F13. Concordant EEG signals recorded with conventional and Ceribell EEG systems

Abstract

Introduction Rapidly obtaining EEG signals in the emergency departments and ICU for high-risk patients can enhance diagnosis accuracy and speed, while cutting down time until treatment. Currently the use of EEGs in emergency settings is limited by technologist availability and lengthy setup time ( ∼ 30 min). Ceribell, a startup company based in Mountain View, California, has developed a portable EEG data recorder and electrode headset with rapid setup ( ∼ 5 min) technology designed to overcome the inaccessibility of EEG in urgent situations when non-convulsive and subclinical seizures are suspected. The purpose of this study is to evaluate the signal quality and performance of the Ceribell system compared to a reputable clinical EEG system. Methods We collected EEG samples both in the laboratory and as part of an ongoing clinical pilot study at Stanford University Medical Center. Laboratory collections on healthy volunteers included simultaneous collection of EEG using Ceribell and Nihon Kohden systems, as well as a split signal that recorded the EEG to both data recorders from the same set of electrodes. For these setups since the recordings were simultaneous, the signals were expected to be very similar in the time domain. In the clinical setting, EEG data was first recorded with the Ceribell system in the ICU on 22 patients and subsequently with the clinical EEG system on the same patients. The data was pre-processed with a [1 Hz, 70 Hz] band-pass filter. The following metrics were used to characterize signal concordance: Hjorth Activity, Hjorth Mobility, Hjorth Complexity, artifact count per unit of time, (defined as # of samples outside of mean ± 6 standard deviation), baseline wander, power of 60 Hz noise, and Kurtosis. At the end, Wilcoxon signed-rank test was run on each metric pair (Ceribell vs. NK) to assess whether the metrics calculated for Ceribell data are significantly different from the metrics calculated for NK data. Results Our findings confirmed that for simultaneous recordings, time domain signals obtained from Ceribell and Nihon Kohden systems closely match each other. For non-simultaneous recordings, results of Wilcoxon signed-rank test confirmed a lack of difference (p > 0.05) in all except one metric. The p-values for different measures were as follows: Hjorth Activity: 0.4264; Hjorth Mobility: 0.0883; Hjorth Complexity: 0.4264; artifact count: 0.2912; baseline wander: 0.1579; power of 60 Hz noise: 0.0005; and Kurtosis: 0.4455. For power of 60 Hz noise, NK recordings have significantly higher 60-Hz noise. For reference, the average power of 60 Hz noise across all subjects and all channels on NK recordings is 764.98 compared to 16.55 for Ceribell recordings. Conclusion Our results indicate that the signal quality of the Ceribell system is similar, and superior in the amount of 60 Hz noise to the Nihon Kohden system.