Abstracts

RATIONALE: Automated spike detections methods often identify a number of differing classes of events. Some classes represent epileptiform spikes, others represent spike-like artifacts. The development of efficient single dipole source model algorithms provides the opportunity to conduct on-line assessment of some features of the dipole model generated from the raw waveforms. This approach may not provide accurate dipole localization, but may provide sufficient information to classify detections into [soquote]good[scquote] or [soquote]bad[scquote] categories.
METHODS: Automated spike detections from prolonged recordings (mean 17.6hr) from 6 patients were reviewed visually. These data were classified into the following categories.
1. SPIKES
2. UNCERTAIN EVENTS
2.1. Waveforms of uncertain significance
2.2. Normal physiological transients
3. ARTIFACTS
3.1. Eye artifacts
3.2. Muscle artifacts
3.3. Electrode artifacts
3.4. Clear artifacts of other, or combined sources
A dipole model was obtained at the maximum of the global field power within [plusminus]35ms of automatic detections. This 4-sphere algorithm can be implemented during data acquisition with current computing technology. The algorithm estimates dipole location and eccentricity (ECC), dipole orientation, residual variance (RV) and whether the solution converges. A separate algorithm is used to calculate signal to noise ratio (SNR).
These data were reviewed to identify parameters that could be used to differentiate detection categories.
RESULTS: There were 19850 automated detections. Visual analysis of those detections revealed that 80.3% represented spike events, 9.4% represented artifact detections and 10.3% represented events of uncertain significance or normal physiological transients.
The distribution of values for RV, SNR and ECC clearly differed between spike and artifact detections allowing detections to be differentiated according to these criteria. Of these, RV provided the best opportunity to remove a significant proportion of artifact detections while retaining most spike detections. Few dipole models of spikes had RVs greater than 20%. A threshold set at this value removed a mean of 34% of artifact detections, and a mean of only 1% of valid spike detections. By using this threshold, 1% of eye artifacts, 42% of muscle artifacts, 32% of electrode artifacts, and 45% of [soquote]other[scquote] artifacts were removed.
Analysis of the dipole generated variables also allowed clear classification of spikes according to location.
CONCLUSIONS: The use of an efficient algorithm for generation of single dipole models of raw waveforms of automatically detected events provides information that can be used to significantly reduce artifact detections. This algorithm may also provide information that will allow for automated spike classification.
Support: Stellate Systems.
Disclosure: Salary - Stellate Systems.