E055 - AI or Not - Marc Fiammante and Pamela Isom

Welcome to "AI or Not," the podcast where we explore the intersection of digital transformation and real-world wisdom, hosted by the accomplished Pamela Isom. With over 25 years of experience guiding leaders in corporate, public, and private sectors, Pamela, the CEO and Founder of IsAdvice & Consulting LLC, is a veteran in successfully navigating the complex realms of artificial intelligence, innovation, cyber issues, governance, data management, and ethical decision-making.

A newborn can look “mostly fine” and still be minutes away from a life-altering diagnosis. The hardest part is that the most effective treatment, therapeutic hypothermia, only works when clinicians can identify brain injury within about six hours after birth. That time pressure, plus limited access to EEG experts, leaves too many families without a clear answer when it matters most.

I sit down with Marc Fiammante, a retired IBM Fellow who became a PhD student at the Paris Brain Institute for one practical reason: legal access to the medical data needed to build better tools. We unpack what newborn EEG monitoring really looks like at the bedside, why black-box deep learning can fail the trust test in medicine, and how “symbolic AI” and signal processing can create decisions clinicians can audit. Marc Fiammante shares how listening to a neurophysiologist interpret EEGs led to a breakthrough approach that models brain dynamics, producing a distinctive “Batman mask” pattern in healthy newborns and enabling high-accuracy, traceable comparisons across patients.

We also get honest about what it takes to move from research to impact: strict data selection within the first six hours, avoiding confounders, navigating patents and publication rules across the US and Europe, and designing a lightweight, lower-cost path toward a portable device that more maternity wards can actually use. If you care about healthcare AI, explainable AI, AI governance, or real-world digital transformation, this conversation offers a grounded blueprint for building systems that augment clinicians rather than replace them.

Additional Links: 

https://www.sciencedirect.com/science/article/pii/S0169260725005826

https://patents.google.com/patent/WO2025186361A1/en

https://foundation.generali.at/en/  

https://www.thehumansafetynet.org 

Transcript

[00:00] Pamela Isom: This podcast is for informational purposes only.

[00:26] Personal views and opinions expressed by our podcast guests are their own and not legal advice,

[00:34] neither health tax nor professional nor official statements by their organizations.

[00:42] Guest views may not be those of the host.

[00:50] Pamela Isom: Hello and welcome to AI or not, the podcast where business leaders from around the globe share wisdom and insights that are needed, needed to address issues and guide success in your artificial intelligence and your digital transformation journey.

[01:06] I am Pamela Isom and I am your podcast host.

[01:10] And so we have a fascinating guest with us today, Marc Fiammante.

[01:15] Marc is a PhD student with Paris Brain Institute and leader of the Newborn NeuroDigital AI Convergence Project,

[01:27] and he's a retired IBM fellow.

[01:30] Marc,

[01:31] thanks for joining me and welcome to AI or Not.

[01:35] Marc Fiammante: Thank you, Pamela. Thank you for inviting me.

[01:39] It's a good opportunity to share my abortivity and why I went on with that in 2019.

[01:50] Colleague of mine in IBM told me there's some interesting project funded by the Generali Foundation.

[01:59] It's an insurance. We were working doing artificial intelligence at that insurance and he said the foundation is funding a project at Paris Hospital Children Hospital to help detect children that have had a brain injury at birth.

[02:17] Because today only one in six of these children get the treatment, which is therapeutic hypothermia. That hypothermia only works if it has been applied within six hours after birth.

[02:35] Europe the fact that babies are undetected leads to 5,000 deaths of infants a year,

[02:43] 10,000 children being impaired for life.

[02:47] And so that's a pretty impactful disease, you know, having the brain injured.

[02:55] And there's a treatment. So why can't doctors have apply that treatment? In fact, there's the only way to detect that children had their brain hurt.

[03:06] There is a clinical examination which is called the Sarnat or Apgar, where you're looking at the baby, if the baby is blue, had the umbilical cord around the neck.

[03:17] Doctors are pretty sure that the baby has been affected and they decide to put the baby in hypothermia. But then some uncertain cases and the doctors hesitate because if you do hypothermia, you have the baby in ICU intensive care.

