Unlocking the Secrets of the Placenta with AI: Conversation with Dr. Jeffrey Goldstein

Each baby is born with a pregnancy diary they've kept while they've lived in their mother's body, otherwise known as the placenta. In the US, only about 20% of placentas are dissected after birth and carefully examined under a microscope, usually to better understand what went wrong in cases of  intrauterine growth restriction or stillbirth; and for the other 80%, (that's more than 3 million births a year) this diary, that contains information about lifetime risk for diseases for the mother, like diabetes and hypertension, and may also include information about likely also holds information about lifetime risks for the baby's future health, for this other 80%, the information is tossed out.

One reason for this desperately poor use of this resource is cost: it costs too much. Today I'm speaking with a researcher and placental pathologist who is part of an amazing team marrying AI to placental pathology to dramatically reduce this cost and make this information more easily accessible and widely available.

Here is Dr. Goldstein's faculty page where you can find publications and bio information: https://www.feinberg.northwestern.edu/faculty-profiles/az/profile.html?xid=41408

To find more about PlacentaVision: http://www.placentavision.com/

To access the show we referenced in this episode with Dr. Mana Parast, 

click here and here

Transcript

[00:00:00] Each baby is born with a pregnancy diary they've kept while they've lived inside their mother's body, otherwise known as the placenta. In the us. Only about 20% of placenta are dissected after birth and carefully examined under a microscope, usually to better understand what went wrong in a case of something like intrauterine growth restriction or a stillbirth, and for the other 80%.

That's more than 3 million births a year in the U. S. This diary that contains information about lifetime risk for diseases for the mother, like diabetes and hypertension, and may also include important information about lifetime risks for the baby's future health.

For this other 80 percent, The information is tossed out. One reason for this desperately poor use of this resource is cost. It costs too much. [00:01:00] Today, I'm speaking with a researcher and placental pathologist who is part of an amazing team marrying AI to placental pathology to dramatically reduce this cost and make this information more easily accessible and widely available.

Welcome to Making Sense of Pregnancy. This show is a new pregnancy reference. I'm finding and talking with experts doing cutting edge work to better understand what we do and don't know about pregnancy and what you can do to better understand your own experience. Each week, I'll be talking to scientists, doctors, and researchers who are trying to uncover the many mysteries that still exist in reproduction, giving you the most current, evidence based way to approach this enormous transition in your life.

I hope it will become your go to source for how to make your pregnancy better. 

Please enjoy my conversation with Dr. Jeffrey Goldstein.

Today, we're lucky to have Dr. Jeffrey Goldstein on the show. He's [00:02:00] the director of perinatal pathology in the department of pathology and an associate professor at the Feinberg school of medicine at Northwestern. His work combines machine learning informatics and science to better understand maternal and newborn health through a deep investigation of the placenta.

Dr. Goldstein, thanks so much for coming on. 

Thank you. Thanks for having me. 

 Did I miss anything in your introduction that I should have pointed out? 

No, I thought that was very kind. 

Okay. Okay. Good. I'm so excited to talk about your work today. I feel like , your discipline should be leading OB and I'm confused about why, why it's not leading ob, but we'll get to that.

 Let's talk about this, paper on the geniusly termed Placenta Vision. , what prompted this work and why do we need placenta vision? 

Absolutely. 

What is PlacentaVision? So, , PlacentaVision is the name of the project. It's also the name of the software that we developed. And what PlacentaVision does is it looks at a picture of the placenta and it [00:03:00] gives you diagnoses of what you would see if you had done a full pathology exam. So, let's say that someone across the street at the women's hospital delivers today.

That placenta goes to our pathology lab. , there's a group of people called pathology assistants that will examine it, they'll measure the placenta, they'll weigh it, they'll take pictures, they'll cut into it, identify abnormal areas. They'll select pieces of tissue to make slides. Those pieces of tissue, they go to, the histology lab.

