Abstracts

Rationale: Continuous EEG monitoring is an important tool to recognize clinically invisible deteriorations like non-convulsive seizure in critically ill patients. However, manually reviewing continuous EEG recordings requires substantial resources, which are often limited in intensive care units. A computational algorithm for automated analysis and trending of continuous EEG recordings from critically ill patients is being developed. The EEG is continuously evaluated according to the ACNS standardized critical care EEG terminology. The results of several hours are visualized graphically on a bedside monitor to allow quick assessment of state and trend of the patient. We compared the performance of this computational algorithm to an expert human reviewer. Methods: Continuous EEG recordings from 71 patients (4773 hours in total) from ICUs in three different clinics are used for algorithm development. In addition, 143 short term EEGs (82 hours in total) from 33 patients are used for quantitative analysis. The EEGs were analyzed by a clinical EEG expert according to the ACNS terminology and compared to the results from the algorithm computation. We define a detection as true positive if manual and computational analysis overlap within a range of 30 seconds before and after the marker. The computation first removes artifacts using the PureEEG (1) algorithm. Then EEGs are split into segments representing normal waves, discharges, or spikes. Localization and morphology of these segments are assessed, and finally the segments are reassembled into groups representing rhythmic activity, periodic discharges or spike-and-wave patterns (Main Term II). The localization of these patterns (Main Term I) together with frequency and amplitudes is determined. Results: The findings of the manual review by a clinical EEG expert showed that 34% of the patients had periodic discharges, 19% rhythmic delta, and 10% spike/wave patterns. The automatic detection algorithm agreed at 84% of the periodic discharges, 72% of rhythmic delta and 74% of the spike/wave patterns. In addition to the ACNS terminology, rhythmic theta and alpha patterns were analyzed and occurred at 6 and 3 percent of the patients respectively. The algorithm detected more than 92% of these patterns.Conclusions: A computational algorithm to monitor the neuronal activity of critical ill patients was assessed. The results show a high agreement between manual and computational analysis on Main Term II of the ACNS terminology. We see a high potential benefit from having automatic analysis of EEG-data for ICU. It will allow an extended coverage of critically ill patients with EEG monitoring in times of limited resources and budget. A more detailed analysis including several EEG experts will be the next step. (1) www.eeg-vienna.com