#407 - 🔵 [NEO CONFERENCE] - How can we build AI literacy among bedside clinicians (Dr. James Barry)

Hello friends 👋

How will artificial intelligence fundamentally change the way we chart, teach, and monitor patients in the NICU? Live from the NEO Conference, Ben and Daphna sit down with Dr. James Barry to explore the critical need for "AI literacy" among bedside clinicians. Dr. Barry draws parallels between driver's education and safe AI use, highlighting the hidden dangers of automation complacency with AI scribes. They also discuss the exciting potential of computer vision in respiratory monitoring and how the CONCERN early warning system is quantifying nursing intuition. Join us as we navigate the guardrails of neonatology's technological future.

Link to episode on youtube: https://youtu.be/ZNloOMP6nLQ

----

Short Bio: Dr. James Barry is an Associate Professor of Pediatrics-Neonatology at the University of Colorado School of Medicine. As a co-founder of NeoMIND-AI, a collaborative of neonatologists, Dr. Barry aims to leverage artificial intelligence (AI) to enhance the quality and precision of clinical care for neonates. NeoMind-AI's goal is to create a future where neonatal care is more personalized, efficient, and effective, and where every child has the best possible start in life. Dr. Barry's educational background includes an MD from Creighton University School of Medicine, an MBA from the University of Colorado Denver, and a BS from Montana State University. He completed his internship and residency in Pediatrics at the University of Michigan Program and his fellowship in Neonatal-Perinatal Medicine at the University of Colorado (University Hospital) Program. As the Medical Director of the University of Colorado Hospital NICU, Dr. Barry provides care for critically ill newborns and their families. His commitment to patient care and education has been recognized through awards such as the Daniel M Hall Teaching Award and Physician of the Year by the Colorado Society of Respiratory Care. Dr. Barry's research focuses on improving outcomes for high-risk infants and advancing neonatal care through innovative approaches.

----

The transcript of today's episode can be found below 👇

Ben Courchia MD (00:00.462)

Hello everybody, welcome back to the Incubator Podcast live at the NEO conference. We are joined by our good friend, Dr. Jim Barry. Welcome back to the show. I think the AI panel from Delphi went out on the podcast channel not too long ago, so you're back-to-back.

James Barry MD (00:09.688)

Thanks for having me.

Daphna Yasova Barbeau MD (00:20.59)

I think you've been to some conferences in between. You've been very busy.

James Barry MD (00:24.6)

I have been. I've been very fortunate to be able to go talk with different people. The last group was actually the Tennessee Hospital Association, and they have a surgery quality collaborative. I got to speak to some surgeons about AI and quality work.

Daphna Yasova Barbeau MD (00:43.0)

How did they take it? Did they listen?

James Barry MD (00:46.062)

I think that they did, but it was pretty interesting.

Ben Courchia MD (00:50.062)

Can I ask you to raise your microphone just a little bit? This way? Yeah, closer to your mouth. Perfect.

Daphna Yasova Barbeau MD (00:55.566)

We don't want to miss anything. Speaking of the AI panel at Delphi, people loved it. A favorite, a highlight of the conference. Definitely, people should check it out because it is available on the podcast.

James Barry MD (01:12.654)

Ben led us in such a great discussion. Just a professional.

Daphna Yasova Barbeau MD (01:17.57)

He's a pretty good moderator, all things considered.

Ben Courchia MD (01:22.318)

I give credit to the French radio for that.

James Barry MD (01:24.782)

I think you guys had a great mixture of people that are in the weeds doing it, and then really big-picture thinkers. Somebody like Ryan McAdams, who always has creative thoughts and is always thinking ahead. So I think you had a great mixture of people.

Ben Courchia MD (01:40.78)

I think so as well. And I think it's a great segue to what we were going to discuss. You were mentioning that you and Ryan are working on AI. You're working with NeoMind AI, a collaborative of neonatologists focusing on artificial intelligence. But you're going a little bit more meta, and you're trying to think: how do we improve AI literacy? How can we deliver information about AI to clinical specialists and neonatologists? And you said you're even working with him on a paper.