 And those are processed overnight to remove the water. Replace that with wax. Then those are put in what we call tissue blocks. Those are cut and used to make slides. And then the pathologist will review those slides for diagnosis. That is not a fast process that takes, I think if you're extremely urgent.

Let's say you have somebody who has a heart transplant. They have a biopsy. , we need to know today whether this person is having rejection so we can [00:04:00] treat them. You could do that in. 6 to 12 hours for something that's a larger organ like placenta, something that's kind of moving routinely. We would say about 3 working days.

 So, a lot of the time by the time that we are in a position to make a diagnosis. The patient has already left the hospital. There's also many things where we wish we had the diagnosis earlier. So there's a diagnosis called early onset neonatal sepsis sepsis So this is a bacterial infection in the infant circulation,

 And these bacteria are usually something that crosses across the placenta during labor and delivery. And so looking at the placenta and seeing inflammation or seeing bacteria can tell you that somebody is at a high risk of early onset neonatal sepsis. But, early onset neonatal sepsis usually shows up by 72 hours, which is to say, three days.

So, this is, I think your previous guest, [00:05:00] Dr. Perez, was talking about this idea that so often we see something and the thing that we see has already happened, and so we're in a position to maybe give some context, but , not to be helpful and intervene. Okay. 

 Yeah to diagnose as the first mover.

 So the pipeline that you described of giving birth, placenta moves to the assistants, et cetera, et cetera.

That's not true for every placenta, right? 

That's correct. We examine about 25 percent of the placentas that are delivered here. , so Prentiss Women's Hospital, we have about 10 deliveries a year. So it's the largest birthing center in the Midwest. I think , there's one hospital in Texas that has more deliveries per year, but it's among the largest birthing centers in the country.

And so we have, , a large, low risk population. And then we have some higher risk patients that are brought in, , by maternal fetal medicine that they don't feel are comfortable delivering in the community. And so in that large group, we examined about [00:06:00] 25 percent of the placentas here. There are some places that examine 100 percent of the placentas.

There's some places that examine. Fewer of the placentas, one of the things we're going to talk about on the paper is this hospital in Uganda umbrella regional referral hospital. They examine 0 of their placentas. So there's a big range. And a lot of that it's not medically driven, right?

It's driven by, what the obstetricians feel, what the hospital feels administrative issues. In in in Brara, it's capacity issues. 

 It feels like obviously in places that are under resourced, this is a real issue, but even in the U S it seems like we're not using placentas to their full advantage in part because we don't want to spend money on it, which seems insane to me.

 So we walk us through the process with placenta vision. , you gave us the, , pipeline for what happens now. What would happen if every hospital was using placenta vision? 

So what would happen with placenta vision would be [00:07:00] that, , someone delivers, that placenta goes to probably to pathology, we would take a picture as we do normally, and then that goes to an AI, and that AI would say, oh, this is somebody, for instance, who has a high risk of early onset neonatal sepsis, and that's information because literally all you're doing is just taking a picture of something, You could have that information within minutes of delivery, well within the time to intervene, make decisions.

 You might have somebody where there's concern for, maybe this is a kid that's going to end up with sepsis, and you look at the placenta, placenta is reassuring. And so then we're not using antibiotics. There's a lot of concern about overexposure of kids to antibiotics when they don't need it.

And so either way, you're really helping to inform that decision. I like to focus on early onset neonatal sepsis because I feel like that's something that is Very, very straightforward and it's very common to worry about, , the value of placenta vision is that there are a [00:08:00] number of diagnoses that we can make from the picture.

So, for the early onset sepsis, we're looking for what's called maternal inflammatory response and fetal inflammatory response, which are changes in the placenta that happen when there is inflammation. And that typically , is a committing infection. The other things that we can see and that we talked about in the paper are meconium staining.

So this is, , just. baby's first poop, essentially, , because they're not eating. There's no food in it. It's mostly, shed skin cells and bile. And that bile is toxic. , and so it's the obese will be able to see was their meconium release. , But Looking at the placenta, we can see how long was it there?

