T70. Automatic clustering of EEG data from ICU patients

Abstract

Introduction Seizures, status epilepticus, and seizure-like rhythmic or periodic activity are common, pathological, and harmful states of brain electrical activity seen in the electroencephalogram (EEG) of patients during critical medical illnesses or acute brain injury. A growing body of evidence shows that these states, when prolonged, cause neurological injury. In this study, we aimed to develop a valid method to automatically discover a small number of homogeneous pattern clusters, to facilitate efficient interactive labelling by EEG experts. Methods In this study, we analysed continuous EEG recordings from 100 ICU patients at MGH. The duration of each recording is at least 24 h, with a sampling rate of 200 Hz. We extracted 592 time and frequency domain features from each EEG recording. Principal component analysis (PCA) with 95% variance retained to reduce the dimensionality for each feature array. We further segment the EEG based on change point detection on spectrogram. In each segment, we associate each 1 s epoch to its nearest exemplar found in a dictionary pre-defined using Textons, and compute the normalized histogram. It is followed by unsupervised clustering method Affinity Propagation, to further split the data into various clusters. Three experts independently labelled all clusters into one of 6 standard pattern categories (seizures, GPDs, LPDs, LRDA, GRDA, and Other). We compared interrater agreement (IRA) indexed by Gwet’s AC1 among experts vs. between each expert and consensus labels using two methods for labelling clusters: (1) “Labour intensive labelling” (LIL): assign the most frequent of 30 expert provided labels; (2) “Labour efficient labelling “(LEL): assign the most frequent of the 3 expert labels for the central sample. Finally, we used the embedding method t-SNE to visualize the data. Results Median [IQR] expert-expert IRA for all label pairs across subjects was 0.65 [0.58, 0.75]. IRA for individual expert labels and the final consensus label was 0.76 [0.70, 0.82] using LIL, and 0.71 [0.63, 0.78] using LEL. Differences between LIL and LEL were not statistically significant (p = 0.34). The t-SNE visualizations of the feature space generally revealed interictal-ictal continuum. Conclusion This research suggests that large EEG datasets can be automatically clustered into a small number of patterns described by standard ICU EEG patterns. We demonstrated efficient cluster labelling by inspecting only the central most representative of each cluster. Furthermore, t-SNE visualizations support the hypothesis of an interictal-ictal continuum.