[03:33] And so you have to sedate the baby,

[03:36] intubate for food and respiration and so breathing, you know, and it's quite heavy. It's 72 hours at 33 degrees centigrade. Not sure how it converts to Fahrenheit, but it's quite cold.

[03:53] The brain will fix itself in 72 hours because you slow the chemistry in the brain which avoids Neurons that had died to propagate.

[04:05] So when my colleague told me can you help with that by the way, the WHO states on its website and I could send you the link that asphyxia is the cause of 900,000 depths globally per year.

[04:23] So that's.

[04:25] So when my colleague told me,

[04:27] do you think that we, because I was doing AI, we could try to do something.

[04:31] We started a project and it was quite complex because there's this restriction about accessing medical data.

[04:39] So we were at the beginning not having direct access to medical data, but some kind of scaffolding because of anonymity and so on.

[04:52] So we were not progressing very fast.

[04:56] And there had been quite a number of research and detection based on electroencephalography,

[05:05] which was the medical doctor,

[05:08] she was a neurophysiologist. And he said,

[05:11] we are five person like me in France.

[05:14] And so we need to replicate what we are doing because we are were too few to address all the births that happen everywhere.

[05:23] There's too few people trained to do what we do to detect the babies that have had their brain injured.

[05:32] Then while we were making little progress and well, I made fellow there was the COVID and I had an opportunity to take retirement because I had been working for 38 years for IBM and three years in the oil industry before in the military service.

[05:50] So I decided I could do something more useful of my life.

[05:55] Trying to diagnose something that has that impact, you know, on lives.

[06:01] The way to have legal access to the medical data was to become a student.

[06:08] So, okay, so as a student in the Paris Brain Institute,

[06:14] that made me entitled to access the medical data.

[06:18] I didn't want to become a PhD student. To have a PhD, you know, that's not. I don't need any more medals, you know,

[06:29] Pamela Isom: I hear you. Good strategy.

[06:31] Marc Fiammante: And so becoming a student, it was. It surprised a few people. They say, but you can't have a honorary PhD based on your experiences.

[06:42] No, but I need that to be officially entitled to the data.

[06:47] And so I got the data and I use all the AI tricks, you know, that we had been using for. Because in IBM we started doing AI quite early.

[07:01] The name that the svm. SVM logistic regression,

[07:06] comparing images, doing some signal processing, type of machine learning.

[07:11] But none of them had the performance. They were limited to like 92%. Which means you're leaving eight on a hundred babies. You're leaving eight babies out of the.

[07:23] Pamela Isom: Yeah, and that matters.

[07:25] Marc Fiammante: And so what made the difference is me sitting next to the doctor and listening to her analyzing EEGs to make her decision about the babies that needed to undergo hypothermia on that.

[07:42] Because when you do healthcare,

[07:45] all what relates to clinical and how doctor make their decision is the crux is essential. And they do not always express that in IT terms or mathematical terms, you know.

[08:00] So then I could see, for example that he was telling me that you see that some theta waves inversion compared to the delta waves, you know, and there was uncroche, frontal or tracealternal, all those clinical terms.

[08:16] Well, I had to study what it meant. And that's when I could get some characteristics to isolate and then isolated a few like a gene of these characteristics from what she was telling me.

[08:31] You know,

[08:33] I still use those support vector machines and logistic regression. I got up to 95% of results. But what I did and which was quite useful is try to visualize those characteristics in three dimensions.

[08:50] Get three of subsample some of these characteristics and do some color display in three dimension to see how they relate together.

[09:00] Where the populations, you know,

[09:02] the moderate and severe cases, how they are shown in space. I could send you a diagram if you're interested, and where the mild are. And this gave me the feeling that I needed something more continuous in terms of analyzing the dynamics of the brain.

[09:21] I got the idea to say, okay,

[09:24] let's try to have a feeling of how the brain power behaves. Because newborn babies, they have something which is called tracer alternate.

[09:36] It's not silly, but it's kind of very fast alternating power brain power behavior where you could have high power followed immediately by a lower power.

