SpectralTrainFig Results Adjudication - nsrr/SpectralTrainFig GitHub Wiki

Despite comprehensive testing and validation of SpectralTrainFig's automatic artifact detection & rejection, artifact-free results are not guaranteed. The majority of issues that could create artifact in SpectralTrainFig's results should be caught during proper data preparation and troubleshooting. However, several varieties of artifact can circumvent the removal process involved in automatic artifact detection and be included in the final spectral ouput. With informed review of each subject's resulting Spectral Adjudication Summary sheet the majority of these artifacts can be quickly identified for removal. This guide contains information on the four presentations of the subject's sleep and EEG data found in the summary sheet and how to quickly review them for known artifacts.

Density Spectra Results
EEG Power Density Spectrogram Results
SWA log10 Density Results
Hypnogram Results
Clean Spectral Adjudication Summary Sheet Example
Example Summary Sheets Containing Artifact

Note: Each investigator should assess whether the methods used in this guide are appropriate for their research hypothesis under investigation.

The Density Spectra is a visual representation of a lead’s power distribution across all bins for the entire study. A typical Density Spectra should present with:

1. A relatively clean curve sloping left to right, generally starting in the 2 to 3.5 uV²/Hz range in the upper left and ending in the 0 to -1 uV²/Hz range in the lower right.

2. Plots for both REM (blue) and NREM (black) densities which will commonly share a similar overall curve. Note that not all studies will include REM sleep.

3. Several common physiological peaks sticking out against the background curve, predominantly in the theta (4-8 Hz) and sigma (12-15 Hz) band ranges.

• These peaks are affected by the percentage of stage 1 and 2 sleep present during recording as well as sleep spindle density.
• These peaks are hallmarks of NREM sleep so it is common for these peaks to be greatly diminished on the REM density (blue) plot.
• Physiological peaks will have an overall rounded shape. Sharp peaks/spikes are typically indicators of artifact and should be further reviewed for removal from results.

1. Increases in the beta+ bands curving upward after ~15Hz.

• Noted in studies with increased arousal, wake time and muscle artifact.
• This upward curve after 15Hz is typically limited to REM or NREM, depending on subject physiology, but can occur during both.
• These values are typically not included in analysis as they occur at frequencies outside expected sleep physiology.

2. Jagged REM results with spikes of varying size across the density plot.

• Typical of studies with very little REM sleep (<20 minutes) causing a very low resolution on the resulting REM density plot.
• While presenting poorly on the Density Spectra these results are still usable data and can be included in analysis.

When adjudicating multiple EEG derivations (ex. C3-M2 and F3-M2) several ideas must be taken into account to avoid incorrectly labeling true EEG activity as artifact. While adjudicating multiple leads from the same ppt:

1. Results should not be expected to pair absolutely between leads due to variations in placement, impedance, pt position, pt physiology, etc.

2. Both REM and NREM plots should share an similar overall shape with any peaks occurring in the same bands with comparable increases in amplitude when compared to results from the derivation of the corresponding region on the pt’s other hemisphere (ex. C3-M2 should only be directly compared to C4-M1.)

3. When two leads recording the same region present completely different curves the results most approximating clean data while still having a similar overall curve when compared to other leads from the same ppt can be kept for analysis. If a more reliable lead cannot be chosen from the pair both should be excluded from analysis.

4. Region-specific variations in multi-lead EEG results can also include the following:

• Frontal leads: Increased sigma density (sleep spindle-range density).
• Central leads: Increased delta density, and increased sigma density (for some pts).
• Occipital leads: Increased alpha density with decreases in sigma and delta density.

The following density spectra are representative of bad data that should be removed from the dataset:

1. Tall spikes on the density spectra, especially if occurring in pairs or triplets of similar height. This is common of electrical artifacts. These artifacts are rare, but very obvious and do not typically require further verification.

2. A very gradually sloping density spectrum with little to no changes in slope during theta and sigma bands. This is typically indicative of a study with large amounts of erroneous slow wave artifact affecting the scale of the density spectra enough to attenuate normal theta/sigma density increases to the point they are unobservable. This artifact should also be verified against the SWA plot.

