feel good, but I'm really not convinced by this paper, even though it didn't give you new data, that it's actually doing better than what we're already doing. Yeah. I mean, it doesn't make me feel good.

And I would go further than saying we have a long way to go for AI and embryo selection and just say, let's put this back on the shelf. You know, I think there are a lot of valid applications potentially for AI in our field, level loading our retrieval schedules in large volume clinics, maybe even a timing of trigger to optimize the number of mature eggs. I can see an application there that would benefit patients.

Maybe even assessing egg quality for patients to understand, have they frozen enough eggs to achieve their reproductive goals? And or is this a good donor egg lot? How many eggs do they need from this donor egg lot? But at the end of the day, patients have a very limited number of embryos to work with, and we're going to sequentially go through the ones that are suitable for transfer. So the potential upside here is actually quite small until and unless we have in vitro gametogenesis, which I hope is not anytime during my career for better or worse. But I think that was a really, really nice description of the limitations of AI in general and for this specific application.

And I think to your question about does it matter, the companies that have created and are bringing these models to market would say, no, it doesn't matter. We're like, who cares what's in that black box if what comes out the other end is an embryo that's more likely to make a baby? But you really highlight the dangers there, that if it is picking based on criteria that are not actually associated with the outcome, and then you change the context of the test significantly, it could do harm. And again, I think the harm is relatively low since we're going to sequentially transfer the embryos anyway, and you laid that out nicely as well.

But I would say don't use it. That's the advantage of being a venture capitalist in this area is that it's really no downside because you're not saying discard the embryo. You're not saying do any harm to it or do anything differently.

All you're doing is changing the order. And probably there's a good embryo there to begin with because you're dealing with good prognosis patients. So you're really, really selling a very, very expensive placebo.

I think that a priori, biologic plausibility is somewhat lacking. We all know that chromosomally abnormal embryos can implant and can lead to live birth. And so looking at these embryos in the laboratory to see which one is going to make a baby, I think is somewhat of a fallacy.

Like you can have a trisomy 13, you can have a trisomy 18, you can have a mosaic trisomy eight that makes it beyond 20 weeks. And so I think that biologically, it's just not really plausible to think that you can predict in the laboratory by five days, which of these embryos is going to make a baby or not. And so I think like, go back to the basics on like what you're trying to study is not consistent with how these embryos act in vivo.

And so I think like it's a flawed premise to begin with. Yeah. And let's end with, there's a ton of papers on AI, right? And if anything, I hope you heard from this diatribe that if you're going to get your paper published in a reputable journal, you need to do more and then just show that AI predicted what I would have predicted with sensitivity or under the curve, because that's just not telling you anything about it.

You're going to need validation. You're going to need validation in an external population. You're going to need to prove that it works because it's no longer novel to say a computer can pick an embryo as well as my embryologist.

I want to give a shout out to the AI research in our field in the last decade. You've probably seen the names Charlie Borman and Nikita Zaninovich on a lot of these papers. I think it's nice when these researchers are being thoughtful about the limitations of the field that they're hoping to build and expand on.

So it's nice to see studies that call that out and are a little bit more critical. All right. And again, two papers in a row, I'm not the naysayer of all technology.

I'm not trying to bring us back to the dark ages. I'm just saying just because it's shiny and new doesn't mean it's better. And you really do need some robust science behind it before you start making claims and selling it when, in fact, it might do nothing or it might actually do harm.

Kurt wants to go back to Zift and GIFT is what I'm hearing. MPG Tito's. Thanks, Kurt.

I think the controversy continues and you, I think, managed to do the best job out of all of us picking apart some of the nitty-gritty methodology and shortcomings of some of these papers. So thanks for that really enlightening discussion on artificial intelligence models. Kate, we're going to get back to the bread and butter of the journal.

We're going back to assisted reproduction, specifically perinatal complications of fresh versus frozen pregnancies. I feel like we've hashed this out a lot of times. What's new in this article that adds to our general understanding of perinatal complications? Well, speaking of shiny and new, I think this paper is also an opportunity to talk a little bit about some trends in methodology, and I'm interested in all of your feedback, especially yours, Kurt.

So Yan et al. out of Beijing evaluated 2,900 pregnancies from 2017 to 2021 that had been achieved among 50 sites in China. And so the group compared the primary outcome of perinatal death or birth defect among fresh versus frozen transfer.