Is it being taken up into the system? Is it something that could potentially be causing damage? So we can look at that. 

Wait, let me stop you right there for a second. You can see that from looking at the surface of the placenta. 

Yes. Wow. So, yeah, so the meconium, so [00:09:00] I think everybody, , is probably unfortunately familiar with color and appearance of vial, , from, uh, well, throwing up, , and that, that bright kind of electric yellow color.

That color will get taken up into the placenta and because , it's sort of on the surface, it's then taken into the surface. And so you can photograph that from the surface. And so when I'm looking at a case, when I'm producing my reports, I'll look at the pictures and I'll see, okay, so there was meconium released during delivery.

 I'm looking at my picture. Do I see, does it, does the surface look discolored to my eyes? And then I'll look under the microscope and say, okay, do I see meconium? What we're saying with placenta vision is, maybe we can get a result that's good enough just from looking at that picture.

 So that's kind of amazing. And this is one of the issues, , that you bring up in this paper is this is the technology that's looking at macroscopic issues. So you're taking a picture of the maternal side and the fetal side, as [00:10:00] opposed to the microscopic issues that a pathologist will examine when the placenta is, dehydrated and waxed and put on slides, et cetera, et cetera.

 So, are we imagining every placenta is photographed and only if there's a red flag on the maternal or fetal side that's macroscopically evidence that it goes to a pathologist? 

I think that that's certainly one paradigm. I think we think about this in a lot of different ways.

I think in a US context and a resource rich context, I think triage is a great use. And. We have to think about why is the placenta being submitted, right? I told you 25 percent of the placentas are submitted. , let's, without arguing about that, let's say that the, that the ones that are submitted, they're submitted for a reason.

And some of those reasons will be things that we can examine macroscopically and some of them won't. So for instance, we're concerned about, a woman who had a fever, high white blood cell count in pregnancy. We're worried that there could be infection. Okay, our tool could probably help you, give you information about [00:11:00] that.

 Somebody who has, for instance, fetal growth restriction. Most of the causes of fetal growth restriction are things where you really need to look at them at a microscope. So this could be things like vascular malperfusion. This could be things about chronic inflammation. So chronic bolitis of unknown etiology, , things like massive perivolous fibrin deposition.

Those are things that you can get a little bit of a hint of them from the photograph, but they're not the kind of thing where the information that we need. Is in those surface photos, 

 Just to give a little more context, these three placental conditions have a few things in common. They all involve abnormal placental structure or function that could compromise fetal blood supply and nutrition. And they are more common in pregnancies that include hypertension or preeclampsia . or autoimmune problems. And importantly, they have a high rate of recurrence, so you'd want to know about them for a possible subsequent pregnancy.

 [00:12:00] That kind of thing is not correlated with the size of the placenta. 

So it's definitely correlated with the size of the placenta.

And that's kind of the caveat of that. The trouble is, let's say you have,, a patient that delivers, there's fetal growth restriction. , we look at the placenta, we take a photo, the placenta is small. That says,

maybe there's something going on with the placenta that is associated with it, but that doesn't tell you what it is. And some of those things may have different recurrence risks. Some of those things will have different implications for instance, if it's if the reason it's small is because maternal vascular problems and then you have Decidual vasculopathy.

So damage to the maternal blood vessels that tells you something about mom's future cardiovascular risk. That's not information that we can get from a photograph. I think, , the reason that I say, I think is

AI models can sometimes give us [00:13:00] information, humans. don't expect them to be able to figure out. , an example of that would be, if I take a picture of your retina, so there's something called like a photoscopic exam, I have really bad vision and so they have to dilate my eyes and take photos of my retinas every couple years.

Those photographs, you can look at that, you can tell how old somebody is, you can tell whether it's a man or a woman. Not you and AI can do that. If you asked an ophthalmologist, they would say, get out of my office. , no, absolutely not. So, the AI can sometimes tell us, find stuff that, that we as humans are not able to find, hopefully.

