Ep. 202: Clinical neurology Meets AI: Are we prepared?

00:00:00: Welcome to EANcast, your weekly source for education research and updates from the European Academy of Neurology.

00:00:15: Hello!

00:00:15: And welcome to EENcast.

00:00:18: my name is Raphael Brum.

00:00:18: I'm part of the EN's communication committee and the AI task force that was newly founded.

00:00:26: Todays topic is clinical neurology meets AI.

00:00:30: are we prepared?

00:00:32: To answer this question i have The wonderful Roland Vist with me today is of course a frequent flyer at EANcast and other EAN media productions.

00:00:45: But

00:00:46: for today's episode, his most important I would say role... ...is that he is the Vice Director of the Department of Digital Medicine at Bären where has many other roles particularly in neuroimaging and radiology.

00:01:01: Roland we're gonna talk about how AI tools are being integrated into clinical practice.

00:01:08: as We speak I think we all feel that this is happening at a very rapid pace.

00:01:16: If you think back about the GPT revolution, then has happened.

00:01:19: I'm going to say three years I don't know maybe it was four feels like he could have been ten.

00:01:23: and How now?

00:01:24: only slowly after the initial excitement of being able to chat to these computers?

00:01:31: influencing our daily life speed in or outside of the hospital.

00:01:36: So my first question to you is, well are we prepared?

00:01:40: And if so what areas that we should prepare for most?

00:01:45: Yeah, so Raphael thank you for this kind introduction.

00:01:48: I'm really happy to be with you here today and you have already addressed i think in a limited way prepared because if you review let's say the production of artificial intelligence tools that have received at this current stage.

00:02:11: The marketing level we count approximately fifteen hundred of them if he were to for example be the FDA boards and it's a general is a journal that where that needs us to me, comes to neurology, I would say then we will identify still as one of the main use cases.

00:02:37: The employment of analytics and clinical decision support in neuroimaging.

00:02:43: this is definitely One Of The One tools that are mostly advanced have been mostly advanced than last time.

00:02:51: We also have domain of neurophysiology.

00:02:55: automated signal analysis has made a lot of advances.

00:03:00: We have the huge domain of automation in clinical workflows and something that is coming, and it's getting more and more interested.

00:03:09: this automated writing off reports studies also what called ambient listening so to means tools that record the exchange between patients and clinicians neurologists and try to sort out the most important content of these dialogues.

00:03:28: This happens, for example not only in the hospital environment.

00:03:32: there are new things to come that also tried to work in a pre-hospital phase.

00:03:37: For example if stroke patient is admitted to a Hospital so good to retrieve The most important facts and informations and sent them to hospitals before the patient reaches the emergency room.

00:03:51: So they're really lots.

00:03:52: And since there are so many applications, it's rather difficult to navigate through this jungle.

00:03:58: as I said before.

00:04:01: Yeah i think you touch on a very interesting point which is the FDA approval.

00:04:07: There is also an EMA approval way that not entirely mandatory right now.

00:04:12: It probably will be at some point in the future.

00:04:16: As a clinician in the field of neurology let say general neurology.

00:04:19: What is your recommendation?

00:04:21: How should try and educate myself about what are actually useful tools right now.

00:04:27: Is there like, another sign that I can look for some sort of stamp approval from someone organization or how do you approach it?

00:04:36: So we have different possibilities to review data.

00:04:42: Of course There is the home page of the FDA where let's see screen All the different technologies that have been cleared in the United States.

00:04:53: as you mentioned this is not something, let's say allowing a legal clearance for Europe but to get information about how these systems has been developed.

00:05:10: that have been conducted and it led to the clearance, you also will be informed about the fast track versus let's say in-depth clearance processes.

00:05:20: There are at least my domain of neuroimaging.

00:05:24: there were different other possibilities.

00:05:26: There're private companies in Europe that offer similar home pages.

00:05:31: where can go through them?

00:05:33: You can see what is level of clearance or purpose for these tools which kind prerequisitions are needed to run these tools in the clinical services.

00:05:44: And most importantly, they also provide a list on literature and evidence that is already there for this tool.

00:05:53: There's similar tool has been built by the American College of Radiology which allows you to retrieve different information but at end it not structured by now, so we have this gap.

00:06:09: We don't have structured recommendations.

00:06:12: which kind of tools can be used?

00:06:13: I think the most important question to answer and first one is always do you have a medical product that is certified?