So this is a retrospective study, but the analysis was conducted one-to-one, and the title includes the phrase hypothetical trial. So it's a one-to-one comparison of fresh versus frozen. So essentially, this was matching in the context of this hypothetical, aka emulated trial.

PGT transfers were excluded. No differences were found in the primary outcomes or in the secondary outcomes, which included preterm birth, SGA, LGA, low birth weight, gestational diabetes, and thyroid dysfunction. The authors did, however, find a twofold increase in the incidence of hypertensive disorders of pregnancy among FET-conceived pregnancies relative to those from fresh ET.

So again, very topical there, and we'll get to that in a moment. The authors state that they addressed this issue in the context of a retrospective study while maintaining the rigor of RCT principles, quote-unquote. In actuality, though, this is a propensity score matching study, and Table 1 describes the hypothetical trial methodology versus what they refer to as trial emulation methodology.

But really, they did not report results from these methodologies separately. Rather, they report the results of a propensity score matching evaluation. And to just kind of spoiler alert on the finding that PIH was twice as likely to occur among the FET pregnancies, here we're talking about 11 cases of PIH among the fresh embryo transfers versus 24 among the FETs.

And this is, again, a study that took place in 2017 to 2021. So I don't really think that we can draw super meaningful conclusions there. And so we don't need to belabor the point.

I think it's a very important question. I think it's a somewhat novel way of looking at it. I do think we need to be really careful about retrospective studies that purport themselves to be equivalent to an RCT in their methodology because they do propensity score matching.

I think this is sort of the study design of the year or of the month. We are seeing more and more of these as editors. It's been a topic that we've been discussing a lot at our statistical meetings, and it's really important, the devil's in the details, in terms of how the propensity score matching is done and whether it's really adding something relative to a very well-conducted retrospective cohort study with really well-done modeling and adjustment for confounding variables.

At the end of the day, whenever you're doing propensity score matching, you are removing cases. You are removing data. And so, you know, we lose something there, and we need to just think hard about whether it's actually adding to the study or is it a cool fad.

And so I think this is a good paper. It's overall reassuring about fresh versus frozen transfers and not finding differences in the perinatal outcomes, but I thought it was a good thought exercise in terms of the methodology. Kurt, I know you're pretty hot on emulated trials, and we discussed this at our annual meeting, at our live podcast.

Do you think we're headed towards a world in which it's really hard to do real trials in our field for, you name it, funding, patient acceptance, moving most of our research out of academic centers and into the private equity world? Are we looking at a world in which we're going to be doing primarily hypothetical trials in our field? Well, I think it is hard to do a really good randomized trial in general and in REI in particular for lots of reasons, and there's lots of shortcomings of them, generalizability and things like that. But Kate really nailed it when she said that an emulated trial can remove a lot of the confounding and bias, and you can get a pretty good answer if it's done right and if it's done robustly. But just claiming that the title is emulated and claiming that you control confounding with propensity score without really doing it well or in multiple different ways is not really helping.

So be careful of the keywords, I guess, is the answer. I do think you're going to see more and more of these trials, and if the rigor is there, I'm happy to look at them in FNS, but please have it be rigorous, not just keywords. I have another propensity score matching paper myself within the ART section, and this one pertains to the differences in outcomes in luteal versus follicular stimulation starts in IVF cycles.

I love a good luteal phase start. It's one of my favorite approaches to ovarian stimulation. It's flexible for patients, flexible for clinics, and if you're a believer in priming for synchronization of follicular development, it negates that issue entirely in cryo-all cycles.

So lots of really wonderful benefits to it. I will share one anecdote where I have been burned by luteal phase starts before. This was in fellowship, an oncofertility case, and the patient had an undetected, unassisted conception in that luteal phase.

We went through a fertility preservation cycle for her before undergoing a chemotherapy start, and she developed terrible OHSS from her dual trigger from that unassisted pregnancy that ultimately did continue into a healthy live birth, but certainly through a wrench in her onco plan as we were focusing on the FERT part of it. Notwithstanding, there is lots of literature that luteal phase starts are efficacious, they're safe, they yield very similar oocyte for patients. This paper is trying to go beyond that.

This is a paper by Barbara Lawrence et al. with senior authors Baris Atta and Human Fatemi, and they attempted to evaluate euploidy rates and embryo quality in luteal versus follicular starts in patients undergoing cryo-all for PDTA. This was a single large Middle Eastern center from 2017 to 2024 that compared retrospectively luteal versus follicular stimulation with antagonist cycles.

