Abstracts

RATIONALE: Several automated seizure detection systems have demonstrated reliable operation over extended periods of time. However, expensive computational and storage requirements, such as those of rule bases and neural networks, make these systems unsuitable for implementation on environments less powerful than personal computers. Olsen et al. 1994 employed 9 signal features and a logistic transformation (tested on 7.5 hours), Qu & Gotman 1997 employed 6 features and nearest-neighbor calculations (tested on 32 hours), whereas Osorio et al. 1998 used a single bandpass-filtered amplitude detector involving convolution, squaring, and median filtering (tested on 55.5 hours). Whether multi-feature seizure detection algorithms could be simplified so much as to allow implementation in the severely limited environment of an implantable device, yet perform robustly, has not been addressed before.
METHODS: We developed tools for validating a seizure detection system based on the following simplified features: half-wave amplitude and duration, sum of absolute differences as an approximation of signal curvelength (in turn a simplification of waveform fractal dimension), and a modified sum of absolute amplitudes as an approximation of signal energy. The detection system for a given patient may employ one or more of the features, and is tuned using special heuristics and/or an automatic optimizer from a training set of 1-3 recorded seizures and typically 9 baselines (3-minute epochs) under various vigilance states. As of this writing, we have trained detectors using only about 30 minutes of intracranial EEG data per patient, and have tested the system over a total of 623.8 hours representing the entire hospital-stay recordings spanning 3-12 days for each of 4 consecutive resective patients with MTLE.
RESULTS: The table summarizes false-negative rate, false-positives-per-hour, average latency in seconds with respect to expertly-marked electrographic onsets, and size of the data sets.
CONCLUSIONS: Seizure detection accuracy exceeding current benchmarks over entire hospital stays for each patient was possible based on a single training session with a relatively small data set, using simplified features suitable for an implantable device. The robustness obtained from combining features was verified: these features often work in synergy by mutually screening each others[ssquote] FPs.[table]