Abstracts

Rationale: Computer assisted EEG interpretation in the ICU strongly assists the use of continuous EEG monitoring in critically ill patients. Its main goal is to detect, at an early stage, derangements in brain function, in particular (non-convulsive) seizure activity and ischaemia, allowing a window of opportunity for interventions. Methods: Eight quantitative features were extracted from the raw EEG and subsequently combined into a single classifier. Features included mean power, Brain Symmetry Index, Burst-Suppression ratio and periodicity. For classification, a decision tree was used. The system was trained by using 41 EEG recordings. All EEG recordings were visually assessed by an experienced electroencephalographer. Patterns included normal, iso-electric, low voltage, burst suppression, hypofunctional, generalized periodic discharges and seizure activity. For evaluation, 20 independent EEG recordings were used. After real-time implementation of the classifier in NeuroCenter EEG (Clinical Science Systems, Netherlands), clinical evaluation in ICU patients was performed. Results: 36 (88%) epochs of the training set and 17 (85%) epochs of the test set were classified correctly. Implementation in the NeuroCenter EEG monitor allowed real-time analysis in the ICU. The user interface presents both trend-curves and text output. At present, we have used the system in over 20 patients in our ICU. In approximately 70% of the patients, the computer interpretation showed good correspondence with the visual evaluation. In the remaining ~ 30% discrepancies were observed. Part of these were caused by artifacts, were the remainder was attributed to erroneous classification. The use of continuous EEG registrations and the real-time classification provided essential information for the clinical decision making in a substantial number of patients. Conclusions: We present our first implementation of a real-time computer assisted interpretation of EEG patterns as observed in the ICU, based on a combination of eight features. At present, the system still underperforms as compared to an experienced electroencephalographer, with a classification accuracy of approximately 70%. At present, we are working towards a further improvement of the classifier, including the use of computer assisted analysis of video to assist in the detection of various artifacts.