In this dissertation, Emil explores the challenges associated with leveraging deep learning technologies for medical data. The continuous accumulation of patient information from diverse healthcare sources, such as hospitals and pharmacies, presents a wealth of essential knowledge and patterns. Extracting these patterns holds the potential to significantly aid clinicians in their daily tasks. Deep learning emerges as a crucial tool for this purpose, excelling at uncovering latent patterns within extensive datasets.
However, significant challenges persist despite the immense potential for automatically learning patterns from medical data. Healthcare data is notorious for its complexity, encompassing issues like data heterogeneity and the integration of domain knowledge. Emil has delved into three key questions addressing these challenges: How can hierarchical domain knowledge be integrated with deep learning technologies? How can we effectively incorporate and learn from the structural and temporal aspects of healthcare data? Which deep learning technologies are best suited for specific medical challenges?
In response to these questions, Emil proposes innovative methods for integrating medical domain knowledge with deep learning technologies. One approach involves extending the multi-label soft margin loss function with a hierarchical aspect, ensuring the accuracy of predictions within the medical ICD taxonomy for diagnosis prediction. Another method, applied to patient electronic health record graphs, extracts the structure of medical domain taxonomies to pre-initialize node embeddings for subsequent medical analytics using graph convolutional neural networks. Lastly, leveraging sequences of medical events related to patient hospitalizations, specific demographic-based threshold values are employed to integrate clinical measurement values into concept representations. This integration is used to train transformer encoders to predict patient hospitalization times.
The findings underscore the significance of integrating medical domain knowledge with deep learning to enhance the performance of medical analytical systems.

Members of the assessment committee are Professor Yuval Shahar, Ben Gurion University (Israel), Associate Professor Catia Pesquita, Lisbon University (Portugal), and Associate Professor Manfred Jaeger (chairman), Aalborg University (Denmark). Supervisor for the thesis has been Professor Katja Hose, Aalborg University. Co-supervisor for the thesis has been Assistant Professor Tomer Sagi, Aalborg University. Moderator has been Associate Professor Gabriela Montoya.

All interested parties are welcome. After the defense the department will be hosting a small reception in cluster 4.