Leveraging AI To Optimize Epilepsy Patient Care

Pierre D'Haese (00:01):

My name is Pierre D'Haese. I'm a researcher and entrepreneur in the United States for about the last 20 years in electrical engineering and neuroscience. My role here at AES has been presenting a new study into the clinical flow about how AI and LLM models can be leveraged in the real clinical efficacy of treating epilepsy patients better.

(00:25):

The study is really about understanding how data can be leveraged for epilepsy care. And we know epilepsy is a very difficult disease to manage. There is a lot of notes in the EHR, and the quality of these notes are really with a lot of narrative and nuances that, over the years, have been tried to be simplified and put in various buckets to create registries of data and these kind of things, but the content often doesn’t present the clinician and others with the right opportunities to take actions. We've all heard about LLM models and AI. We thought that it could be an interesting thing to be able to see if the AI, and AI agents really, could extract from the narrative within the clinical flow some details that will be meaningful, an understanding of what cases would be missing, some opportunities to do MRIs or EEG tests or other assessments. What is one of the things or the patients that could become more surfaced in being candidate for neurosurgery, usually drug-resistant epilepsy or other things along the same lines and following safety.

(01:34):

And one of the things that the AI has done in the context of standard of care is understanding that some patients would benefit from additional assessments like an additional MRI or another EG or neuropsychiatric evaluation, for instance. And this notification can be done at multiple steps during the longitudinal care of the patient. And understanding when that moment is and when the patient would qualify for one and then when the insurance will be actually be able to pay for it and reimburse for it, is something that's not easy. And detecting these times make those numbers (of notifications) actually go high. So you see a lot of patients that could benefit from an MRI sooner, for example. And so this is where we found the solution about being able to bring together all these different contexts.

(02:24):

At the time of care, when physicians or the attending is seeing the patients would actually change their capacity to say, "This is the right moment. The insurance is aligned. Everything is aligned. Let's do it." Everybody's trying to find evidence that AI LLMs models can actually even impact our day-to-day life. It was kind of interesting to know that getting those AI models working within the clinical flow, where there's all these privacy issues, is difficult, definitely not the first place we would think that this would have an impact, but this study represents the first time AI LLM models have been running for about a year now in a real clinical flow. And the impact is thinking that now that these models cannot think for us, but they have a level of comprehension that is at the level of a junior physician; the system can put all these different data and different contexts together to make a real impact for patients.

(03:20):

Validation and trust is a major thing, right? Because AI LLM seems to be a black box for many of us. So understanding that technology can help provide the right answer and not just flagging patients the wrong way, creating false negatives, as we say, is something that needs to be done. And we are doing it already. There is a study on AI and LLM that's about to come that shows not only that it is feasible, but it is more than just doing a ChatGPT on EHR notes. The right teams are getting together. So physician and engineers can work together to provide the right context, and then the LLMs can make actually a difference. Otherwise, they're just going to be prone to errors. So validating this with other institutions, but also within different frameworks and contexts, so we can actually work through the entire chain of value for the epilepsy care centers.

(04:11):

We definitely find a desire from the physician community to understand what AI can do, and many  physicians are trying themselves. A published paper or research is great, but they often don’t produce changes in the real world that you can see in clinical flow. So creating these teams will be important. We see this in many centers that are trying to emulate more of these connections between engineers and experts in the medical field, and we'll see the ones that are moving forward. But the adoption seems to be pretty easy because people are eager to see how AI can become as really an augmented capability for a team, not replacing physicians, but really helping them orchestrating the flow of patients. They've run on more than now 4,000 patients on every visit, on every note, and this is just the tip of the iceberg.