[09:49] And so you have to get an idea of the dynamics of the brain.

[09:54] And so.

[09:56] And now I can talk freely because there's a scientific paper and there's a patent.

[10:01] I said, let's analyze how long the signal stays in a given power range.

[10:09] And so does it stay in that power range for 1 second, 2 second, 3 second, and then make a kind of surface which is duration for one axis and power on the other axis and probability of that happening.

[10:27] With normal babies, you get something like. Which is in 3D. Looks like the Batman mask. Two peaks,

[10:35] like pointed ears,

[10:39] which are one second type of lower power, higher power.

[10:44] And it has a rounded shape. You know, when you do this 3D.

[10:49] And when I showed that to my colleague, because I had been using a kind of bluish purple color for.

[10:55] For the.

[10:56] My baby. Oh, that's the Batman mask.

[10:59] Pamela Isom: The Batman mask, huh? Okay, okay.

[11:02] Marc Fiammante: And so then when I found this, and we search and we searched patent databases and Then nobody had thought about doing that. So we said, maybe we have a potential patent.

[11:15] The Paris Brain Institute, in fact is linked with the Paris hospitals, the National French Research center and the National Medical Research.

[11:24] And they all have kind of patent office that helps them file join patents between the four and the.

[11:33] So Paris hospitals decided to file a patent for that. And so, well, they get. But getting the invention does not mean that it's going to be used in the market.

[11:44] Pamela Isom: I understand.

[11:47] So I want to zoom out a little bit.

[11:49] And so you talked about your self, your career journey,

[11:55] and you gave us insights as to your research and helped us to understand your connection between the Paris Brain Institute and the Newborn NeuroDigital AI Convergence Project.

[12:09] So you kind of started to connect some dots.

[12:12] I would like to understand how AI. I heard some of it, but tell me more about how AI is coming together with this neurodigital convergence to help with addressing this problem and quickly solving the problem faster that you uncover was occurring with our newborns.

[12:37] Because this is a big deal. So tell me more about how AI is assisting with that problem diagnosis and resolution.

[12:45] Marc Fiammante: So artificial intelligence has two flavors, one which is deep learning or machine learning, and the other one which is symbolic AI.

[12:56] And symbolic AI is more about measuring things and having rules and making out doing classifications,

[13:07] I would say in the more classical way, as we were doing in the past. Okay.

[13:13] And I just want to mention that in my approach to solving the problem,

[13:18] I never ruled out any of the approaches.

[13:23] So I didn't want to be like these animals in the car headlights and they do not move anymore. So there were headlights with machine learning during neural networks with the EEGs, which I tried.

[13:39] But a former colleague who went to Nvidia told me,

[13:43] if you can find better characteristics at the beginning, that can discriminate before you do machine learning. That will always be more traceable and trustful than doing machine learning. Because you don't want to have a black box that tells a doctor.

[14:01] And what am I going to tell is a real case where they say the entropy of the signal is not in this range, so the baby should be put in hyperthermia.

[14:11] Does a doctor know what an entropy of a signal is? I don't think so.

[14:16] If you tell the doctor those babies that have not the trois alternum, that's something they know, you know, for normal babies, which means the sleep wake variation, then they understand that.

[14:28] So you need something.

[14:30] And so what I'm doing is taking the signal, processing it in a reversible fashion so you can try them to the signal from what I show.

[14:41] And then so where AI is playing is that now that you have this surface,

[14:48] how do you know that surface is similar to a surface for normal babies?

[14:54] So you have to find the right the appropriate surface comparison process to make sure that the between your reference populations the one that fits that new baby, you know.

[15:11] And let me tell you that I explored like 15 different surface distance measures.

[15:21] And I was not aware that there were so many. You know, you have Bata, Sharia,

[15:26] Kullback, Labor,

[15:29] Jensen, Shannon. You have many of those.

[15:32] You have Earth smoother distance. So you have the ones that come from the statistics and then you have the ones that come from more geographic, you know, terrain,

[15:42] Flying over the terrain and knowing where you are by matching the surface. And so the work which I qualify as symbolic artificial intelligence was to find what is the most appropriate surface distance measure that fitted my need and then applying it.

