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 Pierpaolo Sorrentino, AIX-Marseille University, France
- Second opponent: Professor Sebastien Gros, Norwegian University of Science and Technology
- Third member and chair of the evaluation committee: Professor Mona Elisabeth Skjelland, University of Oslo
Chair of the Defence
Professor Emeritus Reidun Torp, University of Oslo
Principal Supervisor
Ira Haraldsen, Oslo University Hospital
Summary
This thesis is a contribution to the intersection of artificial intelligence and neuroscience. The overall aim was to address methodological challenges associated with applying deep learning (DL) to electroencephalography (EEG), focussing on robustness, generalisation, and interpretability.
Three studies were conducted. The first addressed the challenge of heterogeneous EEG hardware. A method named Region Based Pooling (RBP) was developed to enable DL models to handle varying numbers of EEG electrodes. RBP transformed electrode-level recordings into region-based representations, achieving similar results to spherical spline interpolation when evaluated on down-sampled channel systems.
The second study investigated cross-dataset generalisation. Using five different datasets and age prediction as its example task, extensive experiments were conducted across 1805 hyperparameter configurations. The results demonstrated that DL models can learn age related EEG characteristics that generalise to previously unseen datasets with moderate to strong correlations under favourable conditions. The study further highlighted dataset shifts as an important challenge for clinical adoption and underscored the importance of external validation.
The third study focused on interpretability and biomarker discovery. A novel self-supervised learning framework inspired by EEG reactivity was proposed. Deep learning models were trained to learn normative transition patterns between eyes-closed and eyes-open EEG conditions in healthy individuals, and deviations from these learned expectations were treated as candidate biomarkers. Although the framework did not achieve strong predictive performance, it provided a transparent methodology for residual-based biomarker discovery grounded in neurophysiology.
Overall, this thesis contributes to the development and evaluation of deep learning methods for EEG analysis, with a particular focus on adaptability, generalisability, and interpretability.
Additional information
Contact the research support staff.