James Barry MD (02:09.09)

Yeah, we submitted it to the Journal of Continuing Healthcare Professional Education. They reached out to me and asked for me and Ryan to write a paper about how we should think about AI in professional education in medicine. And it was a good opportunity to really take some time to sit back and think, what should AI literacy be? Because if you Google it or ask ChatGPT, you get a variety of answers.

Ben Courchia MD (02:40.398)

And I'm sure the answer varies based on who you're talking about. What is AI literacy for a layperson? What is AI literacy for a physician?

James Barry MD (02:46.968)

Exactly. I think about it this way: we all drive cars, right? So we're auto-literate. We've taken driver's ed courses. We've actually had to pass a written test, a driving test. That's the way I think about AI literacy. You may not be a mechanic. We go and we start our car. We get our keys out, turn the ignition, or hit the button and the car starts.

Daphna Yasova Barbeau MD (03:07.106)

Definitely.

James Barry MD (03:15.266)

We don't really know how the car starts. Most of us don't. But we know that you put it in gear, push on the accelerator, and go. Then you know there's a bunch of laws that you have to follow. Stop signs, traffic signals, speed limits, guardrails. We've learned those along the way. I think we have to think about AI literacy like auto literacy. Right now, people know how to use ChatGPT or Gemini or Claude, but they don't really know how it works. And they certainly don't know the guardrails and limitations.

That's where I think we really need to focus with AI literacy, to make sure that people understand the strengths and weaknesses of model outputs. So a framework that we came up with, which we actually used from previous literature, is called TU-CA-PA. TU is Technical Understanding of what AI fundamentals are and the ethical principles that govern our use of AI. CA is Critical Appraisal of model output, but also scientific literature with AI—how to actually appraise the literature. And then PA is Practical Application. Using it, but doing it in a safe and effective manner. Those are the principles of how we as clinicians should think about AI literacy. We should think about it as a driver's test that we all need to pass before we really start using them full-fledged.

Ben Courchia MD (04:48.946)

Very interesting because you equate auto literacy and AI literacy, but in driver's ed there are stop signs and signage that we learned, but these don't exist yet in AI. So in one way you're also identifying some of the gaps that are still missing in the AI world.

James Barry MD (05:08.556)

Yeah, and I don't think we know yet, but we need people like you and all of our clinician colleagues to be involved in making sure that we're evaluating these tools. One of the big things that has come out over the past year is AI scribes. Everybody's using them. But if you talk with a lot of healthcare systems, and I've asked a lot of leaders, "How are you evaluating these AI scribes?" Most of the time the answer is, "Well, clinicians like it, and patients like it." And I think those are two very important things. Billing is probably improved; if you have better documentation, your billing is going to improve.

But then I start asking questions about how it works in terms of dialect bias. What are you guys seeing with that? Is it working as well for your Somali patients as your French-speaking Americans?

Daphna Yasova Barbeau MD (05:59.17)

I saw that on The Pit. You don't watch The Pit?

Ben Courchia MD (06:02.146)

No.

James Barry MD (06:03.177)

I know what you're...

Daphna Yasova Barbeau MD (06:04.046)

They were having some difficulty with their dictation.

James Barry MD (06:11.79)

Well, it's true. And I don't think we're capturing that like we should. And then what's the omission error rate? What's the addition error rate? My wife actually uses these AI scribes in her primary care clinic. Before she started using them about six months ago, I told her, "I want you to keep track of how often it's adding stuff to your notes that was never in the conversation." She thinks it happens two to three out of ten times. Wow. You really have to stay on top of looking at the transcription or the output from these models and making sure it's correct.

Daphna Yasova Barbeau MD (06:42.924)

That's a lot.

Ben Courchia MD (06:52.288)

It's interesting because we use a lot of AI with the podcast when it comes to transcription. These audios are auto-transcribed, and then there's a review process that we do with AI. Exactly as you said, when it comes to transcription, AI will take a lot of liberties. I have to do a lot of vetting and say, "You're not allowed to add more things." And Claude or Gemini will apologize. But you're right, and we didn't suspect that. Having used it outside of the clinical setting, it's quite surprising as to the liberties that AI will take.