[16:00] And so I found one that gives me 98% in terms of result.

[16:06] And in fact the two that I made out still needed a hypothermia. I think I might be more correct than the doctor.

[16:15] And so it can be traced back to the signal.

[16:19] And then you have to be very careful about the data we were using.

[16:23] We had 600 EEGs at the beginning,

[16:28] but we selected the Doctors selected only 100. Because you have to take the EEGs that are taken within the first six hours of birth.

[16:39] Babies brain evolve very quickly after those six hours. So if you have EEGs that are a little bit after that they start having different sleep wake.

[16:50] So the shapes are no more the Batman ears shapes, you know.

[16:54] And so you have to be very strict. And you have to select also babies that had not another disease. You know, sometimes they swallow in the lungs some meconium of the birth squids.

[17:08] And so that might type you have multiple other.

[17:11] And you have to be very selective on what you use to be sure that you only have that pure disease, you know.

[17:19] Pamela Isom: And after of course is this research in use then today.

[17:25] Marc Fiammante: So then once we had patented then published the article was published in October. But in the meantime, since we had the patent, we tried contact the companies that build EEG devices.

[17:41] Say okay,

[17:43] what I'm doing can be embedded in your hardware. In fact I should say software because most of the Texas Instruments makes an EEG chip which is eight channel but and it costs only $250 so much.

[17:58] And the devices cost like 20,000 to $50,000. So most is what they put around, which is the software. And that we were not successful at getting those companies. I'm not going to name any company, but then I think they'd rather wait for innovations like that to be used in the market and then maybe acquire a company.

[18:23] So also if you look at all the maternities, you know, having a device that costs like 20 to $50,000 might not be affordable because you need people skill to use the device you need.

[18:39] So change our mind to say okay, we need to do something which is a lightweight approach because I only need four electrodes to detect which are the is happening in the front and the side and maybe a reference electrode on the top.

[18:57] I don't know if you recall beginning of 2025 there was this huge meeting about artificial intelligence in Paris.

[19:05] And so there were a kind of request for proposal for innovative projects. And I submitted my project and we got selected and even better, we got selected to be the one presented to our president.

[19:22] Well, we got one minute with him, but that was enough.

[19:29] Okay, and so what happened then is during that huge meeting we could meet with startup,

[19:37] not studios, but kind of fostering organization that are focusing on healthcare.

[19:43] They know all the, you know, the issue about regulations about getting things used in the market. And so we got one of them interested and,

[19:54] and so some we are currently in those being fostered by one of the. By this organization.

[20:03] We have a future CEO if we get and we get got 100k euros which is a little bit more than 100k k dollar funding to develop the headband and the portable device to be.

[20:16] So it's an internal funding from Paris Hospital,

[20:20] they're currently developing it. And to have something that can be used in all of the maternities.

[20:28] Pamela Isom: I think that is a wonderful example of how you are taking emerging tech and putting it to use to save lives.

[20:39] And we oftentimes hear about the adversarial issues associated with AI and maybe not the good. Now it feels like a real need. And I was listening to you talk about the precision,

[20:57] your focus on quality and precision because lives are at stake.

[21:01] And so I think one of the things that we should really emphasize here is that it is possible to get the precision higher because you do have infants lives at stake and at risk.

[21:15] And so you don't want to prevent the misdiagnosis, which is what I heard you say at first, or even just accelerate the diagnosis because that wasn't occurring,

[21:25] not necessarily misdiagnosis, but just accelerating the diagnosis is what I heard.

[21:30] And the focus on quality is so needed in this example. And so your symbolic AI approach,

[21:39] your use of symbolic AI and getting into that 98%, 100% is better, but 98% is just what we need. Right. We need better quality outcomes because of the risk associated with even that 2%.

[21:54] So I want to say thanks for doing what you're doing and I hope that the progress continues and that we start to see this integrated in our healthcare system today.

[22:10] Marc Fiammante: And we're not ruling out the doctor, it's just a diagnostic aid that we give the doctor.