Like in the retina exam example, there are some age related changes, some damage to blood vessels and things like that, that , are observable and measurable and you can point to them and you can say, yes, I agree that this is a damaged blood vessel. The alternative is that sometimes the AI will hallucinate.

and it will take shortcuts and it will do things that are not, not present in the image. So you don't want [00:14:00] to say that you couldn't possibly see the sigillatory apathy in the photos. But I think that I would be concerned that we were seeing some shortcuts being taken.

Well, I think it's a side benefit, but correct me if I'm wrong, of your project is that you're basically creating a dictionary of placentas matching pictures to pathology reports. And Dr. Parast had talked about a need for some kind of dictionary that we can compare things to, that we, we have this full volume that we can use as a baseline to say, this is how bad 

the growth restriction is right? Something like that, which, , is a reference point. Seems like you're creating that. And given that you're training your AI, I'm guessing that you are seeing the rate of change of accuracy of your AI as you get more and more samples. 

There's definitely an improvement.

 There's also been one of the nice things about this study is that we're looking at placentas from multiple sites. So, the paper that we published was [00:15:00] Northwestern and We also have collaborators at, University of Pittsburgh Medical Center, the McGee Hospital over there. So having data from multiple sites that gives us a better representation, , and better reliability.

, it's super cool. And I'm imagining once this is an app on a phone that the nurse in the delivery room is taking the picture. Is that not what you're imagining? 

 I think it could very well be that way. I think that there are a lot of advantages to that in terms of immediacy.

 I think it's kind of a, at that point it comes down to a question of politics, right? Do you want it to be done by the nurse in the delivery room? Do you want it to go over to pathology and then we take a photo and then that's part of our decision making process? I don't know the answer to that.

 I think we're really excited to try a prospective Deployment to see okay. If we try to introduce this into some kind of clinical workflow, How long does it actually take us to get information to people and are we able to [00:16:00] give useful information? And what's the best way to do this?

. I don't know. 

 I thought because you're looking to use it and under resourced places that maybe don't have a pathology department. It has to be a technology that can be used by someone who's not trained in pathology.

Absolutely. Yeah, I think that, in a low resource setting, that could be. , a setting in the U. S. where people chosen to have a less medicalized delivery, if you will, or in some place like Uganda or their lower middle income countries where there's not a pathology department available, or there's not someone to do that.

 Having midwives or trained birth attendants take these photos and then have that information. I think , that's definitely the way that we would use it there. 

 It's just, it's amazing. This is like one of those things that may fundamentally change. How we think about pregnancy, 

but also how we manage it.

So this is for the civilians out there like me One [00:17:00] thing in your paper you mentioned that you're going to use placenta vision for is to look for retained placenta I was imagining in the delivery room, where people are looking for retained placenta, they're both looking at the placenta, but they're also feeling it for rough spots.

So somehow the AI can take this 2D image and figure out if there's a little piece gone, which seems to me amazing. 

Yeah, the normal surface of a placenta is Gray and shiny and very regular. It's broken into what we call cotyledons. So if you've ever made bread where you have multiple balls of dough and then you cook it and they go together, that's what a placenta looks like.

And they're That appearance is very typical and very regular, and if you have patches of that that are missing, that's certainly something that the, an algorithm is able to detect. 

Amazing. Amazing. Because not every retained placenta has some huge piece missing. That's right. [00:18:00] So, that seems amazing to me.

 I was imagining reading your paper, wow, 

we're using this new technology.

 Right now with photography, which is reflected lights, but couldn't we also use it with ultrasound, which is reflected sound? In which case you could use it during the pregnancy.

Yeah, absolutely. So there's a few ways that ultrasound is being used or can be used. So, ultrasound, , that's a routine part of pregnancy care, looking at the fetus. They will also look at the placenta. Typically, the main information that they're gathering , is the location.

