How AI transforms epilepsy diagnoses

The duration of a routine EEG is about 20 minutes, but more reliable diagnostics are possible with recordings that last up to 24 hours or even three days. Every recording is reviewed by a human expert who browses the data, which is quite time-consuming: for a 24-hour recording, this amounts to up to 3 hours of manual work. “The abnormal events we are looking for, the so-called interictal epileptiform discharges, are essentially the electrical signatures of epilepsy. They are typically quite short, around 100-300 milliseconds, and may occur only rarely. It is like looking for a needle in a haystack,” says Michel van Putten, neurologist and clinical neurophysiologist at Medisch Spectrum Twente. “You know there might be abnormal brain activity, but you don’t know when and where: You have to look through the whole recording.”

Together with Maryam Amir Haeri, associate professor at the University of Twente, Van Putten is working on two closely connected projects to change this reality: AI4EPI and ALICE. Their research uses artificial intelligence to assist experts in EEG diagnostics for epilepsy and make the process faster and more efficient.

The collaboration between Amir Haeri and Van Putten started with AI4EPI: using machine learning techniques to analyse the EEG recordings and detect segments that may contain abnormal brain activity, related to epilepsy. Instead of reviewing hours of data from start to finish, clinicians are guided directly to the most relevant moments. Van Putten: "That's where the expert comes in and checks whether that makes sense or not. This saves a lot of time: instead of reviewing data for three hours, it now only takes ten to fifteen minutes.”

During this project, Amir Haeri and Van Putten decided to take it a step further with a new project. “As we started working together, we came up with the idea of ALICE,” Amir Haeri adds. “An AI assistant in the hospital that explains what abnormalities are detected in the brain recording, what the source might be, and what the most likely diagnosis is.” Both systems are designed as digital assistants; the final judgement remains with the medical expert.

A first-generation version of the AI4EPI system is already being used in Medisch Spectrum Twente. Van Putten: “This first-generation is a spin-off from a PhD project by a student at the University of Twente. That really helped us get these Pioneers in Health Care (PIHC) projects started. It is a natural extension, and thanks to the PIHC vouchers, we were able to further develop these AI systems.”

The feedback from clinicians is already positive: it helps lab technicians and neurologists to review EEGs more efficiently, often reducing review time from hours to minutes. “AI4EPI is generally well appreciated. It saves a lot of time,” says Van Putten. “The next step is to see how well ALICE does her job in explaining the findings and supporting diagnostic decisions.”

For patients, the potential impact is significant. Faster review means faster diagnoses, less uncertainty, and earlier treatment decisions. “I have seen many patients after a first seizure, after which the diagnostic process begins. Is it epilepsy or not? I've seen thousands of EEGs, and I performed the traditional visual analysis myself,” says Van Putten. “We know that EEGs longer than 20 minutes are diagnostically better. An EEG of 24 hours or more would be ideal, but they are simply too time-consuming to review traditionally. With AI4EPI, this becomes possible.”

In the longer term, AI-assisted EEG analysis could improve access to diagnostics in regions where experienced specialists are scarce. “In the Netherlands, we have great healthcare,” Van Putten says. “But globally, there are many places where EEGs are recorded but cannot be reviewed properly because there is a lack of experts. Our software can help bridge that gap.”

The projects are successful thanks to the combination of expertise. Van Putten brings decades of clinical experience and access to large EEG datasets, collected in routine care. Amir Haeri contributes a strong background in machine learning and AI systems, with experience across multiple healthcare applications.

“We didn’t start from scratch, as we already had multiple similar projects running. That was an advantage for AI4EPI and ALICE. But data is usually the biggest problem in healthcare AI. In these projects, that problem was largely solved,” Amir Haeri says.

To train AI models, supervised learning methods require large amounts of labelled data. This means that experts must first mark normal and abnormal brain activity by hand. “That again would cost a lot of time,” Amir Haeri explains. “So for AI4EPI we deliberately explored unsupervised approaches, where the system can learn patterns and detect abnormal brain activity with minimal manual labelling. Michel has already gathered data for over ten years: these datasets are extensive, well-prepared, and clinically relevant. That makes serious machine learning possible.”

With the first-generation slowly being implemented in Dutch healthcare, the vision for the coming years is ambitious. Van Putten: “I hope that in the next three to five years, we will have algorithms that provide a clinically relevant and biologically plausible explanation of EEG data, tailored to the patient’s context. The human expert would oversee the process, but without the burden of reviewing the data by hand as we do now. Perhaps it could even be used to predict epileptic seizures. Even when you can't treat all seizures, patients would still be very happy if you can predict them.”

Amir Haeri sees the AI model being used for other diagnoses in the future. “Ultimately, I see the potential for a foundational EEG model that can be used to diagnose different sorts of diseases or conditions.” Van Putten adds: “For neurodegenerative disorders, psychiatric disorders or neurodevelopment disorders, to name a few.”

To achieve that, the team needs more funding, data, and collaborations with different hospitals and clinicians to work on this model. “What we really need is a dedicated PhD researcher who can work on this full-time,” says Van Putten. “That would allow us to take the next big steps.”

The Pioneers in Health Care (PIHC) Innovation Fund is a collaboration between the University of Twente (TechMed Centrum), Saxion University of Applied Sciences, and the hospitals MST, ZGT and Deventer Ziekenhuis. Each year, the fund allocates €600,000 to support ten innovative projects that use technology in smart ways to advance the healthcare of tomorrow. PIHC brings clinicians and researchers together to develop new technologies for improved patient care, or to explore new medical applications for existing technologies.

drs. M.M.J. van Hillegersberg - Hofmans (Martine) Communications advisor TechMed Centre

+31534896949

 m.vanhillegersberg@utwente.nl

Building: Spiegel Tuin

Building: Technohal 2134

Personal page