[22:17] And since the approach that we have is fully traceable,

[22:21] if the doctor needs to see the signal, we keep the signal and we give him the clues of why we they made out that it was probably a children that needed hypothermia.

[22:34] So the doctor still does his clinical type of examination.

[22:40] There's a blood biomarker like lactates that see that some kind of injury happened,

[22:48] but it could be anywhere in the body. And then the EEG confirms that it was in the brain.

[22:53] So it's something,

[22:56] it's like measuring the blood pressure. You know, you could have a high blood pressure because you just have been running before seeing the doctor.

[23:05] Only the doctors can make out that it was the case. And it's not something.

[23:10] So we're not willing to replace the doctor. We're willing to give an aid to the doctor and give him some more clues about where to look and confirm that we,

[23:23] what we saw is, is the case.

[23:26] Pamela Isom: That's good, right? So that, so we always talk about how we want the use of AI and autonomous systems to augment, not replace. And so this is another prime example and it's the way that we should look at it is there to augment and to provide supplemental information, but relevant information so that the they can make high quality informed decisions.

[23:50] So it's a good thing. It's a very good thing.

[23:52] So tell me, what are some lessons learned that you have run into in your career journey and associated with the newborn health and the neurological monitoring?

[24:07] Are there any lessons learned?

[24:08] Marc Fiammante: I would say the data is the most important. You need good data whether you're doing machine learning or symbolic AI.

[24:18] If the data is not well categorized, defined and you don't have the good clinical information about may not be useful.

[24:33] The other thing is that you should not focus only on the mathematical aspect. And I see people, they take the images of the signal and just trying to do automated classification.

[24:47] You need to get back to the clinical reason.

[24:50] What did the doctor look at to make his opinion, what's important?

[24:57] The doctors, they have a kind of Implicit reasoning, you know, when they see data and you have to make out what they are basing their diagnosis on and try to get them to speak.

[25:09] So you get the people, the business people, if I generalize that,

[25:14] get information from the business people on why they're making the decisions, you know, get the elements and sometimes it might even be unconscious because there's so much training they have done their own deep learning, you know, and so it's part of their body that they tell you, oh, this one look,

[25:35] why tell me precisely where you know?

[25:39] Pamela Isom: Yeah,

[25:40] absolutely. That is really good. So you're bringing up what we often talk about is the human factor and don't just let the numbers rule, but make sure that the humans and the human considerations and perspectives are in that equation.

[25:54] So.

[25:55] And that makes a whole lot of sense. So what's a misconception then about newborn health and neurological monitoring that you wish more business leaders understood?

[26:06] Marc Fiammante: We need to stay humble.

[26:10] We're providing an aid to the doctors. We're not pretending to cure.

[26:16] We're giving something to help,

[26:19] which is AI is not replacing human.

[26:23] AI is demultiplying human.

[26:27] It's helping human.

[26:31] Beware of the biases. You know, when you think there's something to be diagnosed, you will look only at the elements that confirm that it's that diagnostic.

[26:46] So that's a bias, you know, confirmation bias.

[26:50] Beware of the confirmation bias.

[26:54] And there's also a lot of headwinds. You know, in research, people do not always think patentability of things they want to publish.

[27:08] Sometimes they forget that they have an innovation that could be patented because then it will serve as funding to their research.

[27:19] And so think you may have an innovation that's worth patenting. Patenting something costs a lot,

[27:29] you know, the fees.

[27:31] I think if you have a global patent,

[27:34] it's like 20k dollars, you know, per year that you have to pay to keep the patent running. You know,

[27:41] and so it will.

[27:44] You need at some time to get people to license where if you have patent to fund just the fact that you want to keep your patent.

[27:55] That's more on the legal side. There are variations between US patent laws and the European patent laws. And for example, in Europe you cannot patent a process that requires the patient to be always next to the device you're patenting.

[28:11] So you have to.

[28:13] So these are slight various patent autonomous. They would know about that and they tell you.

[28:18] And so you have to think about the way you're formulating your innovation to make sure that you stay not only on the Technical and business side of the usefulness of what you, the innovation.

