Doctoral defence in Computer Science - Ana Borovac

Doctoral candidate: Ana Borovac

Title of thesis: Towards clinically useful neonatal seizure detection algorithms

Opponents: Dr. Alison O'Shea, assistant lecturer at the Munster Technological University in Ireland and Dr. Maarten De Vos, professor at KU Leuven in Belgium

Advisor: Dr. Steinn Gudmundsson, Professor at the Faculty of Industrial Engineering, Mechanical Engineering and Computer Science, University of Iceland.

Doctoral committee:
Dr. Steinn Gudmundsson, Professor at the Faculty of Industrial Engineering, Mechanical Engineering and Computer Science, University of Iceland
Dr. Thomas Philip Runarsson, Professor at the Faculty of Industrial Engineering, Mechanical Engineering and Computer Science, University of Iceland
Dr. Sampsa Vanhatalo, Professor at University of Helsinki in Finland.

Chair of Ceremony: Rúnar Unnþórsson, Professor and head of the Faculty of Industrial Engineering, Mechanical Engineering and Computer Science at the University of Iceland.

Abstract

Seizures are the most common neurological emergency among neonates and if left untreated they can cause permanent brain damage. The current gold standard for neonatal seizure detection is continuous EEG visually interpreted by a human expert. Such expertise is rarely available round-the-clock and reliable automatic seizure detection could fill the gap. The detectors have been in development for about three decades and they are approaching human-level classification performance. In this work, we analyse and improve a detector based on deep learning. First, we study how to utilise the rather small data sets normally available for the development of detectors. In this case, we found that it is important to use the data for which multiple experts agreed on the seizure annotation. The size of the data sets could be increased if data from multiple institutions would be combined. To share the data information safely without breaking data-sharing policies, we propose to use an ensemble of locally developed seizure detectors. Second, we analyse the classification performance of a detector for a reduced number of input EEG signals and results indicate the performance is acceptable even when the number of input signals is small. The detector is further improved by utilizing calibration methods to be able to inform the user about uncertain predictions. Last, we gather user experience with a commercial seizure detector by conducting interviews with nurses. A common finding was that automatic seizure detections need to always be verified, but the nurses still find the detector useful despite many falsely detected seizures.

About the candidate

Ana Borovac was born in Kranj, Slovenia in 1994. She spent her childhood in Ljubljana, Slovenia, where she received her primary education. She earned her Bachelor's degree in Mathematics in 2017 and her Master's degree in Computer Science and Mathematics in 2019 from the Faculty of Mathematics and Physics at the University of Ljubljana, Slovenia. As part of the Erasmus+ programme, she spent the last year of her master's studies at Utrecht University in the Netherlands. In 2019, she moved to Iceland to pursue her doctoral studies in Computer Science at the University of Iceland.