|
Advanced EEG Analysis
- Corticography/Topographic Analysis
- Universal EEG Review
- ICU and EMU Long Term Monitoring
- Spike and Seizure Detection
- Reporting
|
Introduction to Hierarchical Clustering
Introduction
Historically, reviewing inter-ictal (spike) events in EEG recordings has presented a challenge quite different from reviewing ictal (spike burst or rhythmic burst) events. Providing tools to review ictal events is relatively straightforward: they are relatively infrequent, so a navigation bar can be provided with indications of where events are located, and the clinician can click on the event to jump to the event itself. Once there, the clinician can examine the ictal event to determine morphology, onset, offset, etc.
|
Reference Quick Links:
|
The review of spike events is quite different. Thousands may be scattered throughout a recording. When reviewing a recording without the assistance of automated event detection the reader may use a variety of methods to determine if there are multiple focal points, to find the relative number of spikes in the recording, and to reject spike-like artifact upon closer examination.
Because of the complexity of this exercise, translating what humans do with the rather amazing neural-network in the brain has been something of a challenge. Rather than “yes or no” (dichotomous) evaluations employed by some spike detection programs, humans are capable of perception-based evaluations, i.e., rating something on a scale of 1 to 10, and taking into consideration a wide variety of factors and circumstances when making this determination.
By using a software-based neural-network algorithm, SpikeDetector can apply perception-based marks to spikes based on a number of factors, e.g., height, morphology, number of channels, surrounding activity, etc. In this way SpikeDetector will assign a low number to small or ambiguous spikes (0.1) and a large number to unambiguous spikes (1.0). By rating spikes on a scale of 0.1 to 1.0 we are able to use histograms to quickly show the nature of the spikes within any particular group (i.e., how well the spikes in the group are formed). With information collected for each spike that is marked (sign, focus, height, duration, angle, perception, etc.) it is then possible for the computer to find spikes that are alike (because of topology and/or morphology) and present them to the clinician to decide if they are spikes or artifact as a group.
By freeing the clinician from having to arduously review spikes one-at-a-time, the time needed to evaluate a recording is reduced. Also, because the Review Wizard will show if there are significant differences among the spikes in a group, the clinician can be confident that they will not miss subtle, yet clinically-significant, differences. This division of labor makes the best use of the neurologist’s skill and the computer’s capabilities: the computer finds events that are alike, and the neurologist makes the clinical evaluation of these events.
Top of Page
What is hierarchical clustering?
Hierarchical clustering is simply the process of sorting things into groups so that the members of the group are alike. There are a number of analogies that can be used to illustrate this: one might imagine a restaurant receiving 100 cases of wine, only to find that in exchange for a great price, the distributor had mixed various kinds of red and white wine in each of the cases. The task of sorting all of the wine into meaningful groups (merlot in one, chardonnay in another, etc.) would fall to someone who might make a list of what is in each bottle to find out what they received, then decide how to stock it (how many groups), and what they might want to send down the drain right away (artifact). The first level of the hierarchy might be red vs. white. The next level might be by type: is this a bottle of merlot or burgundy. The next level might be by region and then finally manufacturer. The number of varieties and the level of distinction desired will determine the number of groups.
Sorting of spike events is much like this: there may be different focal points, clinically-significant differences in morphology, and often there will be confounding artifact mixed in. SpikeDetector marks these events, and with the information collected, SpikeReview will allow you to automatically sort them into like groups so that you can decide what kind of spikes you have, and to send the artifact groups to SpikeReview’s own recycle bin.
To illustrate how the cluster tree works, we have created a set of simplified, simulated traces with similar spike events grouped together in the illustration. Of course, actual recordings would not have events grouped in time or have this morphology; however, these simulated traces will more clearly illustrate how spike clustering works.
In the scenario that follows we will manually perform the tasks that SpikeDetector and SpikeReview do automatically so that you will see what is done each step of the way. After examining this simulated trace, we recommend examining the sample recording of real EEG that is installed with the EEG Suite: the sk000.lay file found in the C:\Program Files\Persyst\Insight\Samples folder. The SpikeDetector Quick Tour in the help file (Start|Programs|Insight|EEG Suite Help) will take you step-by-step through the process with this recording.
Top of Page
Detection
The simulated recording that follows has four channels labeled A through D with a series of spike events that we have created. An electrode map illustrating their placement follows. The first step in the process of clustering is to assign a “perception” value to each of the spike events in the trace. We have placed the spikes in three groups one after the other, and have labeled these “Group 1” through “Group 3” on the trace.
Traces and channel map
The first group of spike events (Group 1) appears in multiple channels, and the focal channel (B) is the channel with the greatest height. Because this spike stands out well from its background, is well shaped, and appears in multiple channels, we will assign a perception value of “1” to these spikes.
The spike events in Group 2 also appear in multiple channels; however, they only exceed the height of the background activity in channel “B” (the minimum criteria in this example). The spike events in channel B do not stand out well from the background activity and are not well-formed, so we will assign a perception value of .2 to these spikes.
The spike events in Group 3 are similar to those in Group 1; however, in this case the focus is at channel “C”. Also, they are not as well-shaped and there is more confounding background activity so we will assign a perception value of .7 to these spikes.
