Epilepsy is a chronic noncommunicable disease of the brain affecting 50 million people and making it one of the most common neurological diseases globally, according to WHO.
With proper diagnoses and treatment, 70 percent of people living with epilepsy could live seizure free, making access to appropriate care and detection of upmost importance.
Seizures can create challenges for the independence and day-to-day lives of people living with epilepsy. They can also lead to driving collisions, with 0.2 percent of traffic accidents linked to a form of seizure. A team at University of Sydney, led by Dr. Omid Kavehei, set out to answer an important question, “Can we improve the accuracy of seizure detection in epilepsy and can we predict a future seizure?”
According to the law in New South Walkes, Australia – home to the University of Sydney, people with epilepsy must be seizure free for at least 12 months to drive. This seizure free declaration is often based on a rough conversation between a patient and their clinician, with the clinician certifying they have been seizure free for a set period of time and patient reports. Given it’s not uncommon for patients to not remember seizures, or not have a family member or caretaker around with them, the certification process can lead to inaccurate outcomes. The researchers saw an opportunity to challenge the status quo and help clinicians make data-driven decisions.
Despite a great deal of research and development in the last few decades, statistics of epilepsy remain almost unchanged. For example, still about 35 percent of diagnosed or confirmed epilepsy patients are going to have a long and difficult journey towards falling into the epilepsy treatment gap, where there are no available nor suitable treatment for them. “The percentage has not changed over time, it’s the same in 2022 as it was in 1990s. The current methods are clearly not working if we have not been able to improve that statistic despite some major understandings about the underlying disease and the brain,” Omid shares.
The aim became predicting seizures with a good level of accuracy, leading to further investigations in the field. Should a seizure be predicted before it occurs? Would the administration of an anti-epileptic drug, that otherwise might not have been helpful, trigger a different outcome? A whole sphere of ideas was shaping up as they undertook the research. Their project was awarded an AI for Accessibility grant to investigate seizure prevention and prediction with the help of AI.
The first hurdle was access to data. As Omid explains, “We need democratic access to data. No data, no research.” By partnering with the Royal Prince Alfred Hospital, the researchers were able to access this data, with proper patient consent. “Data should belong to the patients and we should be considered as custodians of it. Equally, research should be facilitated and most patients have no problem with that concept. If we limit access to data inequitably, we limit the opportunity for the next breakthroughs, which can often come from an unexpected source and person. This is doubly important if the data in question was produced using a form of publicly funded research or development. Exclusivity of that data beyond a certain point is neither understandable nor fair to the patients who volunteer so that data becomes available.”
For successful seizure prediction, the team at University of Sydney needed to extract specific biomarkers from this data, specifically detecting biomarkers that indicate abnormalities in the brain activity. The rate of those incidents and density would represent the onset of a seizure. Existing documentation shows the brain senses when a seizure is about to happen, and it works in order to stop it. From a practical standpoint, clinicians and researchers should be able to seek repetitive and consistent patterns of data in the brain before a seizure. In developing their AI model, they partnered with the head of neurology at the Royal Prince Alfred Hospital, as well as additional neurologists and epileptologists to test the performance of the model and understand if it led to any conclusive results.
In developing the AI model, the problem they overwhelmingly faced was the retrospective research; creating issues in training the algorithm with the same dataset as the one they wished to test on. They switched to a perspective approach, which takes the view of no knowledge of when a seizure is happening with no data available. The forecasting systems start out as less accurate, but combined with detection systems, it improves over time and more importantly it becomes patient specific. Two papers were published on the research: Continental generalization of an AI system for clinical seizure recognition and A multimodal AI system for out-of-distribution generalization of seizure detection.
Current data collection is through the help of a high number of electrodes connected to the patient. The next dilemma for this body of research is to understand if they can reduce the number of electrodes and compensate for it through the AI model. “We wish for seizure prediction to continue to improve. The research we are doing is impacting real people, it’s not just a theoretical puzzle to solve. We are in touch with people from all over the world, their stories are profound and real – from a patient who need supervised assistance going up or down the stairs, to another who fears holding their infant baby over potential accidents during a seizure. What we wish to give back to the community is independence,” states Omid.
Enabling increased independence for those with epilepsy is top of mind and heart for Francesca Fedeli and Roberto D’Angelo, co-founders of FightTheStroke Foundation, who set out to help their son, Mario. At the 2019 Microsoft Hackathon, they developed MirrorHR, an epilepsy research app available in 13 languages and is used daily by hundreds of families across 31 countries.