Automated Detection and Prediction of Seizures Using Probing Neurostimulation
Citation
Ebrahim, Senan. 2019. Automated Detection and Prediction of Seizures Using Probing Neurostimulation. Doctoral dissertation, Harvard University, Graduate School of Arts & Sciences.Abstract
To study and prevent seizures in patients with epilepsy, clinical neurophysiologists seek effective methods to detect seizures, and ideally predict them for timely intervention. EEG monitoring is the clinical gold standard, while single pulse electrical stimulation (SPES) has emerged as a modality for active probing of brain states.Seizure detection on EEG data is employed in neurocritical care, epilepsy diagnosis and management, and novel therapies such as closed loop stimulation. A detector with both high sensitivity and specificity is necessary for clinical use. We introduce a generalized linear model built from a set of 141 custom features for classification of seizures in continuous EEG. In 16 rats with epilepsy exhibiting 1012 labeled seizures, we built a pooled classifier with an AUROC of 0.995.
We also aim to automate multiple additional EEG labels in neurocritical care settings. We developed a robust method using 592 features extracted from the EEG data of 97 ICU patients. An affinity propagation (AP) method was used to generate 30-50 clusters for each patient; clinical EEG experts labeled clusters by observing the medoids. We observed a 60-fold reduction in expert labeling time, without a significant change in interrater agreement.
Seizure prediction by analyzing continuous EEG is a therapeutic goal for closed loop seizure preemption. We developed pooled and individualized predictors using the following methods: (1) support vector machines (SVM); and (2) multilayer perceptrons (MLP). We then assessed model performance using epileptic rat data. MLP yielded the highest AUROC of 0.88 on our pooled dataset of 1012 rodent seizures.
We finally studied whether SPES into the hippocampal focus of rats with focal epilepsy yields novel predictive features, as compared to EEG monitoring. In the induced seizures for three subjects, we found multiple features across time and frequency domains, including evoked HFOs, that significantly change in preictal periods. In the two subjects with multiple spontaneous seizures, we trained SVM classifiers that perform at AUROCs of 0.94 and 0.98.
These results offer new insights into the mechanisms underlying seizure initiation, and may help improve diagnostic and therapeutic approaches for patients suffering from focal epilepsy.
Terms of Use
This article is made available under the terms and conditions applicable to Other Posted Material, as set forth at http://nrs.harvard.edu/urn-3:HUL.InstRepos:dash.current.terms-of-use#LAACitable link to this page
http://nrs.harvard.edu/urn-3:HUL.InstRepos:42029517
Collections
- FAS Theses and Dissertations [5815]
Contact administrator regarding this item (to report mistakes or request changes)