Abstracts

Rationale: Seizures are not always accompanied by convulsions and outward signs, especially after the patient has been treated with drugs. The electroencephalogram (EEG) can provide accurate seizure identification, thereby enabling prompt and aggressive treatment within a time-critical window to prevent permanent disability and death. However, emergency personnel typically cannot interpret the EEG, requiring the consultation of a neurologist. We have developed a Seizure Vector (SV) algorithm to analyze EEG in real-time and automatically identify ictal activity. This algorithm can be incorporated into a seizure monitor to provide automated screening of seizure activity, without expert consultation. Methods: SV is an EEG “digital fingerprint” based on 5 spectral and temporal features, specifically selected based on their response to seizure activity. This combinatorial method yields a 5-dimensional vector designed to differentiate ictal from normal activity. EEGs were collected from 40 human adults admitted to Johns Hopkins Hospital with a variety of seizure types. A sixteen channel recording montage was used. Each recording lasted at least 20 minutes in length. A blinded epileptologist classified the EEG into “normal” or “seizure” categories in 15 second epochs, on a channel by channel basis. A total of 2035 episodes of seizure and 3867 episodes of normal data were used. Half of this data set was used to train each feature and the combined SV, storing mean values for each category, and the other half of the EEG database was used to evaluate the discrimination performance of each of the features and the combined SV. Results: EEG epochs from patients were used to test the ability to discriminate normal vs. seizure for (1) each of the five selected feature sets, and (2) SV algorithm combining all five feature sets. We used Receiver Operator Characteristic (ROC) curves to demonstrate the relative ability of different classifiers to accurately discriminate one pattern from another. SV’s combination of features yields an Area Under the Curve (AUC) that is better than individual feature classes for the human data. A comparison from the second highest individual feature to the Combination represents the ability to detect differences between classifiers with power of 0.8 at a significance level (alpha) of 0.05. Similarly, the comparison of the best performing individual feature to the Combination has a power of >.99 and alpha=.01 to detect the differences between classifiers. These increases reflect the fact that the combination of features has better discriminative ability than any single feature. The optimized SV algorithm differentiated between “seizure” and “normal” with 95.0% sensitivity and 95.2% specificity, with the Combination vector outperforming each of the individual features. Conclusions: These results show that the optimized SV algorithm can provide the necessary ictal waveform discrimination capabilities for potential use in a seizure screening system to be used in the field or in a hospital.