00:06:23: if it's a medical products or environment where these products are running either as an on-prem solution within hospital firewalls as a cloud-based solution, it needs to have a certification of the medical product.

00:06:40: Yeah I think that's very good point.

00:06:41: if you don't see that CE or FDA certificates better not use for sensitive medical data

00:06:49: right.

00:06:49: and this brings me one important points.

00:06:51: so sensitive data on chat GPT for example to write a report or even just to condense information because this is really something that's dangerous and not within the legal frameworks of any country.

00:07:07: Yeah, I think.

00:07:07: yeah i think it's very good point.

00:07:09: you racist again if he put some thing into chat gpt explicitly give them The allowance and the right to train on this data so this will be saved.

00:07:18: So it's not like googling something.

00:07:20: you actually put these data into.

00:07:22: if he puts your own personal information in, or that has had their own risk but if you put your patients' informations then does violate privacy laws?

00:07:31: You mentioned another thing I'm really curious about which is EEG or neurophysiology.

00:07:36: pattern recognition.

00:07:38: It's something from my point of view has been surprisingly slow when we look at.

00:07:44: It's not that much signal, I guess.

00:07:47: But for some reason it seems to be eluding sort of the pattern recognition models?

00:07:53: What is your take on?

00:07:54: how is the future looking for automated EEG?

00:07:59: To be honest i have not reviewed most recent advances in that field but from me its quite straightforward because you can train systems relatively easy on signal recognition.

00:08:13: So therefore, I think it makes a lot of sense to have this kind.

00:08:18: specifically if you have long-term EG recordings for example or sleep recordings these systems can be employed.

00:08:25: i don't honestly Have current let's say information about the last most recent information in literature But from the use case, I would say this is a very nice one.

00:08:37: And it's kind of in some instances quite similar to what we do in ure imaging because its pattern recognition and classification.

00:08:47: Yeah i imagine if you look through as you said like at twenty four hours recording something that just filters out The areas where there's nothing happening and just directs you To those that need attention?

00:09:03: Maybe just to add something today because there is a similar principle that's currently emerging in the field of neuroimaging.

00:09:09: That's the anomaly detection, so instead of training The data on pathologies you train the data under normal or under normative appearance Of certain even signals are images and then these kind of anomaly vectors Detectors, they spot the differences to normal data.

00:09:31: This gives one opportunity to train this data on really large datasets because it's much easier to get let say a normal data into training system and then it will spot any anomaly without giving you classification And this allows you to review this anomaly again to justify or determine if this is relevant finding or not.

00:09:54: Yeah that makes so much sense right?

00:09:56: Because I think we're always in this space, also having to confront a little bit the fear of providers that eventually they'll be replaced.

00:10:06: And I think these goes much more into direction of co-pilot analogy where there is something flying along with you and it's gonna highlight or steer maybe at some point can take over wheel for couple hours but eventually its you who isn't driving seat and making decisions.

00:10:22: You mentioned ambient listening before Something i've just seen anecdotally pop up in advertisements much more.

00:10:29: I know that the Americans used it quite a bit, they've also already had the profession of the medical scribe something that i'm not aware as much in Europe.

00:10:38: is this something that's happening?

00:10:40: In your practice?

00:10:41: do you see that coming soon to other fields?

00:10:44: well at the moment it's at our level of being fully clinically implemented but there are lots research projects going on.

00:10:52: I think it's a huge difference where you want to use.

00:10:57: If we wanted an emergency room, for example... Where there is a lot of noise and distraction in between people talking with each other or maybe even patients that can either hardly speak or speaking in the different language then still something that isn't completely solved.

00:11:15: There are lots of artefacts and errors within these systems but if you do this a shield that are protected environment.

00:11:23: I'm just thinking about more lengthy communications, for example in the domain of psychiatry or in patients with chronic disorders that you see for let's say the fifth to tenth time within your reward and then we're discussing them there.

00:11:41: progression from certain disorders than this works already pretty well so it is rather dependent indeed On the use case, I would say it's pretty much advanced now in the field of chronic disorders.

00:11:55: And we have more extensive discussions with the patients whereas there is still some need for improvement in the actual rooms of an emergency environment.

00:12:08: Yeah that makes sense!

00:12:10: i come from the field cognitive neurology.

00:12:11: so talk to patients with dementia a lot and think they can work wonderfully.

00:12:16: And it's also, I think one of the good examples where if we use this technology in the right way.

00:12:23: It will actually free up time and headspace for us to go back to the human-human interaction.

