Teaching AI to predict seizures using long-term brain activity recordings
Explore Pilot Study
£39,986
Dr Dominic Burrows
King's College London
Many people with epilepsy live with constant uncertainty about when their next seizure will happen. Even with treatment, many people continue to have seizures, and the lack of warning can be one of the hardest parts: it drives anxiety, restricts work and daily activities, affects sleep and independence, and increases the risk of injury. This project aims to tackle that core problem by improving how we use EEG to give earlier warning of seizures.
"One of the greatest burdens of epilepsy is the unpredictability of seizures, which can severely restrict independence and quality of life. Recent advances in AI now make it possible to learn long-range patterns in brain activity over time - we hope to use this technology to design models for more accurate, clinically useful seizure forecasting.
A major reason seizure prediction has been so difficult is that most computer models need doctors to label seizures in EEG recordings. That takes a lot of time, so only a small amount of data can be used, and the models often don’t work well outside the original hospital or device. Newer AI takes a different approach: it can learn from very large amounts of raw data without labels first, then be ‘taught’ a specific job with a much smaller labelled set. This is similar to how systems like ChatGPT became useful – they first learn general patterns from huge amounts of text, then are adapted to particular tasks. We will use the same idea with long-term epilepsy EEG, so the model can learn the normal background patterns and the subtle changes that may happen before seizures.
We will do this in two steps. First, we will train the model on large amounts of EEG using simple learning games (for example, hiding parts of the signal and asking it to fill them in, or asking it to predict what comes next). Then we will train it further using EEG where seizures have been confirmed by clinicians, and test whether it can i) spot seizures reliably and ii) give any useful early warning of higher-risk periods minutes to hours ahead.
If seizure warning can be made more reliable, the impact could be substantial. A practical warning could help people take protective steps (for example avoiding risky situations, asking for support, following an agreed safety plan, or using rescue medication when appropriate). Over time, better warning signals could also support clinical care by providing a clearer, objective picture of seizure patterns, rather than relying on diaries and occasional clinic visits.