Luteal starts were confirmed with P4s being elevated above 1.5, and in total they had 3,500 cycles in the follicular start group and 552 in the luteal phase start group. Average age was similar, 37. Average BMI was similar, 26 to 27, and the one difference that we'll call out is that patients with luteal starts had slightly lower median AMHs, 1.5 versus 1.8. So because of the lower ovarian reserve parameters, the luteal phase group did have a slightly smaller number of median oocytes retrieved, 10 versus 11, but there were no differences in mature egg count and no differences in euploid blast counts in both groups.

Now, I did mention that this was a propensity score matching study. They matched on age, BMI, AMH, and AFC to see if that difference persisted, and really what they noticed is that the luteal phase stimulation lasted on average one day longer. Patients consumed one day extra worth of gonadotropin, so about 600 to 700 units extra, but there were no differences in the egg yield, mature egg yield, blastulation rates, biopsy, blob blasts, and ultimately euploid blasts.

And if you're marching this out to what happens to these embers when they're used, the live birth rates are also no different. I think this continues to provide nice biologic plausibility for us that there is no difference in the eggs that we retrieve in the follicular versus the luteal phase and no difference in those embryos or how they behave when you go to use them. There's always been this theoretical concern about what about that progesterone exposure in the luteal phase and is that somehow affecting these embryos, but I think we now have a preponderance of data that this is legit, it is safe, it is efficacious, as long as there's not that unassisted pregnancy in the luteal phase where I have at least been burned before.

But Eve, Kurt, you guys use luteal phases, do you reserve them only for your ongo fertility patients or are you more liberal with it? We use them a lot. We've been burned more than once with existing pregnancies and pretty severe ovarian hyperstimulation, so I think that's a very, very real, not a one-off. I do want to highlight though that these data and I think my own anecdotal data have shown a longer stimulation and more gonadotropin dosage being used.

And so I think for the self-pay patient, I mean it was about a thousand dollars more of gonadotropins per day or per stimulation cycle with an extra one to two days of stim, so depending on the dose, but I think it begs the question of like who gets the burden, right? In our program, we need more flexibility on timing starts. We really try to map out who starts on what day and what type of cycle and so luteal really is a scheduling convenience for us, but we are passing that convenience burden off to the patient by having a longer stimulation requiring more gonadotropins. And so I'm not sure how fair that is, but it is the reality of being part of a larger program.

I think there probably is some difference that I've noticed on where within the luteal phase you start and do you absorb that extra day or not. I think if you go like in the immediate post-ovulatory stage, yeah, you're going to see a day, day and a half longer. But if you go in the late luteal phase, the pre-follicular, in my experience at least that I haven't tabulated this, but those stim at about the same rate and don't really, you don't see that extra day, day and a half of stimulation.

But it's a great point. This is a type of stimulation start that for the self-pay patients may not make good financial sense, but for those who aren't experiencing that added cost can make a big difference for flexibility for them and for clinics. We're going to continue now.

We're going to get back into the nitty gritty. We're going back to epidemiology. And Kurt, I'm sorry, you could not get this article.

This one went to Kate. We're talking about neonatal characteristics of children conceived via IVF or IUI compared to sibling births from unassisted conceptions. And I will point out to the listener, unassisted conceptions is the preferred terminology by Kurt Barnhart, MD, editor for Fertility and Sterility.

If you use spontaneous conception, I believe he will reference the New Testament for you about the only known spontaneous conception. Unassisted conception it is. And I really like this paper and hope that Kurt will grace us with some of his epidemiologic pearls as well.

So a reader and colleagues from the University of Utah conducted a sibling cohort study to address the quintessential, really essential question as to whether age and or fertility treatment and or the genetic underpinnings of subfertility are what accounts for the differences observed in children conceived in either low-tech fertility treatment, high-tech fertility treatment, or via unassisted conception. And so they used a linkage between their university-based REI clinic, the Utah Birth Defect Network, and the database of Utah birth and death certificates. Naturally conceived children were compared to their IUI and IVF conceived siblings.

So there had to be at least two children born, one of which was naturally conceived. They specifically looked at the outcomes of preterm birth, low birth weight, large for gestational age, and major birth defects. They adjusted their analysis, in my opinion, appropriately for maternal age, infant sex, the year of birth, gravidity, and birth order.

And it's really important to note again that this study was using data that's kind of old. So these were births that occurred between 1985 and 2018. And at the end of the day, very few of these embryos were actually transferred in a single embryo transfer.

