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: Resarch fellow Roman Hornung, Augsburg University Hospital, Germany
- Second opponent: Head, Computational Systems Biomedicine Lab Benno Schwikowski, Pasteur Institute, Paris, France
- Third member and chair of the evaluation committee: Associate Professor Marieke Lydia Kuijjer, University of Oslo
Chair of the Defence
Professor Vessela N. Kristensen, University of Oslo
Principal Supervisor
Professor Tero Aittokallio, University of Oslo
Summary
This thesis investigates how artificial intelligence can improve the integration and interpretation of heterogeneous patient data for precision oncology, with pancreatic cancer as the main application. It addresses three central challenges: how to stratify patients using complex multi-modal data, how to analyse datasets with missing modalities, and how to make machine learning models sufficiently interpretable to generate clinically meaningful insights.
Four interconnected studies were conducted. First, an explainable computational framework was developed for PDAC patient stratification using molecular data. Multiple explainability methods consistently identified DNA methylation features as major contributors to model predictions, highlighting epigenetic alterations as potential drivers of clinically relevant patient differences. Second, iMML, an open-source software framework for multi-modal machine learning with incomplete data, was developed to support reproducible research across several data analysis tasks. Third, multiple clustering methods were systematically benchmarked across several multi-modal datasets, including datasets with complete information and datasets with missing modalities. The results provide practical recommendations for selecting clustering methods according to data characteristics, missingness and analytical objectives. Finally, these methods were applied to six molecular layers in pancreatic cancer, identifying two patient subgroups with significantly different overall survival, recurrence-free survival and tumour mutational burden.
Overall, the thesis demonstrates that robust handling of incomplete multi-modal data, systematic method selection and model explainability can improve the reliability and applicability of AI-based patient stratification.
Additional information
Contact the research support staff.