00:12:30: because right now what is happening?

00:12:32: Just thinking you're going your physicians... What's gonna happen is that they'll be on a keyboard typing.

00:12:38: They are looking at screens or typing away while talking with them.

00:12:42: so something can just pick up your conversation and make something meaningful out of this, it would free up a lot of space.

00:12:48: I think that will be beneficial to both

00:12:50: parties.".

00:13:13: Well, like for example being empathic listening let's say to subtle information that is provided by the patient.

00:13:20: Whereas they're purely analytical process can easily be done by algorithms and then two work together as a team at the end of course in best-of-all cases than human physician has additional information make decisions.

00:13:38: but I think we should invest much more on this vision of having men machine interfaces instead of these discussions about replacement.

00:13:49: Yeah, no I fully agree and i think... Of course there is this pull towards automating things also for productivity gains.

00:14:01: but I see the most important use cases right now as you already mentioned in things that are extremely tedious don't require a lot.

00:14:10: What we actually good at right it's not the physical examination is gonna be automated.

00:14:14: It's not talking to a patient and having like, I mean there are lots of research into that when people have got feeling it is actually them taking in so much information on then having a hunch based off something that's tangible.

00:14:27: i'm really hoping they can take away all these tasks in my daily practice that extremely repetitive and to be frank just bit stupid.

00:14:38: I want to build on that.

00:14:39: for the next question, because we're talking about the matter of preparedness today.

00:14:43: How can say i'm in a practice or say i am in hospital and... ...I want to be ready for these tools when they come?

00:14:53: Is there anything I could do with my data right now?

00:14:57: Or is it any way I can prepare how I collect data?

00:15:00: so far as being ready for AI?

00:15:05: It depends a lot on the environment where you're working.

00:15:10: So I think many hospitals are currently trying to solve this problem by having kind of, yeah let's say data... so i hate these words but they call it data warehouses.

00:15:23: You have structured environment where you can save this data.

00:15:26: Then if you want to retrieve these data, make a formal application for the Data Protection Officer then it's clear how we use that data.

00:15:35: but in general I think everything is standardised and somehow digitalized should be Digitalized.

00:15:43: so that means the clinic clinical hospitals need to rethink about their formats.

00:15:49: one easy way or not.

00:15:52: Easy it's an expensive way, but maybe some reasonable ways of course move forward.

00:15:57: electronic health records this is currently done in many.

00:16:02: Centers in Europe where you have some kind of pre structured data.

00:16:07: cause you should try wherever its possible implementations of structured reports, and if you don't have this kind of structured report then with tools like large language models.

00:16:22: You can try to condense the information into structure it in a secondary way.

00:16:28: You need some let's say storage protected storage where we could also save the laboratory data where they can save their imaging data.

00:16:41: But I think the key element there is a strategic thinking.

00:16:45: It's the strategic thinking about implementing a protected research environment that fulfills the criteria, so it means they cannot be entered and can not be hijacked from external sites... ...it needs to have governance in place.

00:17:01: who are allowed use this kind of data?

00:17:04: then you need access.

00:17:06: This is long-term project!

00:17:08: And then it will not be solved within half a year.

00:17:11: So for now, I would really if you want to prepare your data and do have certain use case in mind Then the only advice i can give is try to structure It as good as possible from the very beginning.

00:17:26: Right You mentioned LLM's obviously talking about AI.

00:17:34: When we think about commercial AI applications right now, they all happen in the cloud.

00:17:38: Right?

00:17:38: So I go to any of the big names and I will enter my whatever it is query or prompt And then it would calculate somewhere else.

00:17:46: obviously For healthcare data We would like to do this more local.

00:17:52: Do you think that This Is something which should also prepare for on a hardware site when need To buy lots of GPUs right Now?

00:18:00: Yes i think this is The ongoing discussion.

00:18:04: What are the arguments for having on-premise solutions versus cloud based?

00:18:10: and I think there it's a very differentiated way How we should try to analyze this.

00:18:17: because of course if It comes to safety then the best solution will always be To have on prem Solutions.

00:18:24: This comes

00:18:25: with

00:18:26: certain shortcomings Because you need to build infrastructure which is still doable, so you need to have your own service within the hospital environment.

00:18:47: But if there are perfect solutions for starting with AI implementations?

00:18:59: options and we have been following that now for several years, is to go into a different direction.

00:19:05: To have platforms with integrated AI solutions that are working cloud-based.

00:19:12: then of course you need to fulfill national requirements regulatory requirements.

