Search
Close this search box.
Search
Close this search box.

Harnessing artificial intelligence to understand refractory epilepsy

Professor Mark Richardson

Patient data research

- Harnessing artificial intelligence to understand refractory epilepsy
- Epilepsy Research UK & Angelini Pharma
- Innovations in Healthcare project

Professor Mark Richardson has been awarded funding to accelerate innovations in refractory epilepsy through patient data, as part of the Epilepsy Research UK and Angelini Pharma strategic alliance. Professor Richardson’s team at King’s College London are developing artificial intelligence (AI) tools to analyse and annotate electronic NHS health records. Here, Mark explains how this technology will be used to predict refractory epilepsy, which could lead to novel treatments and prevention in the future.

Jerome K Jerome was a humorous writer, perhaps best known for the novel ‘Three men in a boat’, who died in 1927. During the summer of 2020, a new short story appeared with his name as the author and written in his particular style. Was this a previously undiscovered work? Given that the theme of the story was the importance of being on Twitter, clearly it could not have been written by him.

In fact, it had been entirely generated by an Artificial Intelligence (AI) system called GPT3, created by Google, without any human input other than giving the AI system the name of the author, the title of the story, and the first word. Everything else, including plot, characters, dialogue, and the subtle wit and humour of Jerome K Jerome was entirely generated by computers.

More recently, another Google AI system called LaMDA, was incorrectly reported by a Google engineer as having become conscious. The reason the engineer had reached this conclusion is because LaMDA is able to engage in real-time conversation with humans, responding to the verbal cues and directions of the human, and generating strikingly human-like responses, including introspective reflections on its own self-hood. But again, this is all generated by computers.

These prominent and newsworthy advances in how computers can understand and manipulate language in a highly sophisticated manner are directly relevant to gaining a better understanding of epilepsy.

A key challenge in improving outcomes for people with epilepsy is to understand why some people with epilepsy respond to their medication, but approximately one in three do not – this failure to respond to medication is known as ‘refractory epilepsy’. Several carefully conducted studies have analysed medical records to try to identify features in people who have recently developed epilepsy and predict whether they will develop refractory epilepsy in the future. If we could identify the reasons why some do not respond to medication, we could focus on developing new approaches to treatment for these people right from the start of their epilepsy. These previous studies have required the laborious review of entire medical records from hundreds of people, going through thousands of pages of text and large databases of results from blood tests, EEGs, brain scans and much else. Because a human researcher is required to read and understand the records, the number that could be examined is limited to hundreds rather than thousands or millions. These studies have revealed some hints about the causes of refractory epilepsy, but much remains unexplained.

Medical records contain vast amounts of information about each person’s medical history, including descriptions of how their health conditions started and developed, how they were treated and whether treatments were effective, information about risk factors and potential inheritance of diseases. Mining this rich information should reveal crucial explanations and predictors for many health outcomes. It’s likely that mining thousands of records would reveal important discoveries missed in the smaller studies conducted previously.

At the current time, most hospitals and GPs use electronic (rather than paper) records. Two factors make it difficult to extract information from electronic medical records from many thousands of people. Firstly, a trivial but nonetheless difficult issue, which is that each medical record consists of thousands of pieces of information which are held in a wide variety of databases, documents and images, often in formats that are difficult to access. Secondly, most of the really valuable information is in so-called ‘free-text’, that is, in whatever form of words the healthcare professional chose to use as they were making notes in the record. The recent extraordinary advances in computer understanding of language provides a breakthrough for mining medical records – for the first time, we can now automatically extract meaning from free-text and can use this understanding to create structured health information about an individual, translating the free-text into medical diagnoses, symptoms, investigation results, treatments, and outcomes.

Researchers in my team at King’s College London have developed a comprehensive system for accessing, extracting, understanding and organising vast amounts of information from thousands of electronic health records using AI. This system has been refined and optimised to interrogate electronic records from people with epilepsy, and the stage is now set to mine tens of thousands of records to identify the features of people with epilepsy that predict future refractory epilepsy.

You can read more about Prof Mark Richardson’s Innovations in Healthcare project, supported by Epilepsy Research UK and Angelini Pharma, here.