Master’s thesis presentation by Sofie Wållberg, MPSYS
Welcome to the master’s thesis presentation “Classification of epileptic seizures – Accelerometer based detection of hypermotor seizures” by Sofie Wållberg, MPSYS
When? 14:00, Thursday, June 15 2017
Where? Landahlsrummet (room 7430), Hörsalsvägen 11, 7th floor
Examiner: Tomas McKelvey
In order to provide objective data for physicians to evaluate, accelerometry can be used to monitor movements of patients suffering from epilepsy. The focus of this thesis is to evaluate data gathered at Sahlgrenska Academy with the purpose to build models which detects nocturnal hypermotor seizures (HMS). This thesis extends work from a previous study focused on detecting generalized tonic-clonic seizures (GTCS), and the ultimate goal is to combine the results to create multi-class classification algorithms to be implemented with wearable electronics. When compared to GTCS, HMS commonly have a larger variability with varying motoric responses and durations among the patients. To overcome the increased difficulty of HMS detection, two major approaches are considered in this thesis. The general approach where the classifiers are constructed using data from multiple patients and the patient specific approach where models are created for each patient using their individual data sets for both training and testing. A set of features from the previous work is considered and extended to incorporate common features used in literature regarding HMS detection. The feature space is used to train and evaluate the classification methods logistic regression, k-nearest neighbors (KNN), support vector machines (SVM), random forest and kernel density estimation (KDE). Additionally, pre-processed accelerometer data is used to train artificial neural networks (ANN) without first calculating features. It proved to be a difficult task to achieve perfect accuracy using the current data set, the listed methods and the implemented feature space. It was possible to achieve high sensitivity but at the cost of a large amount of false positives. By visual inspection of the accelerometer data and video recordings a set of subtle or atypical seizure could be identified. The results of the classification showed that these seizures were difficult to detect as expected from the inspection, they were either too subtle or too atypical for accurate classification. Their similarity to normal non-seizure movements are most likely affecting the performance in a negative manner and several post-processing steps had to be considered in order to reduce the amount to be within acceptable limits. Furthermore, the feature space was evaluated to find differences between GTCS and HMS, the results indicate that there are considerable differences and the feature space could serve as a basis for multi-class classification in future work.