Abstracts

RATIONALE: We acquire up to 128 channels of EEG simultaneously, from both scalp and intracranial electrodes. To reduce the volume of data requiring interpretation, we have developed spike and seizure detection algorithms, until now only non-invasive. With a view towards implementation of a fully implantable brain stimulator to abort seizure activity, we have developed and tested a seizure detection algorithm for invasive electrodes. Although not implantable, we can evaluate our seizure detector and sub-thalamic nucleus (STN) stimulator in a fully closed-loop mode.
METHODS: All EEG channels are processed using the third level detail coefficients from the Daubechies-4 wavelet transform. After smoothing, the ouput power is compared to a threshold, which is constantly updated by a statistical measure. A minimum duration criterion and channel spatial relationship criteria are employed to reduce false triggering. The invasive seizure detection program runs concurrently with data acquisition/display and analyzes the EEG from subdural, depth, or STN electrodes in real-time. Detection of a seizure prompts the ordinary sequence of diagnostic actions (seizure onset marking, saving a predefined segment before and after seizure onset, and optional alarming), [underline]plus[/underline] triggering of an external switch to activate a controllable STN stimulator.
ROC curves were obtained to optimize the four adjustable parameters (number of taps on the smoothing filter, threshold multiplier coefficient, minimum duration above threshold, background forgetting factor) on a training set of 50 seizures acquired invasively from 6 patients. Validation was carried out retrospectively on a testing set of 101 EEG seizures and 60 clinical seizures from 11 other patients. True positive detections were based on a 25 - 400% overlap between computer detection and physician identification of seizure onset and end.
RESULTS: Of the EEG seizures, 76% were correctly identified, with 24 false positives (FP) and 8 false negatives (FN), with detection approximately 8 seconds after human marking. Clinical seizures were identified 75% of the time, with 15 FP and 6 FN, and a prediction time of 5 seconds before clinical onset.
CONCLUSIONS: This work differs from other efforts in that a variety of wavelets were tested for suitability, electrical field criteria improve accuracy, and the threshold is continuously statistically updated. In this research phase, the DBS stimulating leads are externalized so that we can record from them directly, the recording and analysis computer is external to the patient, and triggers the DBS through an external controller. After further refinement, seizure detection can be reduced to silicon and made a part of the implanted stimulator.
Support: Medtronic and Cleveland Clinic Foundation.
Disclosure: Grant - Medtronic