Abstracts

Rationale: The urgent need to detect and treat Non-Convulsive Status Epilepticus (NCSE) in patients with coma or altered behavior has been widely recognized. The objective is to develop automatic EEG analysis tools for use on unconscious or comatose patients in hospital ICU and emergency departments. Their purpose is to monitor brain state and provide warnings, by analysis of EEG, that the patient's brain is in a potentially dangerous state of NCSE. Methods: The proposed hybrid system is a combination of two Artificial Neural Networks (ANN) and rule-based algorithms. ANN-1 is trained to detect paroxysmal epileptiform activity in 1-sec epoch and distinguish it from various non-epileptiform EEG patterns and artifacts. ANN-2 is trained to classify 1-minute EEG characteristics or EEG states. In order to enhance ANN-2 detection of NCSE we have included four other EEG states (SLOW, FAST, BSP and ARTIFACT) that are either commonly seen in EEG from patients in coma, or need to be distinguished from the epileptiform activity seen in NCSE. Of these states, the most frequently observed is SLOW, with theta or delta activity. FAST is associated with higher frequencies such as alpha-beta activity. BSP (burst suppression) indicates severe depression of brain function in coma. Artifacts need to be detected to avoid errors in classification of EEG states. Results: The developed algorithms were tested on recordings ranging from 3 hours (Patient # 4) to 84 hours (Patient # 5). The total of 241 hours (14479 minutes) of EEG data was recorded from 10 patients. These recordings include 5 comatose patients and 5 patients admitted for long term epilepsy monitoring. First four patients were diagnosed clinically and electroencephalographically as NCSE. Patient #5 was an ICU patient in coma without signs of NCSE and the length of study was approximately 4 days. Patients 5-10 were used to evaluate the classifier s specificity and false positive rate. One minute epochs from 9 training and 10 test records were expertly scored into one of the five EEG states listed above. Training of both ANNs was done using 9 EEG records from patients not included in the test set. Out of 14479 minutes of EEG, 2883 minutes were marked by expert as NCSE minutes. The algorithm correctly classified 2762 NCSE minutes (Sensitivity = 95%) and 11175 non-NCSE minutes (Specificity=96%). NCSE by an often-accepted definition lasts at least 30 minutes. Figure 1 is a running sum of the proportion of one- minute NCSE detections in a 30-minute window from 10 patients. Note that the 4 NCSE recordings exceed 23 detections/30-minute window most (78%) of the time. In contradistinction, the 6 non-NCSE patients showed only brief periods of paroxysmal activity, never exceeding 21/30 (0.7). The separation of NCSE from non-NCSE patients was surprisingly robust, given the small size of the training data set and the lack of full optimization of the detection algorithm. Conclusions: These findings suggest the potential for highly accurate detection of NCSE in the unconscious or comatose patients. (Partially supported by NIH/NINDS SBIR Grant R44NS039214).