Persyst Application Notes:
Optimizing Spike and Seizure Detection for Long Term Monitoring and Event
Notifications II
Introduction
Persyst spike and seizure detection has advanced dramatically since the first application note on this topic was published in the spring 2000 Insights. Since that time, Reveal has replaced our previous SpikeDetector product and larger data sets and better neural network training methods have been developed. Because a “perfect” spike and seizure detection algorithm aims to be one that is indistinguishable from a panel of human experts, detection performance is affected by the same factors that affect human experts. Fortunately, all of these factors can be controlled, and are fundamental to any good EEG Recording. They are:
1) Recording quality.
2) Montage selection.
3) Channel selection.
4) Detection parameters.
In an era of diminishing reimbursements, performance has become more important than ever; consider that even a “free” spike and seizure detector may be no bargain at all if the performance is lacking, because revenue gained from providing “digital analysis of epileptiform activity” can be quickly consumed with time lost dealing with poor sensitivity (missed events) or high false positive rate. Performance measures made with small data sets are of little help in predicting real-world performance (an example of a large data set is provided here) and the more difficult the recordings, the more pronounced any performance differences will become. The net of this is that best performance is achieved when internal performance factors (algorithm performance) and the external performance factors listed above are properly managed.
Recording Quality
All EEG labs take measures to obtain the best quality recordings possible—extremely noisy recordings with bad electrodes are difficult for humans and automated detection systems to evaluate. Still, when presented with an unavoidably noisy recording SpikeReview provides unique capabilities for extracting information via hierarchical clustering of events that share similar topology and/or morphology (see Introduction to Hierarchical Clustering, Insights, Spring 2001).
Amplifier reference contamination is important to check for as well, and is easily done. Remember the Bio-Cal used with analog EEG to check the dynamic response of the pen writers? All of the channels were set to “FP1-O2”, and were examined to see that all pens were responding equally to a biological signal.
Today, performing a traditional Bio-Cal on modern digital EEG systems is considered by many to be insufficient; with the advent of modern referential digital EEG amplifiers, elimination of multiple discrete amplifier stages, and the elimination of mechanical pens, a traditional Bio-Cal will often test just two amplifier channels: FP1-Ref and O2-Ref. Digital EEG introduces dynamics common to all EEG channels that were not present in analog systems: a common amplifier reference. In a digital EEG amplifier, the amplifier reference is Grid 2 for all of the EEG electrodes. Also important (but not new) is the amplifier “ground” electrode which usually determines where the amplifier “sits” electrically vis-à-vis the patient.
So why is amplifier reference contamination important? Because remontaging the referential EEG into a bipolar (or traditional Bio-Cal) montage can obscure problems with amplifier reference and/or amplifier ground that will contaminate the EEG, because the reference is subtracted from G1 and G2 in a bipolar montage. The result can include unusual artifact that can confound analysis by human experts and automated detection alike.
The solution is thankfully quite simple; include a raw referential montage with filters “off” (or fairly “wide-open”) as a Digital Bio-Cal to supplement any traditional “FP1-O2” Bio-Cal at the start of the recording. Any problems with reference contamination will be easily spotted as artifact common to all channels, and individual bad electrodes can be spotted quickly as well. Reapplying the reference will often clean things up immediately, and can bring down the impedance of all the EEG electrodes at once, too.
Montage
Proper montage selection is fundamental to getting good results, yet this is easily overlooked because traces can be viewed on the recording system in a montage completely different from the one selected for detection.
However, this is easily addressed simply by having a look at the EEG with the montage used for spike detection (or the “Digital Bio-Cal” described above). If there is significant artifact obscuring the EEG throughout, the problem can be addressed directly by fixing the problem electrodes. The importance of high-quality recordings is common ground among expert EEGers, as dislodged/high impedance electrode connections rarely yield clinically useful EEG.
Detection Montages for Long-Term Monitoring Recordings
· If you are using standard 10-20 channel labels, then the “Default Scalp” spike and seizure detection protocol will be best—no other changes or adjustments to detection settings should be required.
· If you are using additional channels, e.g., a mix of 10-10 and 10-20 channels, a detection montage that uses your custom channels should be used. No other changes or adjustments to detection settings should be required.
· If you are recording grid and/or strip electrodes, a detection montage that uses these custom channels should be used. Reveal’s Automatic Sensitivity will set the spike detection gain (µV/mm) for you automatically. (Version 2004.01.28 or later includes automatic sensitivity for seizure detection as well.)
Detection Montages for Ambulatory Recordings
Ambulatory recordings present a different challenge. Despite heroic efforts to secure electrodes to the scalp, sometimes one or more electrodes become dislodged after the patient leaves the EEG lab. The good news is that the same technique used to read problem recordings can be used to perform spike detection with the best results: Use a reference that provides the best view of the EEG, and turn off channels that have become dislodged early in the recording. Here’s how:
· Take a look at the first page, the last page, and a page in the middle of the Ambulatory EEG recording to get a quick “birds-eye view” of how well the electrode connections fared.
