Clustering Seizures to Find Groupings With Common Clinical Characteristics

Abstract

Seizure clustering and classification are open-ended problems in neurophysiology. The aim of this work was to cluster seizures using unsupervised methods to create groupings that share clinical characteristics. Efficient and accurate clustering of seizures allows for automated labeling and the ability to make inferences about clinical characteristics of a seizure given only an electrocorticographic (ECoG) signal representing the event. Three different feature sets were computed for clustering: (1) A set of 25 of time and frequency domain features commonly used for seizure detection and prediction; (2) Means specifying a latent distribution constructed by a variational autoencoder, and (3) coefficients for a delay differential analysis (DDA) model developed for ECoG data. These three feature sets were clustered using the K-means algorithm once the ’correct’ value for K was estimated using the elbow method, silhouette score, Davies-Bouldin score, and the Bayesian Information Criterion. The Affinity Propagation algorithm was also used to create clusters. The resultant groupings formed by each clustering were then compared to labels manually assigned to each seizure: which patient the seizure originated from, the onset region of the seizure, the visual pattern of the seizure, and its etiology. The Adjusted Rand Index was used to quantify the similarity of these labels to the six different groups of clusters produced. While none of these methods was shown to be able to reliably sort seizures sharing certain characteristics into distinct clusters, TFD clusterings and DDA clusterings more consistently sorted seizures into groups corresponding to labeled clinical information.