Daphna Yasova Barbeau MD (07:26.35)

Yeah, and then it really scares you about what it could add or omit from the clinical setting.

James Barry MD (07:31.426)

Well, because automation complacency is going to be a big issue. When us humans start seeing something all the time and it seems like it's working well most of the time, we're just going to automatically click sign, sign, sign. Exactly. We're not going to read every single word. But if you look at our baseline error rate, I don't know if you guys have ever come on service after one of your colleagues and gone through all their notes to see a baby who was extubated two weeks ago...

Ben Courchia MD (07:43.864)

We're responsible for every word.

Daphna Yasova Barbeau MD (07:59.81)

Go on.

James Barry MD (08:00.558)

The physical exam still says intubated, absolutely. And then somewhere in the plan, it still says intubated.

Daphna Yasova Barbeau MD (08:06.894)

Even on a day where I said I knew I edited this, I still couldn't catch all the errors.

James Barry MD (08:12.802)

Well, part of it is because it's in three different spots and you catch it in two of them. Then you find it later on. So we're probably no better than these models, but we are holding them to a higher standard, which I think is necessary. We still have to understand the error rate, and I don't think that we're close to detecting that.

Daphna Yasova Barbeau MD (08:33.806)

When will the virtual masterclass on AI literacy be available?

James Barry MD (08:38.795)

You know, we're developing it. I've been a little busy the last couple of months, but I have the template. In our group, NeoMind AI, we have a subcommittee that's going to be focused on education, so we'll be working on that. I think there are probably others out there now because everybody's starting to get involved. We want to make sure that it's really specific to neonatology and pediatrics.

James Barry MD (09:08.256)

Because I think in this AI wave, pediatrics and newborn medicine are kind of being forgotten. So we want to make sure we're able to focus on that and support our colleagues.

Daphna Yasova Barbeau MD (09:19.342)

I love that. You were given a very difficult topic to speak about. Artificial intelligence, which you know everything about, but specifically neonatal respiratory monitoring. Promise or hype?

James Barry MD (09:33.878)

Well, I can just tell you in a short period of time, it's not ready for prime time. Yeah, I would say it's a promise. There are promising developments, and it's interesting. But I think the interesting parts will come from other areas, like computer vision. Physiological monitoring data and understanding patterns that exist within there haven't really been applied yet to respiratory monitoring, but I think it's evolving in potential.

Ben Courchia MD (09:39.221)

It's promise.

James Barry MD (10:00.734)

Not just the data from the EHR, your pulse oximetry data, your FiO2, your mean airway pressure. I think we'll go beyond that with patterns that we humans don't see at the bedside. That's what I'm really excited about. The technology has evolved with computer vision in particular, and that's something we're going to see in our professional era.

Ben Courchia MD (10:32.398)

This is truly very exciting. There are lots of possibilities for sure.

James Barry MD (10:39.726)

There are. I think the one area I'm skeptical will move quickly, but where a lot of focus is, is predictive analytics. When you're predicting things, it's really going to depend on your own unit. There are so many different nuances to your unit that a predictive model might work well in New Jersey or Hawaii, but it's not going to work in Colorado. And then you have to have an infrastructure to develop or fine-tune that model so it will work in your unit. I don't think most healthcare systems are going to have that capability.

Ben Courchia MD (11:17.674)

Yeah, and I don't know if we even need predictive algorithms, to be honest. I think we just need reminder algorithms, as we've talked about. If you could get to the bedside and the algorithm can remind you of the risk stratification of your patients, saying, "Hey, this patient was advanced on feeds recently and has had some issues in the past," and bring you the information—then you do the prediction. Even that alone would be very beneficial.

James Barry MD (11:41.502)

I think I've told you guys the publication I love the most over the past couple of years was by nurse scientists out of Columbia. They created something called the CONCERN early warning system, where they created a model that quantified and qualified nursing intuition or nursing surveillance. We all know we have a 20-year experienced nurse who comes, taps you on the shoulder, and says, "Jim, there's something wrong with this baby. I can't put my finger on it, but there's something wrong with this kid." I go running to the bedside.

Daphna Yasova Barbeau MD (12:18.082)

Because you know that nurse did not want you at the bedside otherwise.