So that's relevant when we interpret these data. That said, the main analysis was restricted to singleton live births, of which 2,705 births had at least one putative, quote, naturally conceived sibling that was available for analysis. Of the 460 IVF singletons that were included in the analysis, only 67 of them, 15%, were actually from single embryo transfer.

What they found in their analysis was that there was an absolute difference, and this is in their adjusted analysis, of 4.3% in preterm birth, 3.2% absolute difference in large for gestational age, and 2.3 absolute difference in major congenital abnormalities among IVF-conceived children. So there were differences consistent with what others have found that said the absolute differences were relatively small, again, ranging from about 2% to 4% for those outcomes. The difference was also found for IUI-conceived pregnancies, and that was also a very small difference.

So when we're looking specifically at congenital abnormalities, they found that there were 12 IUI babies, 14 IVF babies, and 21 naturally conceived babies who had major congenital abnormalities. And so the confidence interval that these authors found for IUI crossed one, so no difference there, 0.7 to 2.7 confidence, and 95% confidence interval versus unassisted conception. And the confidence interval for IVF versus unassisted conception was from 1 to 3.8, 95% confidence interval.

So again, we do see this very small difference. That said, we have to take into account that these are pregnancies where almost all of the embryos for IVF were two embryos having been transferred, perhaps accounting for the differences in birth weight. And then we have to also, I think most importantly, consider that these findings are statistically, but maybe not clinically that significant, especially in the setting of parents who may not have otherwise been able to start or expand their families.

So that's always a question that we come down to. We want to be able to counsel patients with accurate information about potential risks of these technologies, but at the end of the day, the small delta may be a pill that many of these patients are very much willing to swallow to be able to expand their families. So I think a really well done analysis by this Utah team and adds to the body of evidence that we have for patient counseling about what they can expect or worry about, for lack of a better word, when they choose to use these technologies to expand their families.

Yeah. So I agree. I thought it was a really nicely done study.

There are two issues that I wish were expanded upon further. So one of them is what was the birth order of the medically unassisted conception versus the assisted conception? And what is the impact of age on that? And not to say that age influences anomalies, but we know that older age certainly can influence birth weight. And so that may be a confounding factor if the assisted conception followed the unassisted conception and the maternal age was older as well as paternal age.

The second point, and they touch upon this a little bit, is they did use their own internal registry, but certainly not everyone seeks IVF or ART care at the University of Utah. And so they did not have the ability to capture which patients that they may have called naturally conceived that actually used ART or IUI outside of their own center. So I think, you know, yes, the center does the predominance of births in Utah, but I think that's sort of a flawed assumption that every patient that didn't cycle there is considered in the medically unassisted group.

And so I think it makes the data a little bit more murky when you consider that the groups may not be characterized perfectly. Yeah. I think those are all great points.

That said, I do think that they did as good a job as you can with the data that they have. They also cross-referenced SART to look to confirm whether they were missing any assisted conceptions. And in terms of the piece about birth order and maternal age, they did per birth control for those variables.

So if, you know, obviously the mother was older and her gravidity, so prior births also changed, actually, the gravity and parity they looked at. So I thought they did a very good job with what is obviously a very challenging question, right? One that we're always talking about, is it the infertility or the age or is it the technology? So I liked it and also glad to interpret with that caution. The best part of this paper in my mind is this question's been asked lots of times.

The question is, you know, what's the answer, right? But if you can actually get it in siblings, that's a very different population. And the strength of association, if there is one, should be stronger if it's in the same family. You remove a lot of the other confounding from, you know, population data.

Again, this is reassuring. No epidemiologic study is perfect. There are flaws.

You can't take this as the definitive answer, but I thought this was well done and adds to the literature because of that fact about the sibling. But, you know, like we said earlier before, that's a different population. You know, people that have one child without assistance and another child with assistance is not a typical population either.

So, you know, we can still think about it. Eve, I want you to bring us home on this podcast. We're going to end with something that I feel very passionate about, polygenic embryo screening.

I think we've talked about it at length on this podcast, the shortcomings of it and why it's not ready for primetime. So I'm glad to see a paper talking about the governance of polygenic embryo screening. Tell us a little bit more about this qualitative study.

This is interesting. It's certainly a loaded topic, and I feel like it's becoming more and more mainstream in terms of conversations, newspaper articles, et cetera. And I have my own biases and hesitations.