00:19:19: That means you need put your server in safe areas.

00:19:23: You cannot put it everywhere.

00:19:24: In the best case To have a process for de-identification, you need to build security concept that also describes what would happen if there is data breach or an attack on the system.

00:19:44: From an economical perspective let's say not in many cases but with arrangement of the company and paper use then usually the cloud solutions are the ones that are more economically implementable.

00:20:02: Because I think, uh... The costs for the service and specifically the cost of personal should not be underestimated if you want to build a really high throughput system like For example in national stroke analytics System tries to find for example large vessel occlusion And where have relatively let's say limited pre-test probability.

00:20:28: You can have the stroke mimics that are running through this system also, now then you will end up in a high number of data to be processed by these systems and then cloud based solutions may be rather expensive.

00:20:43: there an on-prem solution is installed within hospital may save money so you cannot say one's better than other.

00:20:51: You might either consider the safety aspects, which are definitely on the side of one-prem solutions where in a low to midsize range for data computing.

00:21:03: The cloud solution is more economically implementable

00:21:07: right?

00:21:07: So it really depends on your use case.

00:21:10: I think we've covered a lot of ground last couple minutes.

00:21:15: i think i want to quickly emphasize the main points you made and then ask if you wanna add something.

00:21:20: so I think we can see that AI and clinical neurology is an area that's going to intersect even further in the future.

00:21:29: We'll see more applications, you mentioned that in imaging were maybe shifting towards flagging of abnormalities but also mention a large vessel occlusion which flags specific pathology?

00:21:48: Taking notes as you speak is something that's actively being developed and I think very useful right now if your talk a lot to the patient.

00:21:56: so things like I actually think dementia thing outpatient visits.

00:22:01: That are frequently around the same topic but might need some more work in busy environments talked about pattern recognition in neurophysiology which making i think also great improvements.

00:22:15: we talked about that.

00:22:15: if u wanna Prepare if you want to be one of the early adopters.

00:22:19: The most important thing You can do right now is just structure your data and store it locally in an accessible way And then for the future, It might depend on Your use case.

00:22:31: If you wanna go safety Go local.

00:22:34: But if you have something that's not as easily scalable Then maybe a cloud-based solution That also obviously Is within all regulatory premises Might be Something thats more economical For you.

00:22:47: But if I hear you correctly, You're also quite convinced that the future Will see AI being much more tightly integrated into our clinical environment.

00:22:58: Yeah.

00:22:58: I think clearly i have no doubts That this is a next and very disruptive step but it's on us physicians to structure It To organize it but also build regulations for use of artificial intelligence.

00:23:15: and this brings me maybe to my final point.

00:23:17: we should invest now into the education.

00:23:22: Of our younger colleagues, you need to have some kind of baseline as he mentioned when it has some kinds of baseline training how is a high working water use cases?

00:23:31: what are also there?

00:23:32: let's say that they're legal environmental differential and ethical aspects that I needed to run this safely in hospital environments, so we need to think much more now about structured educations even about certifications that are recognizable and open new career paths for young physicians.

00:23:56: And of course you should take the opportunity not just show up as customers or companies but actively seek this public-private partnerships and to open networks together with developers within the streets, with universities in order to shape developments of new tools according to needs that we as neurologists or physicians have.

00:24:24: And really make sure these tools are concepted fulfill requirements for clinical practice.

00:24:35: That's a great.

00:24:36: last point.

00:24:37: want to emphasize that people should look to EAN for exactly that.

00:24:44: I think we're actively developing, thinking about these sorts of educational material but also certificates in the future and this is what the AI task force has been set up to do And i agree it needs something where were an essential part of driving forces.

00:25:06: So with that Roland, I thank you very much for joining me on this cast today.

00:25:11: Wish you a very pleasant day!

00:25:14: May not many helicopters fly over you and i hope to see you soon.

00:25:19: Thank You for the moderation.

00:25:20: it was a pleasure to be here.

00:25:29: This has been EANcast Weekly Neurology.

00:25:32: Thanks For Listening.

00:25:33: Be sure To Follow Us On Apple Podcasts Spotify Or Your Preferred Podcatcher For Weekly Updates From The European Academy of Neurology.

00:25:41: You can also listen to this and all of our previous episodes on the EAN campus, E&Cast weekly neurology is your unbiased and independent source for educational and research-related neurological content.

00:26:13: Although all the contents are provided by experts in their field, it should not be considered official medical advice!