· Next, selecting a montage for scanning ambulatory recordings is as easy as selecting one for review. If the default Laplacian montage looks good, just scan using the “Default Scalp” protocol. If the amplifier reference looks better, select the “Default Ref” protocol instead.
After launching Reveal, select Settings, then Protocol. The Spike Montage tab lists the channels used for detection. To change the reference, select the first and last channel with the mouse pointer while holding down the Shift key. Next, click the Reference button and select the same “clean” reference used for review.
Reveal Spike and Seizure Detection Montages
Reveal uses different montages for spike detection and rhythmic burst detection for optimum performance:
· Reveal’s default spike detection montage uses a Laplacian reference (i.e., each electrode is referenced to an average of the electrodes directly surrounding it). A Laplacian reference shares benefits with bipolar and referential montages—it is effective for minimizing the effects of moderate artifact (vis-à-vis bipolar montages), and allows comparison of amplitude between channels for mapping (vis-à-vis referential montages).
· Reveal’s default rhythmic burst detection montage is bipolar, because voltage mapping is not required, and bipolar montages perform well for minimizing and isolating the effects of moderate artifact.
Reveal Artifact Detection Channels for Scalp Recordings
Montages for spike and rhythmic burst detection should include artifact-detection channels, e.g., FP1 and FP2 are used for eyeblink rejection, and temporal channels have a “Temporal” code assigned to reject muscle artifact. Reveal uses these channels to improve artifact rejection, so they should not be “switched off” in your scalp EEG scanning protocols. (Version 2004.01.16 and later considers the complete topology for 10-20 recordings, replacing the use of eyeblink/temporal codes.)
Effects of Montage Selection
The following example illustrates how montage selection and recording quality can affect detection. Using an amplifier reference we see overwhelming, continuous artifact at P3, A1, and A2. The continuous artifact is confined to those channels, leaving the other cerebral channels unaffected—this recording should be scanned using the amplifier reference for spike detection.
Figure 1
The same page viewed with the Av17 montage (average of F7, Fz, F3, F4, F8, T3, C3, Cz, C4, T4, T5, P3, Pz, P4, T6, O1, O2). The overwhelming artifact at P3 has contaminated all of the other channels:
Figure 2
And finally, with a Laplacian reference. Note that while most channels remain unaffected by the artifact in P3, the Laplacian references that include P3 in their average (T5-aT5, C3-aC3 (and P3-aP3). Because each electrode in a Laplacian reference only contributes a fraction of the grid-2 weight (e.g., P3 is ¼ of aT5), moderate artifact in any single channel will largely be “averaged out”—unless it is continuous, overwhelming artifact.
Figure 3
T5-aT5 Laplacian reference includes T3,
C3, P3, and O1, each contributing ¼
of the grid-2 reference:
Figure 4
C3-aC3 Laplacian reference includes F3,
T3, P3, and Cz, each contributing ¼
of the grid-2 reference:
Figure 5
Why use a referential montage for spike detection?
First, a referential montage offers better topographic mapping (and clustering for SpikeReview) than a bipolar montage. (To understand why bipolar montages are difficult for topographic mapping, consider this question. What value should be plotted at Fp2 on a topographic voltage map when the montage contains both Fp2-F4 and Fp2-F8?)
Secondly, the average reference removes the possibility that an “active reference” is used. For example, you don’t want to use an A1+A2 reference for patients with temporal spikes because this causes maximum deflection away from the spike focus. Thus, Reveal’s default scanning montage, which includes 10-20 electrodes and uses a Laplacian reference, is a good choice when there is no a priori knowledge about the epileptogenic focus.
The Laplacian average reference performs well in environments where low-to-moderate artifact can be expected (e.g., any in-lab or inpatient scalp EEG recording), because each channel in a Laplacian montage is referenced to an average of the electrodes surrounding it. With each surrounding electrode contributing ¼ of the average, moderate artifact is largely “averaged out”.
There is, of course, a limit to how much artifact a 4-channel Laplacian average reference will average out; figure 2 shows how a single bad channel can completely contaminate a 17-channel average reference as well. Given the limits of average references for moderating artifact, the Laplacian again shows its advantage by moderating reference contamination through averaging, while keeping the effects of a bad channel localized if the average is overwhelmed.
Therefore, when scanning ambulatory recordings with severe artifact in one or more channels, an amplifier “Ref” (or other quiet reference location) will yield better performance.
Montage Conclusions
· The spike and rhythmic burst detectors work best when they can “see” the EEG activity clearly—if you’re have trouble reading the EEG, the detectors will have trouble reading it too. If this is the case, using a detection montage that gives you the best view of the EEG will ensure that Reveal’s spike and seizure detectors will have a good view of the EEG as well.
