BORIS Theses

BORIS Theses
Bern Open Repository and Information System

Artificial intelligence techniques for studying neural functions in coma and sleep disorders

Aellen, Florence Marcelle (2023). Artificial intelligence techniques for studying neural functions in coma and sleep disorders. (Thesis). Universität Bern, Bern

23aellen_fm.pdf - Thesis
Available under License Creative Commons: Attribution-Noncommercial (CC-BY-NC 4.0).

Download (43MB) | Preview


The use of artificial intelligence in computational neuroscience has increased within the last years. In the field of electroencephalography (EEG) research machine and deep learning, models show huge potential. EEG data is high dimensional, and complex models are well suited for their analysis. However, the use of artificial intelligence in EEG research and clinical applications is not yet established, and multiple challenges remain to be addressed. This thesis is focused on analyzing neurological EEG signals for clinical applications with artificial intelligence and is split into three sub-projects. The first project is a methodological contribution, presenting a proof of concept that deep learning on EEG signals can be used as a multivariate pattern analysis technique for research. Even though the field of deep learning for EEG has produced many publications, the use of these algorithms in research for the analysis of EEG signals is not established. Therefore for my first project, I developed an analysis pipeline based on a deep learning architecture, data augmentation techniques, and feature extraction method that is class and trial-specific. In summary, I present a novel multivariate pattern analysis pipeline for EEG data based on deep learning that can extract in a data-driven way trial-by-trial discriminant activity. In the second part of this thesis, I present a clinical application of predicting the outcome of comatose patients after cardiac arrest. Outcome prediction of patients in a coma is today still an open challenge, that depends on subjective clinical evaluations. Importantly, current clinical markers can leave up to a third of patients without a clear prognosis. To address this challenge, I trained a convolutional neural network on EEG signals of coma patients that were exposed to standardized auditory stimulations. This work showed a high predictive power of the trained deep learning model, also on patients that were without a established prognosis based on existing clinical criteria. These results emphasize the potential of deep learning models for predicting outcome of coma and assisting clinicians. In the last part of my thesis, I focused on sleep-wake disorders and studied whether unsupervised machine learning techniques could improve diagnosis. The field of sleep-wake disorders is convoluted, as they can cooccur within patients, and only a few disorders have clear diagnostic biomarkers. Thus I developed a pipeline based on an unsupervised clustering algorithm to disentangle the full landscape of sleep-wake disorders. First I reproduced previous results in a sub-cohort of patients with central disorders of hypersomnolence. The verified pipeline was then used on the full landscape of sleep-wake disorders, where I identified clear clusters of disorders with clear diagnostic biomarkers. My results call for new biomarkers, to improve patient phenotyping.

Item Type: Thesis
Dissertation Type: Cumulative
Date of Defense: 31 March 2023
Subjects: 000 Computer science, knowledge & systems
500 Science > 510 Mathematics
600 Technology > 610 Medicine & health
Institute / Center: 08 Faculty of Science > Institute of Computer Science (INF)
Depositing User: Sarah Stalder
Date Deposited: 01 Feb 2024 09:08
Last Modified: 31 Mar 2024 22:25

Actions (login required)

View Item View Item