[28:34] But on the legal side and more on,

[28:37] you want this to progress, be able to have your patent file and making revenue from it.

[28:46] Pamela Isom: So you're saying that we should remember this type of work is patentable.

[28:53] And then I heard you also say that there are global implications once you think about the patents, that there are global implications. So for instance, in,

[29:05] so in one country, for instance, after you create something,

[29:09] I believe is Europe,

[29:10] after you create something, you have one year to patent it. Is that correct?

[29:16] Marc Fiammante: Well,

[29:17] in US when you create something, if you have made it public, you still have one year for patenting. It's not the case in Europe or other parts of the world.

[29:28] So there are different rules applying.

[29:31] And basically you cannot publish an article in Europe before you have filed the patent,

[29:40] so you can publish your article the next day or even you cannot send the article for review before you have patented.

[29:49] So you're doing some research, you think you have an innovation, you have to go through the patent cycle and that takes six months. That was the case for me before the technical and business people think, yeah, there might be some value of that.

[30:05] The patent review boards, like there was in IBM,

[30:10] then the patent is file. And when you get the file date, you know the submission date to the patent offices,

[30:20] then only you can start sending articles to peer reviewed journals or even discuss it outside.

[30:27] So there are some restrictions like that.

[30:30] And even if you're in US if you want to have something global,

[30:35] if you have not respected the European routes, you cannot get the European patents. So in fact it's more restricted that applies here.

[30:44] Pamela Isom: I was listening to you talk about how you were able to sit next to doctors to understand how human experts diagnose with data.

[30:53] I don't think that we really understood to the extent to which that occurs.

[31:00] And we always talk about how we want to collect data,

[31:04] reliable data and trustworthy data.

[31:07] And I hear you describing an example of how you did that, right? So how you sat with the doctors, how you were solicited by a doctor to help you with this, but then how you sat with them to really understand and to understand their expertise when it comes to diagnosis.

[31:24] Not just telling them, here's what I'm finding, here are the patterns, but also understanding their expertise. I believe this is insights that we can use. But most of all, I'm hearing the applied perspectives of some of the governance that we've been talking about when it comes to AI.

[31:40] Right, so you've given good examples of how we are Actually applying this to help our newborns, which is a true blessing.

[31:50] So I don't know if there's anything else you want to say about that, but before I ask you, before we close out and I ask you to share advice or words of wisdom or a call to action,

[32:01] I just wanted to know, is there anything else you wanted to share before you address that topic?

[32:07] Marc Fiammante: Well,

[32:08] what I want to share is you're never too old to do something useful.

[32:17] Some people are innovating very young and I was 69, 68 when I filed my patents.

[32:25] There's still possible innovation when you have experience, you know, and it's.

[32:30] And I thank IBM because what I got from IBM is think out of the box. And using that symbolic AI and signal processing took me out of the box of the machine learning.

[32:47] That's so the technical realm.

[32:50] Pamela Isom: Right, so you're saying it took you out of the technical realm and into reality. Practicality.

[32:56] Marc Fiammante: Practicality, yeah.

[32:58] Pamela Isom: So would you be willing to share advice, words of wisdom and. Or a final call to action for the listeners?

[33:05] Marc Fiammante: When I was out of the university, the engineering school,

[33:10] I always wanted to do something useful for my fellow humans. So doing something for good is always possible.

[33:19] Use your experience.

[33:21] Might not be like I did because I had some,

[33:24] but in some other domains and hopefully my karma will be good.

[33:32] Pamela Isom: Okay, so do something for good and it's never too late.

[33:35] Marc Fiammante: Yeah.

[33:36] Pamela Isom: All right,

[33:38] well, thank you very much for taking the time to talk to me today.

[33:42] I really appreciate it. It's been a fascinating discussion and you are doing some wonderful work, which is why I'm so glad that you were able to join me today to talk about it.

[33:53] There is always that possibility to do something and it is never too late. And like, you really took off once you retired.

[34:01] Right. So. And also I like your strategy to get to the data and you had a technique to get to the real data that would help you make the right difference.

[34:13] So I sincerely appreciate it.