Due to copyright issues, an electronic copy of the thesis must be ordered from the faculty. For the faculty to have time to process the order, the order must be received by the faculty at the latest 2 days before the public defence. Orders received later than 2 days before the defence will not be processed. After the public defence, please address any inquiries regarding the thesis to the candidate.
Trial Lecture – time and place
See Trial Lecture.
Adjudication committee
- First opponent: Professor Sebastien Nicolas Gros, Norwegian University of Science and Technology
- Second opponent: Associate Professor Erik Westhall, Lund University
- Third member and chair of the evaluation committee: Professor Elisabeth Gulowsen Celius, University of Oslo
Chair of the Defence
Professor Emeritus Reidun Torp, University of Oslo
Principal Supervisor
Group Leader Ira Ronit Hebold Haraldsen, Oslo University Hospital
Summary
Artificial intelligence (AI) has the potential to support clinical decision-making, but its integration into healthcare remains limited by concerns related to trust, reliability, and robustness. In neurophysiology, electroencephalography (EEG) is widely used, yet interpretation is complex and partly subjective due to high signal dimensionality, variability and complexity. AI-based analysis may help identify subtle patterns related to disease risk and progression, including dementia, but clinical use requires reliable and trustworthy methods.
The aims of this thesis were to develop and evaluate AI-based methods for EEG analysis with a focus on robustness and uncertainty, and to investigate how uncertainty information can support safer clinical use.
The research first addressed practical bottlenecks by developing and validating an automated 3D-scan based pipeline for electrode localization, enabling more advanced and scalable EEG analysis. The core work focused on the reliability of AI predictions. By implementing deep learning models and an uncertainty-based selection framework, the research investigated whether a model’s own confidence could be used to improve performance. The results demonstrated that classification accuracy increases when the most uncertain predictions are excluded, providing a mechanism to flag doubtful cases for expert review.
Furthermore, the thesis investigated how dataset shifts, common when moving between different hospitals or equipment, affect AI performance. By comparing various AI ensemble methods, where multiple models collaborate, the research showed that certain types of ensembles provide more reliable uncertainty estimates than others. Crucially, their uncertainty and calibration increases in tandem with declining accuracy.
Overall, the findings suggest that for AI-based EEG tools to be trustworthy, they must be defined not only by their predictive power but also by a transparent and reliable understanding of their own limitations.
Additional information
Contact the research support staff.