James Barry MD (12:21.078)

Exactly. So they created this model that qualifies and quantifies that, and they evaluated it in four hospitals on the East Coast. It was on the general floor, and they showed decreased mortality rates using this because it flagged the patient as high risk, medium risk, or low risk. What was cool about it is they used just common things. In this world of AI, we're always thinking about fancy things. What they used was a nurse who went and saw her patient more often than they normally would, at unusual times, missing a dose of oral medication because maybe the person was getting sick, and then some words of concern in their documentation. Just those simple things they were able to qualify and quantify. If you think about it, that's just brilliant.

Daphna Yasova Barbeau MD (13:15.054)

We're going to move forward.

James Barry MD (13:18.848)

Not everybody's going to be Nurse Val, where I'm going to go running to the bedside with her. Instead, it'll be somebody who's been there for 10 years and maybe people dismiss that nurse. But if you have something that has a red light, yellow light, green light that pops up on your screen, it's going to change your behavior. I think it's just brilliant.

Daphna Yasova Barbeau MD (13:39.477)

I wanted to make sure people knew it was episode 395, "How's Artificial Intelligence Transforming Neonatal Care?" so they can hear us live from Delphi. My last question for you: you just posted about questions that you're getting at conferences. Was there a good audience or panel question that you think we should all hear the answer to?

James Barry MD (14:02.83)

Well, the main thing—and I posted that on LinkedIn—was what should we be teaching our trainees about AI? What do you guys think? What should we be teaching them?

Daphna Yasova Barbeau MD (14:13.694)

Well right now they're getting nothing. But I'm recognizing, for example, that even my elementary school-aged child is getting lessons on AI. Both when to use it, when not to use it, and how to evaluate it. Really recognizing that the students are going to use it, so we should teach them to do it right. If they're not using it, they may fall behind some of their classmates. I've been really impressed by how they're integrating it into elementary school education. Certainly, our residents, trainees, and fellows should learn it.

Ben Courchia MD (14:50.21)

We should tell trainees to absolutely delve into AI. There are obviously the three pillars of medicine: clinical care, research, and education. I think the use of AI in these three domains is going to differ a little bit. Our job is to let them run wild with AI, but our responsibility is those guardrails. Take a picture of an X-ray and play with AI and see what it gives you. De-identify, and please make sure that you are safely using it. Because at the end of the day, I consider myself to be AI-savvy, but I still don't know everything. I will be taught things by the trainees, but I just need to make sure that we don't slip and divulge patient information, and that we are careful from an ethics standpoint. It's very exciting, I believe.

James Barry MD (15:44.61)

Yeah, so what you just said about your daughter, that's auto-literacy. That's AI literacy, right? So they're in driver's ed for AI.

Daphna Yasova Barbeau MD (15:54.047)

Yeah, and I think if they can teach those kids, then surely neonatologists can learn as well.

James Barry MD (16:00.526)

But it's been funny. I think many of my colleagues are still somewhat resistant to it. When I talk with them about it, sometimes they roll their eyes or yawn. That's fine, but this is a tool. It is getting bigger and more complex. If you at least just started wrapping your arms around it a little bit, you'd be so much better off. Then you can teach your fellows and residents the right way to use it and not use it.

Ben Courchia MD (16:24.438)

Unfortunately for the slow adopters, the option to let this be a fad is no longer there. It's happening, and you can either decide not to use it and just practice the old way, but eventually, it's going to compound and you're going to be the one who's left behind trying to catch up.

Daphna Yasova Barbeau MD (16:53.678)

It's also interesting. We read about and go to workshops about how we can incrementally change one thing like BPD or incrementally change one thing like ROP, but this is an intervention for every single patient. When you think about it that way, we have an ethical obligation to use it for good, as they say.

Ben Courchia MD (17:25.314)

Jim, thank you so much for dropping by the booth.

James Barry MD (17:26.956)

Thank you for having me. I'm such a big fan of you guys. Keep up the great work that you're doing. What you guys have created and the service that you're providing for neonatology is just fantastic. Keep up the great work. Thank you.

Daphna Yasova Barbeau MD (17:29.185)

Always. Thank you.