Before the advent of AI, medicine was more preventive and curative. Since the introduction of machine learning, a sub-field of AI, a new paradigm known as predictive medicine has emerged. Researchers have developed and are still developing several machine learning tools to predict adverse medical events, enabling physicians to take early action and prevent the worst. This thesis is a continuation of this work. It aims to use machine learning tools, namely deep learning models, to predict adverse medical events during patients' medical stays. The medical events addressed throughout the thesis include the prediction of mortality in intensive care units, the detection of depression, and the prediction of unplanned hospital readmissions. Extensive experiments and comparisons with cutting-edge models have shown that the newly introduced deep learning models improve the accuracy of tackled medical tasks. Overall, the thesis findings demonstrate that machine-learning tools, specifically deep-learning models, are ideal candidates for predicting or detecting adverse events during a patient's medical stay. Although that is the immediate thesis goal, the long-term goal is to improve these models to be more accurate, robust, and explainable. This would facilitate their adoption in real-world medical settings and help physicians in their daily decision-making processes.

The thesis is published in the Digital Collection of TalTech Library.

Supervisor Prof. Sadok Ben Yahia and co-supervisor Prof. Nédra Mellouli-Nauwynck.

Oponents:

• Prof. Vannary Meas-Yedid Hardy, Institut Pasteur-Paris, France;
• Prof. Henning Christiansen, Roskilde University, Denmark.

Follow public defence in Zoom.

Meeting ID: 910 4754 1742
Passcode: 665177