This study aimed to examine how key stakeholders, so U.S.-based reproductive endocrinology and infertility physicians and U.S.-based IVF patients perceive the governance and potential regulation of polygenic embryo screening to estimate polygenic disease risk as well as non-medical traits. So the study design I thought was great. It was a qualitative study.

They used individual semi-structured interviews, and the goal was to elicit a wide range of views rather than generate statistically generalizable findings. And I think it's important as somebody who's done some qualitative research, I think it's really important to understand that you're not trying to generalize these views to the overall population. You're really trying to understand what are the different themes that emerge from these semi-structured interviews.

So the study included 53 participants. There were 27 REIs and 26 IVF patients, and those IVF patients either were currently undergoing IVF or had done so within the preceding five years. So participants were recruited in kind of a narrow time frame between January and December 2022.

So, you know, I feel like the conversations have really snowballed since that time, but they used a combination of sampling methods. Physicians were selected using random sampling and direct outreach, and then somewhat through SREI, a random sample of 250 SREI members were sent email invitations. Data was collected through these one-on-one interviews.

They were done by a single interviewer who was trained in qualitative methods. Interviews were conducted on Zoom. They were recorded, and then they were professionally transcribed.

The researchers used these semi-structured interview guides, and this is really key. How do you develop your interview guide? And this was developed by a multidisciplinary team that consisted of bioethicists, social scientists, and statistical geneticists. Interview questions explored views on polygenic embryo scoring use, or PES, perceived benefits and harms, and then attitudes towards regulation, like general attitudes, specific attitudes.

They also showed some mock embryo reports that were modeled after commercially available PES reports, and these were incorporated to simulate the information that patients will realistically receive from the scoring. And then the interview guides were piloted before they broke them out for the real interviews by two clinicians and two patients, and the interview guide did not change throughout the study period. So the transcripts were coded using thematic analysis.

They used DEDUCE, which is a qualitative software program that allows you to read each interview and tag the themes that you are seeing emerging, and so they use both deductive and inductive coding. Deductive coding is based on the interview guide topics, and then inductive coding is what emerges as you are reviewing the transcripts, like what themes are you seeing again and again. Each transcript was independently coded by at least two researchers, and discrepancies were resolved through discussion.

These sub-themes were reviewed and refined by additional team members, so I think the methodology was very solid. What did they find? And I thought this was interesting because I have my own thoughts and feelings about PES, and I have my own thoughts and feelings about government regulating reproductive health care, and I thought it was interesting. So clinicians generally supported some form of regulation, mostly what they called soft regulation, such as professional guidelines.

They cited concerns about clinical utility, data validity, misuse, and patient harm. Some clinicians noted concerns about PES having limited clinical utility and being more socially rather than medically beneficial. Some supported regulating the commercial advertisement of PES, given that the FTC can police against deceptive advertising of PES by requiring commercial entities to make sure that their ads are truthful.

Now, interestingly, patients opposed regulation, emphasizing procreative autonomy and the perceived intrinsic value of genetic information. And so the authors identified a central tension between two themes. One is what they called a welfarist model, which is favored by clinicians, and that's regulation to promote well-being, and the use of PES should be allowed for the selection of traits associated with well-being and the selection of traits against—selection against traits that reduce well-being, and what is well-being is hotly debated.

Is that medical disease? Is that height? And then a libertarian model was favored by patients, which was minimal or no restrictions. And so, you know, I overall support this as a very well-done qualitative study. I think the authors demonstrate this clear divergence in clinician and patient perspectives on PES governance.

I think the findings really highlight the importance of stakeholder engagement, particularly from my physician lens. I think that you have to have physician lens involvement that suggests that professionally developed guidelines may represent a more acceptable regulatory pathway than formal legislation, especially in current U.S. context and current U.S. political stakes. And again, as I alluded to earlier, I have huge reservations about formal legislation.

I personally don't believe that government should have any role in regulating the physician-patient conversations, but I do think that there needs to be good oversight in the development of tests and the need for clinical trials and or validation studies prior to bringing these to market. And so I think that where I personally believe there should be regulation is in approving tests for clinical use. I don't want a government to regulate like what I can and can't offer patients for clinical care.

So I thought this was really beautifully done, very well written, easy to follow. Some of the diagrams that showed the divergence of opinion and divergence of immersion of themes was really nicely done. And I think it's a good read for all.

So, Eve, sorry, who should make the decisions who are regulated, if not the government? Well, I think that the government, the FDA has some oversight. And when you develop tests that affect consumers, I think we have a pathway for that. But I think that they really need to have stakeholders, both those who develop the tests as well as those who will utilize the test.

