Midway evaluation in Computer Science - Ana Borovac

Live Stream

Thesis title: Towards clinically useful neonatal seizure detection algorithms

Doctoral candidate: Ana Borovac

Doctoral committee:
Steinn Guðmundsson, Professor in Computer Science, University of Iceland, Tómas P. Rúnarsson, Professor in Industrial Engineering, University of Iceland, Samspa Vanhatalo, Professor in the Department of Physiology, University of Helsinki.

Abstract

In this project, several properties of a neonatal seizure detector based on a convolutional neural network were investigated. We found that by utilizing annotations from multiple human experts when limited training data is available, results in improved performance, compared to using annotations from a single expert only. Furthermore, an ensemble of detectors trained on small disjoint data sets performed similarly to a detector trained on the union of the data. This suggests a practical strategy for training a classifier on data from multiple institutions without having to share potentially sensitive EEG data between institutions. For a seizure detector to find widespread use, it needs to work with a varying number of EEG channels. We found that the performance of such a detector was in line with the performance of a human expert, a lower number of channels, results in a lower, but still sufficiently large, number of detected seizures. Finally, we found that using dropout during training and prediction, leads to more reliable seizure/non-seizure probability estimates. As a result, users can be notified when the detector is not certain about its predictions, leading to increased trust in the system.