· Ambulatory recordings are usually scanned off-line, and if so, this can be used to your advantage. Take a quick look at the first few pages, the last few pages, and a few pages in the middle—did the electrodes remain attached during the recording? If so, the Default Scalp detection protocol will be optimal. If one or more electrodes were completely dislodged during the recording, an amplifier reference for spike detection will perform better. (And if enough electrodes along with the amplifier reference have come off, then at that point it’s no longer an EEG recording.)
Figure 6
In the “Bad P3” recording example(figure 6), we’ve selected all of the EEG channels (Shift-mouse select). Next, right-click on the Montage Bar and select Change Reference to select any EEG channel, average ref, etc. (we’ve selected “Ref”). Save the new montage and it can be used as a spike detection montage in Reveal.
Figure 7
A quick “one time” reference change can be done right from Reveal(figure 7). From the Reveal menu, select Settings, then Protocol. We’ve selected all of the EEG channels (Shift-mouse select). Next click the Reference button to select any EEG channel, average ref, etc. (we’ve selected “Ref”).
Channel Selection
· Scan the channels that are being recorded:
o If a mix of 10-10 and 10-20 electrode placements are being used, create and save a montage that includes all of the cephalic channels that are being recorded, and use that montage for scanning.
o If scanning grids, strips, or a mix of both, save the “Recorded” montage with the patient’s name, and use that montage for scanning.
· Don’t scan channels that are not recorded:
o Empty channels have loads of artifact, so de-select these in Reveal (from Reveal’s menu select Settings|Protocol), or create a montage with only the channels used.
· Record and include FP1, FP2, and temporal channels in your scalp protocols—these are used for detecting and rejecting spike-like artifact.
Detection Parameters
Reveal provides one simple adjustment for spike detection sensitivity, and the default “0.5” setting (on a scale of 0.1 to 1.0) will provide the optimum sensitivity vs. false-positive rate. This can be adjusted for special applications if needed. The ROC (receiver operating characteristics) graph below was developed with an extremely large data set, and is a useful predictor of performance.
The default Spike Burst and Rhythmic Burst detection sensitivity is at the maximum “0.1” sensitivity by default, and will provide optimum performance for all clinical EEG recordings.
Reveal provides separate detection and notification sensitivities (select Settings|Protocol from the Reveal menu). In most cases only the Rhythmic Burst “Notify” Duration Minimum would be changed (e.g., to adjust notifications when monitoring patients that have an unusual degree of inter-ictal rhythmic activity).
Figure 8
Reveal spike and seizure detection performance (ROC) curves:
Figure 9: Tested on 18,503 spikes in 266 hours of EEG. The SDv5 and SDv6 (Reveal Generation III) curves are generated by applying a perception thresholds of 0.1, …, 1.0. (Highest sensitivity and FP/min at threshold=0.1). The Sensa curve was created by varying the amplitude threshold from 4 to 6.
Figure 10:ROC (receiver operating characteristic) curve for the Reveal algorithm and the single point values for Sensa and CNet. The curve was generated by increasing the perception threshold from 0.1 to 0.9, which has the result of lowering both the sensitivity and false positive rate. The sensitivities are calculated on 670 seizures from 426 epilepsy patients for a total of 1,049 hours of EEG. The false positive rates are calculated on 33 records from 33 patients deemed to not have epilepsy for a total of 465 hours of non-seizure EEG.
Conclusion
The process of recognizing spikes and rhythmic bursts is both perceptual and probabilistic, and humans are equipped with pattern-recognition “neural networks” more powerful than what can be brought to bear by even the most elaborate supercomputer. Nevertheless, the ability of human experts to recognize clinically important patterns can be confounded by excessive artifact and inappropriate presentation of data (e.g., poor montage/gain/filter selection).
Regardless of review and analysis method used, maintaining recording quality to the extent possible is job #1. After that, appropriate selection of scanning montage, reference, and channels will ensure the best possible performance for human experts—and computer neural networks alike.
Relevant Publications
Seizure detection: correlation of human experts, Wilson et al., Clin Neuro., 114 (2003) 2156-2164 (reprints available).
Spike detection: a review and comparison of algorithms, Wilson et al., Clin Neuro., 113 (2002), 1873-1881 (reprints available).
Spike Detection II: automatic, perception-based detection and clustering, Wilson et al., Clin. Neuro., 110 (1999), 404-411 (reprints available).
Introduction to Hierarchical Clustering, Guess et al, J. Clin. Neurophys. 2002 (reprints available).
Confirmation of Two Magnetoencephalographic Epileptic Foci by Invasive Monitoring from Subdural Electrodes in an Adolescent with Right Frontocentral Epilepsy, Otsubo et al., Epilepsia, (1998), 608-613 (reprints available).
Computerized brain-surface voltage topographic mapping for localization of intracranial spikes from electrocorticography, Otsubo et al, Journal of Neurosugery, (2001) 1005-1009.