Abstracts

RATIONALE: The advent of implantable devices for detection, termination, prediction, and prevention of seizures is becoming a reality. The goal of this study was to evaluate the NP seizure onset detection system in a simulated on-line operation, under conditions resembling the clinical scenario that an implantable device for automated seizure detection/termination would face.
METHODS: Intracranial EEG (IEEG) signals available from recordings taken during the hospitalization of four patients under evaluation for resective surgery were analyzed in a prospective simulation with NP algorithm and an adaptive threshold classifier. All patients had medically intractable partial seizures of predominantly mesial or neocortical temporal origin. A total of 332.2 hours spanning the full-hospitalization stay were used in the simulation. An epilepsy expert properly marked electrographic onsets following the UEO (unequivocal electrographic onset) definition of Litt B. et al. 1999. The system consists of four main blocks, the early detection block, the data collection, the parameter optimization, and the user-interface. Only the line length tool (IEEE-EMBC 2001) from the tools available in the early detection block was evaluated at this point. The system is initialized by collecting a variety of IEEG segments and adjusting its parameters as a highly sensitive detector. After initialization IEEG segments are collected in response to each detection. As expected in a real clinical situation following device implantation, initially every day, the physician accesses the system and labels collected data as seizure or baseline. Labeled IEEG segments are used by the parameter optimization block to retune the adaptive threshold. It is expected that as the hospitalization time progresses the adaptive threshold will be adjusted to decrease the false positives per hour (FPh) over time, so the threshold optimization is required less frequently.
RESULTS: All clinical and subclinical seizures were detected with the exception of two brief focal subclinical seizures. An overall rate of 0.054 FPh (worst case including postictal detections and flat artifact detections was 0.13 FPh), sensitivity of 92.6% (only 2 FNs which corresponded to two sub-clinical focal seizures, 6 second duration, out of 27 seizures), and an average UEO detection of 2.9 seconds with a range of [-44.18 sec., 7.86 sec.] were observed. The evolution of the FPh over time was assessed by comparing the FPh during the first 12h for each patient with respect to the FPh in the last 12h. The average reduction of the FPh for all patients was 91.6% corresponding to an average of 0.25 FPh in the first 12h and 0.021 FPh in the last 12 hours. The FPs exhibit a clear tendency to decrease over time as the system [dsquote]learns[dsquote]; the patient signals and is tuned to them.
CONCLUSIONS: The adaptive classification scheme combined with the line length tool tuned to each patient as data were collected demonstrated an outstanding performance in the prospective simulation conducted. Further evaluation is under way to validate this and the other NP tools. This study demonstrates the plausibility of an automated seizure detection system for an implantable device under the causality constraint required for any real-world scenario and illustrates how the adaptation over time can lead to better performance.
[Supported by: NeuroPace, Inc.]; (Disclosure: Salary - R. Esteller, J. Echauz, B. Pless, and T. Tcheng, are currently full-time employees of NeuroPace. Their salaries are confidential information., Stock - R. Esteller, J. Echauz, B. Pless, T. Tcheng, and B. Litt have been awarded stock options (each less than 0.25% of the company s total value) in NeuroPace Inc, resulting from licensure of patents to the company. These patents are all owned, singly)