We can now use this information to create histograms to tell us about the spikes marked in the recording. The first histogram shows the relative population of spikes from the beginning to the end of the recording:
Time Histogram
In the Time Histogram illustration, the X-axis is time (from recording start to end), and the Y-axis is event density (proportional to the number of events in the group). In our example, the number of spike events is greatest at the beginning, middle, and end of the recording, and the peaks in the histogram reflect this.
Next, we can create a histogram showing the relative proportion of low, medium, and high perception spikes in the recording:
Perception Histogram
In the histogram above, the X-axis indicates perception from 0.1 to 1.0, with .1 for small or ambiguous spikes and ranging to 1 for clearly defined spikes. The Y-axis indicates the relative quantity of spike events at any given perception level. In our example the number of spike events with a perception of “.2” is 7, of “.7” is 25, and of “1” is 9.
The Time and Perception histograms are used to give us an overall impression of what kind of spike events are reflected in any given group of events, i.e., at what part of the recording did they occur, and what kind of spike events make up the group. This is important not only when separating spikes into “like” groups—it is also important when filtering events, e.g., taking 1500 events for analysis out of 10,000 marked. The Time and Perception histograms help to verify that the sample is a good representation of all the spikes that were marked. (Returning to our wine analogy, one sip will let you know what kind of wine it is—without having to consume the whole bottle!)
Top of Page
Clustering: Sorting Spikes Into Groups of Similar Events
Now that we have marked the spike events in our example trace and examined our histograms, we can separate the spike events into meaningful groups. To do this we will use the following:
- Topographic perception maps to see the spike focus.
- Perception and Time histograms to see the makeup of the group.
- The “n” number to see the number of events in the group.
- The “d” number for an indication of how “different” the events in the group are from each other.
The topographic maps are similar to traditional voltage maps—with an important advantage: by selecting “perception” (the default) rather than “height” (µV) for the maps, the clinician is presented with maps that use SpikeDetector’s neural-network to more clearly show the spike focus in any particular group. This provides better detail when opening the cluster tree (or using the Review Wizard) to see if there are clinically-significant distinctions yet to be made.
By default, the “d” number is based on both topography and morphology, so significant differences in either or both of these criteria will yield a large “d” number. As this number approaches “2” the events in a group will be visually similar. (Note that the “d” number is a relative measure of difference that is distinct for each patient/data set.)
In the cluster group below (the “parent” group) there are 41 events, and the “d” number of 20 indicates that they are quite different from each other.
Because we have a fairly large “d” number we will open the tree—essentially splitting this group into two new groups based on the largest distinction among the members of the group. Whatever the largest distinction is among the events in the group (either topology or morphology) is where the split will be made:
In the example above, the largest distinction among the spikes in the “parent” group is topology: between the spikes at channel “B” and at channel “C”. As seen on the example trace, the events at channel “C” are all very much alike, so once the bottommost group is broken out we see the “d” number drop to “3”.
The spikes in the other “child” group with a focus of channel “B” have the same topology; however, they have different morphology, so the “d” number dropped from “20” in the parent group to “10”—meaning that there is still significant difference among the members of the group.
How do you know when the “d” number is small enough that the cluster tree does not need to be opened further? While a “d” number of 2 indicates visually-similar events, it is often the case (as in scalp recordings) when a higher “d” number does not yield clinically-significant differences. For example, if the neurologist observes that the maps, histograms, and average trace in the new groups are not very different from one another (e.g., same topology, similar morphology, etc.), and decides that the new groups do not provide new information (e.g., a focus that continues to be at F4), then there would be no need to open the cluster tree further.
Essentially, the “d” number is a guide to what is in the group—if you open the cluster tree another level and do not see a clinically significant difference, you can close the tree by clicking the “-” sign to the left of the “parent” group. (An animation of this is shown on the Persyst CD in the EEG Suite Applications|SpikeDetector|SpikeDetector Video page).
Once we open the “B” group into two groups we see the “d” number drops significantly, meaning that the spikes comprising the group are similar in appearance. Note the Perception Histograms for these groups—the “n=9/d=4” group has a very high perception, and the “n=7/d=3” group has a very low perception. Upon examination of the average trace for the low perception group we decide that this is artifact/not useful and we delete this group from the tree.
After deleting this group, we are left with the remaining two groups that were at the end of the tree (the “leaf” groups). Here we see that there are two spike foci—one at channel “B” and one at channel “C”.
Top of Page
Conclusion
This example uses a simulated EEG trace to step through the process of clustering; however, the process is the same for real recordings too. We recommend that you first follow along with the Quick Tour of SpikeDetector in the EEG Suite Help File (Start|Programs|Insight|EEG Suite Help), using the sample recording installed with the EEG Suite software (sk000.lay). This file clearly shows bilateral spikes and produces predictable results in SpikeReview.
The Quick Tour of SpikeDetector also describes the ReviewWizard. This new feature in SpikeReview not only automates the process of opening the cluster tree, it also provides quick verification that the spikes in a group are alike by presenting average traces of the largest distinction among the spikes in a group. (Refer to the help topic “Using the Review Wizard” for more information.)
After learning how to cluster spikes with this example, we recommend reviewing the clinical application note “Optimizing SpikeDetector Parameters for Long Term Monitoring and Event Notifications” (available on the Persyst CD, Website, and by request from Persyst).
Top of Page
|