So is the placenta someplace where. , it's sitting over the cervix, and so then that's somebody who should not have a vaginal delivery. Those patients need a c section. Is it sitting over the anterior, of the uterus where we would normally cut to have a c section? So those patients who will, , need to have a different kind of operation or hopefully avoid a c section.

 Does it look like it might be abnormally adherent? , and then there are a few [00:19:00] abnormalities such as something called a corangioma, which you can see on ultrasound. And so there are research groups that are looking at can we use ultrasound to identify other diagnoses in the placenta before delivery?

I think that's one way to address what I would say is a big problem in placental pathology which is that. We only assess the person after their pregnancy is over. With ultrasound, you could be assessing it beforehand. And that's a very powerful technique. , and there's I'm not incredibly familiar with it, but we know that that's potentially coming down the line.

There's also a group of people that do ultrasound after delivery, so they will examine placentas that way. I think that, that's, you can get information like that.

 There's, certainly , a prejudice among pathologists [00:20:00] that, what we do and what radiologists do are not equivalent. And we like our way better. And so I would say, well, if you are going to ultrasound or plus end after delivery, you could just slice it. 

Yeah. Yeah. 

And, that's a much better way of looking at What's under the surface and what's there cross sectionally,

yeah.

 So what I want you guys to do is use your AI technique to look at every placenta and connect it with fetal origins of health and disease kind of stuff, because it seems like it's a treasure trove of information we're mostly just throwing away. 

It absolutely is. And I would certainly agree with that.

, one of the uses that we've talked about. , is in research. We know that there are these large birth cohorts that are done in different countries. There's a study in the U. S. called the ECHO study that looks at, things prenatally, pregnancy, following the kids afterward, and sometimes you have placental pathology done in [00:21:00] those studies and sometimes you don't.

And if you go to the people that design the studies and you say, why aren't you doing plus methodology, they'll say 2 things. 1, it's too expensive. And 2, we can't get anybody to look at the slides. The cost, what I would say is that I've spec'd out the cost of a placenta exam for a research study.

It's somewhere between 150 and 200 just to get to make the slides. And then however much you want to pay a pathologist, which I would say is probably somewhere in like 30 to 50, would be reasonable. That's a good chunk of money. In the scheme of these studies, it's not make or break usually, but taking the photographs is.

Markedly less expensive and that's a way to get , include something about placenta in every study and because we're generating the diagnoses automatically, you're not [00:22:00] necessarily having to go find a pathologist or make time to get the slides. Right? 

 I don't know what your response is to that.

Mine is outreach. Expensive compared to what, compared to the amount of money the woman will spend on her diabetes treatment when she's shocked by it. Cause she didn't know that her placenta showed signs of diabetes and she could have been eating differently., I'm outraged on your behalf. Based on your work, do you have any advice for pregnant people today? I do. It's 

completely unrelated to what we published. , you should get vaccinated against SARS CoV 2.

 Why? Give us the why. 

So there are placental pathologies that we see associated with SARS CoV 2 infection. There is also a, rarely you can have the virus infect the placenta and cause stillbirth. , the thing about this cause of stillbirth, which is devastating and we should be trying to avoid, nobody that has been [00:23:00] vaccinated has ever had a stillbirth due to SARS CoV 2.

And people, we've been tracking this over the last five years, vaccination seems to be 100 percent perfect protective, there's nothing that is that good. It's almost unbelievably, but it is, it's, it's, that's the case.

 I think there's a lot of value in placental pathology. I think that, we should be encouraging people to have placental pathology to read it to understand what's going on to discuss it with their hobbies. I'm somewhat conflicted about that. This is a guess, but there's probably about a hundred sub specialty trained perinatal pathologists that are active in the U. S. right now. So most placentas that are examined in the U. S. are examined by non specialists. , and every specialist will say this, that it's better to get examined by a specialist than a non specialist, but in placental pathology this is particularly [00:24:00] true.

