Florence Aellen is the deep learning specialist at the Cognitive Computational Neuroscience Lab of the Institute of Computer Science, University of Bern. For her PhD she works with a very interdisciplinary research team to unravel interrelations between electrical brain activity and states of consciousness. Herself originally coming from a background in mathematics and theoretical physics, she now puts her computational expertise in the service of clinical applications, learning a lot in the process.
Flo(rence), can you tell us about your research?
My project is embedded within the computational platform for the Interfaculty Research Cooperation “Decoding Sleep” at the University of Bern and the Inselspital, Bern University Hospital. Apart from background methodological research, I have investigated indicators for sleep disorders and most recently have extracted positive predictive markers for coma patients out of EEG data. In our interdisciplinary team, I am responsible for data analysis via machine and deep learning.
Today, assessing the state of coma patients in critical care is very difficult, focusses mostly on negative markers, allows for inter-scorer variability, and leaves a third of patients with an unclear prognosis. We have trained deep learning models to discriminate coma survivors from non survivors. So, if the network outputs a value above a defined threshold, it is a positive indicator. Focusing on such positive markers can be a critical piece of additional information for clinicians to consider and it can help family members if they receive some positive news in such extreme situations. We are one step away from having a fully automated pipeline based on complex EEG data, which if validated on new patient cohorts and hospitals, could help to predict survival objectively and quite accurately, even for patients where this was previously difficult.
I like solving puzzles. This is an aspect I especially appreciate about coding.
Why are you using Artificial Intelligence for this?
So, for this study we have EEG data from 134 coma patients from four hospitals across Switzerland, collected within the first 24 hours after falling into a coma. Patients were presented with 20 minutes of sounds and while their brain’s processing of these sounds was recorded. This first phase of coma is critical to predict the outcome of patients. Deep learning was especially apt for these calculations because unlike machine learning for neural data it does not need to focus on a narrow aspect of the data. With it, you can explore the whole EEG response in different electrodes at the same time. It can also account for the great variation in patient EEG responses: some respond at a neural level quickly, others slowly to the presented sounds. So, deep learning can capture more patterns within the data and provide a meaningful overall picture.
How do you perceive your research environment?
EEG data has been explored with deep learning many times, but most of this research stays purely academical. What I like about Bern is that it offers the possibility of bringing deep learning into the clinics and into an application. There are many initiatives here and a supportive environment to apply your research clinically.
I think this project is especially suited to be translated. We have worked with a good, validated data set. And our work is not the determining factor but would be part of a bigger clinical future evaluation.
I like solving puzzles. This is an aspect I especially appreciate about coding. I wanted to specialize in computer science and was quite interested in machine and deep learning, so when I found out about this specific interdisciplinary project, it really motivated me. I feel that I can get so much out of it, learn so much about fields I didn’t know. And, importantly, I know that my research can be helpful and make a difference in clinical practice.
Florence is a PhD student at the University of Bern, in the Institute of Computer Science under the supervision of Ass. Prof. Dr. Athina Tzovara. She has a Master of Science in theoretical physics from the University of Bern, where she focused her studies on symmetries in general relativity.
Prior to that she earned her bachelor’s in mathematics also from the University of Bern.
Her work in her PhD focuses on AI for neural signals and clinical applications. She is involved in multiple projects ranging from methodology for basic and applied research to clinical applications in the fields of sleep disorders and coma after cardiac arrest. In her project on coma after cardiac arrest she used deep learning to predict outcome of patients in a comatose state based on their neural data during processing of sounds. The work shows state of the art performance in predicting survival of patients and could, if further validated in future implementations in the clinic be used as additional information for physicians. Her work shows multiple advantages compared to traditional measures, such as standardization and automation of protocol, as well as reducing time used for the analysis and good performance on patients that with current used methods are without clear prognosis.
The results of her project predicting outcome for coma patients