Abstracts

Rationale: Information on the frequency of epileptiform activity and its evolution over time can help in the evaluation and management of patients with status epilepticus (SE). Manual spike counting in patients with Electrical Status Epilepticus in Sleep (ESES) is subjective, labor-intensive and cannot be performed in real-time. We have developed a novel wavelet transform based algorithm for the automatic quantification of EEG spikes in patients with ESES. Our method improves previous attempts at spike quantification by using a double threshold model and inter-channel comparison to more effectively isolate spikes and reject artifact. Here we aim to test the ability of this method to quanitfy EEG epileptiform activity by comparing its performance against the results of gold-standard quantification by an experienced epileptologist.Methods: We applied our algorithm to three EEG sets in patients with ESES. Per EEG set, spike counts were obtained by the automated method and by visual review by a board-certified clinical neurophysiologist. In the first data set (N=5 300 sec samples from 5 patients) we assessed the sensitivity and specificity of the automated method. In the second data set (N=75 100 sec samples; 3 samples per each of 25 patients recorded in wakefulness, NREM 2, and NREM 3) we compared the manual and automated methods in each stage of the sleep-wake cycle. In the third data set (N=6, 6-10 samples from 6 patients) we assessed its capacity to detect lateralization of spike activity. Results: In the first data set, the automated method detected 882 of 1231 spikes marked by the human reviewer, yielding a sensitivity of 72%, with 95% confidence interval [0.62;0.82], and a specificity of 99.9%. In the second data set, automated and manual counting correlated in every stage of the sleep-wake cycle: Spike-wave percentage (mSWP: the manually counted percentage of one-second bins with at least one spike-wave complex) and spike frequency (mSF: manually counted number of spike-wave complexes per 100 sec) correlated in wakefulness, NREM2 , and NREM3 (Table 1). mSWP and automated spike frequency (aSF: algorithm counted number of spike-wave complexes per 100 sec) correlated in wakefulness, NREM2, and NREM3 sleep. mSF and aSF correlated in wakefulness, NREM2 sleep, and NREM3 sleep (Table 1). The percentage of spikes detected by aSF (sensitivity) was 94.2% of the one-second bins with spikes detected by SWP, and 54.5% of spikes detected with mSF. In the third data set, automated analysis lateralized spiking to the correct side in all cases, detecting1836 of the 2,409 spikes marked by the human reviewer, yielding sensitivity of 76%. Specificity was 88% (Table 2).Conclusions: Our data demonstrate a high correlation between the automated and manual methods for quantification of epileptiform activity. Our wavelet transform algorithm can be feasibly applied in clinical practice to allow for rapid and objective quantification of epileptiform activity, leading to patient-tailored evaluation. Future improvements include application of machine learning algorithms to increase sensitivity of detection.