I mean, I was very struck and I don't want to name the company, but I had a kind of a small group discussion with one of these companies at ASRM and they completely developed the test and then they bring it to market and just expect that REIs are going to offer IVF for non-medical needs solely for polygenic risk scoring. And I think that, you know, look, as we discussed in the last paper, IVF has risks. And so it is not without risk.

And patients who conceive with IVF often have some of these inherent risks. And so are you trading off, you know, your inch in height that you might gain with a lower birth weight because you've used ART? And how do you balance those risks? And so I think that, you know, this idea of development versus implementation, there's a, to me, there's a big struggle between that. Like I'm certainly not going to offer a test that doesn't have medical benefit.

And I think, you know, I'm quite honestly, I'm disappointed that many of our colleagues feel differently and strongly that patients should be able to choose whatever they want to choose. I think that ART is not without its own set of risks and you are incurring those risks in somebody who may have the ability to conceive without medical assistance. And now you're introducing an additional harm in the process.

And I think for those patients who cannot conceive without medical assistance, like they're willing to take on those risks as we just talked about. But in those patients who can conceive no problem, you're trying to sell something for well-being, but you may be reducing well-being in other ways. And so I think it's complicated.

It has to be the FDA. It has to be. But on the other hand, maybe not the current FDA, though.

Not the current FDA. And the other thing is that they've applied their regulation in such a spotty way. And the cat is out of the bag for PGT.

And so it really is on our eyes. And unfortunately, these technologies, though not ready for primetime, are just inconsistently offered. And it's the Wild West for PGT.

But, you know, regulation for AI is extremely tight. And those technologies have not been able to clear the burden of the FDA, even IVF itself when we back that far. Right.

So I mean, even as we talked about last month, like some drugs are held to a different regulatory standard than other drugs. And so like it feels like the Wild West right now. And my hope would be that our eyes would self-regulate better.

And I think it's disappointing to see that we are not. And I think that the onus is upon us as clinicians. And I think that the onus is really upon clinics to have firm policies.

Like in our clinic, we do not do IVF without a medical indication. We won't do it simply for sex selection. And I know people have strong feelings about that.

And patients certainly do in terms of autonomy and reproductive choice. I think about PES in a similar vein that, you know, isn't medically indicated maybe for schizophrenia and type 2 diabetes, maybe, maybe, maybe, but certainly not for traits. Well, if you're white, North American and of European ancestry, I think type 2 diabetes and schizophrenia, maybe, maybe.

But the minute you're not a white North American with European ancestry, all of the predictive value of that really falls apart. And I think if you're a listener who is interested in learning more about this topic, there is a fabulous paper in the New England Journal of Medicine entitled Problems with Using Polygenic Scores to Select Embryos by Patrick Turley, geneticist in a group of maternal fetal medicine specialists, geneticists, ethicists. It was published in June of 2021.

I have printed it out more times than I can imagine for patients who come asking about polygenic risk scoring at my clinic. So I think this is something that unfortunately we're going to be having to talk about more and more as a field. I'll have kind of some shared knowledge.

A shameless plug. I did a fertile battle on this back in, I think, 2022 and some juicy info in there and some juicy perspectives. And so I think like I've certainly gone back to read it and I don't know if the battle continues.

This is something we should think about as a field because it's going to, you know, it can really help our field or it can really hurt our field. Well, unfortunately we don't have time to dig into all of that on this short podcast and we've covered so many wonderful articles already. Eve, Kate, Kurt, thank you guys for a wonderful rich discussion.

Really interesting articles in this month's issue. Micah, we miss you. I know you're listening.

We'll see you hopefully for the next month's podcast. And of course the conversation continues beyond this podcast. We have FNS Roundtable and FNS Unplugged for our listeners, our other two sister podcasts.

And again, if you're looking for more content, follow us on Instagram, Facebook, LinkedIn. We're there. We're continuing the conversation beyond today's podcast.

That's all the time we have for today. Until next time. Bye-bye.

This concludes our episode of Fertility and Sterility on Air, brought to you by Fertility and Sterility in conjunction with the American Society for Reproductive Medicine. This podcast is produced by Dr. Molly Kornfield, Dr. Adriana Wong, Dr. Elena HogenEsch, Dr. Selina Park, Dr. Carissa Pekny, and Dr. Nicholas Raja.