Abstracts

Rationale: The real-time detection of seizures using scalp electroencephalograms (EEG) has been intensively researched upon. In effect, a low cost, non-invasive, reliable, robust and easy-to-use automated seizure detector can be a very useful addition to the monitoring capabilities of any emergency vehicles, emergency departments, field hospitals, etc. Our group has been involved in the development of such a device, referred to as NeuroMedic^TM. At its core, the NeuroMedic embeds an innovative population-normed wavelet-based seizure detection algorithm. The purpose of the present study was to assess its real-time performance using a limited number of EEG channels (2, 4, and 8) from a comprehensive database of clinical EEG recordings. Methods: After obtaining institutional approval, 126 seizure records from patients monitored at the Cleveland Clinic Foundation (CCF) using the standard 10-20 montage were obtained. Two channels were selected from each of the frontal, parietal, occipital and temporal region of the brain. Each EEG record contains periods of ictal activity, and/or no-seizure activity. All the records were reviewed by an EEG technologist who annotated the start and end times of each seizure based on off-line review of the data. A subset of the database was also reviewed by another independent EEG technologist. The sensitivity (resp. specificity) of the first marker with respect to the second marker was 85% (80%). In addition to the CCF database, 37 1-hour long sleep data recordings were added to the database in order to accurately calculate the algorithm specificity. The output of the NeuroMedic algorithm is an index that increases during ictal activity and resets to 0 during normal EEG activity. A seizure flag is triggered whenever the index crosses a pre-defined threshold. In order to assess the performance of the algorithm, we derived the Receiver Operator Curve (ROC) corresponding to various thresholds and for various channel configurations. The 2-channel configuration uses only 2 frontal channels (FT9-Fz and FT10-Fz). The 4-channel configuration adds 2 parietal channels (P3-Fz and P4-Fz). The 8-channel configuration further expands on the 4-channel configuration and adds temporal and occipital channels (T7-Fz ; T8-Fz ; O1-Fz and O2-Fz). Results: Results are summarized in Figure 1. By targeting a sensitivity of 85%, we can achieve a specificity of 85% (8-channel), 84% (4-channel), and 76% (2-channel). The area under the curve was 0.92 (8-channel), 0.89 (4-channel) and 0.88 (2-channel). As expected, adding more channels improves the performance of the algorithm. Conclusions: The NeuroMedic algorithm is providing a level of performance comparable to that of a trained human operator. In addition, the detection is performed in real-time and uses a limited number of channels. Using a ruggedized and portable EEG acquisition device, the NeuroMedic can be a useful adjunct in emergency medicine. We expect to further improve these results by combining an automated artifact detection and filtering method (see, e.g., [1]). [1] Zikov et al. "A wavelet-based de-noising technique...", IEEE EMBS, 2002.