Recent years have witnessed an unparalleled surge in deep neural networks (DNNs) research, surpassing traditional machine learning (ML) and statistical methods on benchmark datasets in computer vision, audio processing and natural language processing (NLP). Much of this success can be attributed to the availability of numerous open-source datasets, advanced computational resources and algorithms. These algorithms learn multiple levels of simple to complex abstractions (or representations) of data resulting in superior performances on downstream applications. This has led to an increasing interest in reaping the potential of DNNs in real-life safety-critical domains such as autonomous driving, security systems and healthcare. Each of them comes with their own set of complexities and requirements, thereby necessitating the development of new approaches to address domain-specific problems, even if building on common foundations.

In this thesis, we address data science related challenges involved in learning effective prediction models from structured electronic health records (EHRs). In particular, questions related to numerical representation of complex and heterogeneous clinical concepts, modelling the sequential structure of EHRs and quantifying prediction uncertainties are studied. From a clinical perspective, the question of predicting onset of adverse outcomes for individual patients is considered to enable early interventions, improve patient outcomes, curb unnecessary expenditures and expand clinical knowledge.

This is a compilation thesis including five articles. It begins by describing a healthcare information platform that encapsulates clinical, operational and financial data of patients across all public care delivery units in Halland, Sweden. Thus, the platform overcomes the technical and legislative data-related challenges inherent to the modern era's complex and fragmented healthcare sector. The thesis presents evidence that expert clinical features are powerful predictors of adverse patient outcomes. However, they are well complemented by clinical concept embeddings; gleaned via NLP inspired language models. In particular, a novel representation learning framework (KAFE: Knowledge And Frequency adapted Embeddings) has been proposed that leverages medical knowledge schema and adversarial principles to learn high quality embeddings of both frequent and rare clinical concepts. In the context of sequential EHR modelling, benchmark experiments on cost-sensitive recurrent nets have shown significant improvements compared to non-sequential networks. In particular, an attention based hierarchical recurrent net is proposed that represents individual patients as weighted sums of ordered visits, where visits are, in turn, represented as weighted sums of unordered clinical concepts. In the context of uncertainty quantification and building trust in models, the field of deep evidential learning has been extended. In particular for multi-label tasks, simple extensions to current neural network architecture are proposed, coupled with a novel loss criterion to infer prediction uncertainties without compromising on accuracy. Moreover, a qualitative assessment of the model behaviour has also been an important part of the research articles, to analyse the correlations learned by the model in relation to established clinical science.

Put together, we develop DEep Evidential Doctor (DEED). DEED is a generic predictive model that learns efficient representations of patients and clinical concepts from EHRs and quantifies its confidence in individual predictions. It is also equipped to infer unseen labels.

Overall, this thesis presents a few small steps towards solving the bigger goal of artificial intelligence (AI) in healthcare. The research has consistently shown impressive prediction performance for multiple adverse outcomes. However, we believe that there are numerous emerging challenges to be addressed in order to reap the full benefits of data and AI in healthcare. For future works, we aim to extend the DEED framework to incorporate wider data modalities such as clinical notes, signals and daily lifestyle information. We will also work to equip DEED with explainability features.