So we have studies that show that generalist pathologists, miss things, miscall things. And,, we have a residency training program here, so we have a program that trains pathologists. They come through our section. We try very hard to train them. We're trying to raise that general standard.

 But the standard in practice right now is not as high as we would like it to be. , one of the advantages of AI, of placenta vision, is that you don't have to be delivering at Northwestern or University of California, San Diego or Mass General to have that level of expertise. that's something that we can build and that's something that we can deploy broadly.

But having AI and using that AI to supplement expertise to sort of upskill people, , that's something [00:25:00] that's really important. 

 That's an amazing way to leverage, the particular skills of your, , training. I'm interviewing people across the spectrum who do all, have all kinds of specialties.

And, , , I interviewed, a, reproductive endocrinologist , Dr. Nicholas Macklin. And he says, we absolutely have to take care of the placenta. It's so important. It needs to happen.

Like three months before you get pregnant and then I interviewed an OB in Texas who's looking at aspirin therapy, to try to ward off preterm birth and maybe preeclampsia. And he says, Oh, the placenta is so important. We got to get that aspirin in at 10 weeks. 

Uh, 

sooner is better, right? , we need to focus more on how that placenta gets laid down as opposed right now we're just fixing what is abnormal as opposed to figuring out how to lay down the best placenta.

Absolutely. I mean, the best time to act is before there's a problem and the second [00:26:00] best time to act is now. 

Yeah. 

Now, as soon as you realize that there's something, something going on, 

 So the change that you'd like to see in pregnancy care would be

a tough question. The change that I'd like to see is I'd like to see placenta vision get used and go out and be used to. At least help us figure out which placentas are going to need microscopy, because I think we don't do a great job at that right now. And that would, if we're not going to examine everybody, , I think that's a great way of figuring out who is at the highest risk and, , getting information to people.

 And in places where we don't have access to pathologists, that's a great way of getting some information to people who otherwise would not get, get anything. 

 And in terms of pregnant women's 

Request for a placental pathology. 

if I give birth at somewhere, can I ask for a pathology? How does that work?[00:27:00] 

I don't know. , what I would say is that there are a few patients that have contacted me and I've spoken to them. , and I know that. When for instance, they have a pregnancy, something abnormal happens, when patients advocate to their obstetricians that I think I need to have the placenta for this pregnancy examined and here are the reasons. You know, and maybe they have had a previous delivery and the placenta showed something very abnormal and they want to see if it's happening again.

Maybe they've had a particular experience, a particular diagnosis that we know that there's something we should be looking for in the placenta. When patients are able to advocate for themselves and potentially educate their OBs about what they could be looking for, , that seems to be fairly effective.

, , if I have gestational diabetes or gestational hypertension, [00:28:00] don't I want my placenta looked at to see what my risk is going forward? 

I think that you would. There are findings that we can have in the placenta that will say, okay, here's your risk of recurrent gestational hypertension or recurrent gestational diabetes.

Or, here is your risk of lifetime diabetes or lifetime hypertension. 

Okay, great. Well, I hope to increase demand on the pregnancy side. We'll do it one way or another. 

That's fantastic. Thank you. 

 Thank you so much for coming on and sharing your work.

I look forward to following. It's, ascendant success. 

Thanks very much. 

Thanks again to Dr. Goldstein for walking us through the state of placental examination after birth that exists in hospitals today, and what we hope it will look like in the near future with placenta vision. Given that pregnancy is a stress test, We could get real health information from the placenta that could impact how we live our [00:29:00] lives going forward.

But as of yet, we don't regularly take advantage of it, basically discard it. But I wonder if an investment in placental examination would represent a giant net gain if it was understood in the context of cost savings from limiting the effects of living with hypertension or diabetes for many years to come.

Thanks for listening. If you like this episode, please share it with friends. We'll be back next week with more amazing research.