Fake data helping real brains

Assessing brain activity using EEG, which measures the brain’s electrical activity using sensors on the scalp, is a complicated and time-consuming process. For this reason, M20 PhD researcher Clemens Kaiser is working on creating an AI tool that can generate fake EEG data. Researchers can then use this data to make better predictions based on EEG results, such as better estimating the consequences of a concussion.

Text: Fardou Haagsma (Science Communication student at the University of Groningen)

Brain activity can be measured using the electroencephalogram, better known as the EEG. This is a useful tool with many applications, such as predicting whether someone who suffered a concussion will experience long-term consequences from the injury. However, EEG is not without its issues. ‘The process of collecting EEG data isn’t an easy one,’ according to Kaiser. ‘It takes at least half an hour to set up for each person.’

The reason for this needed time is many small electrodes need to be attached to someone's head using an adhesive gel, the PhD student explains, as he is pointing to the parts of his head where the electrodes would be placed. Besides, the actual measurement of the EEG is never perfect. ‘You have to be very still whilst having an EEG done. That means don’t blink or speak, or anything. When you blink, it will cause a huge disturbance in the data.’ 

Training a model

That is why Kaiser is looking for a solution. He works on training models using EEG data. A model is a computer program that simulates something complex in a simplified way, helping researchers understand and predict how it works. In his project, Kaiser will use such a model with EEG data. The model will, once it is trained, recognise changes in the patterns of the EEG data. These changes can for example indicate when someone with epilepsy may have a seizure. In addition, it can help predict whether someone is paying attention to a task or train AI to recognise human emotions.

However, there is a catch. To train a model well, the model needs to see a substantial amount of EEG data, requiring many hours of collecting EEG data for researchers and having electrodes on their heads while sitting still for participants. That is not even mentioning the process it takes to find the right people to be included in such a study, according to the PhD student.

Therefore, Kaiser is working on making this process, to train and improve a model, less time-consuming. He is generating fake EEG data that can be used to train this model. ‘We use a combination of real EEG data and data generated by an AI tool,’ he says. This way, fewer real EEGs have to be collected for his model to make good predictions.

Challenges 

The PhD researcher generates this fake EEG using Generative Adversarial Networks (GANs), which is a type of AI able to make ‘new content’. ‘For example, ChatGPT can generate a text. We want to do something similar, that’s maybe a bit less exciting: generating fake EEG data,’ Kaiser says with a grin.

However, this process involves certain challenges. One of his main obstacles right now is to find an algorithm that can produce EEG data. This is not the same as generating a text. ‘It’s more difficult to produce something like EEG data.’ Kaiser also mentions that determining whether the data such an algorithm generates looks real, is difficult. ‘With a picture, you can see whether or not it looks real, but with EEG data, even when you’re familiar with how it looks, it is hard to say whether it can pass as real EEG data.’ Furthermore, GANs are very hard to train, the PhD researcher stresses. ‘They can be quite unstable. The tool may end up generating EEG samples that are very similar to each other.’

‘There is another challenge,’ Kaiser says. The model consists of two parts that need to work together and are dependent on each other to learn and become better. They therefore need to learn at approximately the same speed. However, if one part becomes better than the other much more quickly, the other part will no longer keep up and won’t learn anything. If that happens, the AI can no longer determine whether the fake EEG sample it produced sufficiently resembles the real EEG data, or it will have difficulty generating data that looks real. Therefore, it is searching for the right balance.

Interdisciplinary research

For his PhD project, Kaiser received an M20 grant from the University of Groningen. The funding is dedicated to interdisciplinary projects made possible by a donation from the Ubbo Emmius Foundation (UEF). With a bachelor’s degree in Business Administration and Psychology, followed by a master’s degree in Computational Cognitive Neuroscience, Clemens has quite a diverse background himself already. ‘I never expected to do a PhD in AI,’ says Kaiser, humbly mentioning that he does not have ‘a proper AI background’. 

The project itself is a mix between cognitive neuroscience and artificial intelligence, the PhD researcher says. Therefore, Kaiser is also supervised by colleagues from different departments. ‘I have two supervisors: Marieke van Vugt from the AI department and Natasha Maurits from the Clinical Neuroengineering department at the UMCG.’ Kaiser mostly sees benefits to this interdisciplinarity as he receives input from different perspectives. Nevertheless, interdisciplinary work also has its challenges. ‘Natasha will sometimes have suggestions that aren’t immediately obvious to me. So then I need to do a bit more research, but that’s also beneficial.’

Hopeful for results

Kaiser is currently in the early stages of his research. ‘At this point, I don’t even use EEG data.’ That, however, does not mean he is not thinking about the future. ‘Ideally, I hope to have a model that works at the end of this project.’ This model should lead to reliable predictions about when people with epilepsy will have a seizure. He is also realistic in his outlook on future results. ‘It is realistic to expect quite some issues. If we find a way to add some knowledge to the model in a way that improves the model a little bit, that would also be a good outcome.’