3. High density at 0 Hz. The gradually sloping studies mentioned above should also be checked for abnormally high density at 0 Hz. Density at 0 Hz should be in the 2-3.5 range with studies beginning in the 4+ range indicative of pervasive slow wave artifact.

4. Clusters of short spikes with waxing/waning amplitude occurring in groups along the density spectra. This artifact has been noted in studies containing electrical artifact but is most commonly assosciated with ECG artifact and can be affected by the pt’s BPM, HRV and length of QRS complexes. As a result of these variations spikes can occur across most of a pt’s density spectra. In general, spiking in the delta-alpha bands is indicative of wide complex ECG bleed and sigma-beta bands indicative of narrow complex ECG bleed. This can typically be verified by the presence of horizonal striping on the pt’s EEG Power Density Spectrogram.

Density Spectra examples available here.

The EEG Power Density Spectrogram represents each signal’s frequency distribution as a function of time. All vertical white bands represent epochs marked as artifact that were removed from analysis and reporting and can be ignored. Horizontal white bands at 12 and 15Hz mark the sleep spindle frequency range for easier visualization. A pt’s EEG Power Density Spectrogram is more difficult to use for visual adjudication due to subtleties in its presentation, but should contain some expected elements to aid in adjudication as well as several clear indicators of artifact.

1. The colors representing frequency density will vary slightly by cohort due to changes in EEG activity determined by cohort age, health, etc but have general guidelines. The majority of EEG activity should present in the -10 to 10 range (dark blue to light green), with gradual variations caused by sleep stage changes. Abrupt changes in density color/distribution are indicative an EEG lead that has come loose/off at that time.

2. If the EEG Power Density Spectrogram is dominated by a single color (typically red, yellow or orange) this is not a clear sign of artifact. These types of results are typically caused by the density color scale attempting to include an extreme value on the EEG lead. These extreme values commonly occur for a fraction of a second when the recording begins or leads are plugged in and are very difficult to identify on the spectrogram but will have little to no effect on spectral results.

3. Density in the delta range (bottom of spectrogram) should increase during scored stage 3 sleep and show increases to yellow, orange and red densities. During adjudication it should be verified that all areas with this increased SWA line up with scored stage 3 sleep on the hypnogram. This is a general guideline and should not be expected to lineup 1:1 with the hypnogram due to delta during stage 2 sleep, artifact etc.

4. Most participants will show a noticeable density increase (changing from dark blue to light blue) in the theta and sigma bands due to the predominance of stage 2 sleep and its hallmarks. This will appear as one or two lighter horizontal stripes that stick out against the background. When these bands may be indicative of artifact are when there are numerous, thin stripes covering some or all of the frequency range. This is indicative of electrical artifact or ECG bleed with location varying based on lead placement, BPM, HRV, QRS complexes and when/where leads become loose or broken.

Spectrogram examples available here.

The SWA log10 Density represents signal density in the 0.5-4.5Hz range as a function of time. The SWA log10 Density is best used to recognize possible low frequency artifact rising from issues such as loose leads or sweat. Physiological SWA density should generally fall in the 2.5-4.0 uV2/Hz range (increasing to 3.0-5.0 uV2/Hz for children) with artifacts rising above this range and plateauing in the 3.5-5.5 range. SWA may (but cannot be expected to) follow a clean distribution during stage 3 sleep with all increases in SWA density during stage 3. Using SWA log10 Density for adjudication includes:

1. Verify that all SWA falls into a realistic physiological range. Plateaus may be of various heights due to normal changes in stage 3 sleep but should follow a gradual rise and fall and not appear as short, independent spikes, which can represent lead popping.

2. Verify that all clean plateaus of SWA line up with scored stage 3 sleep. If a study contains three SWA plateaus of similar density but the third is not scored as stage 3 it is typically a sustained period of electrode popping that regains contact at the end of this “SWA” period.

3. Even if all periods of SWA do line up with scored stage 3 sleep these plateaus should present similar densities. If there are periods of sustained SWA whose density varies by over one order of magnitude (i.e plateaus at 3.0 and 4.5 uV2/Hz) it is typically indicative of artifact during one or both periods of SWA and is worth further inspection or removal from he dataset.

SWA & Hypnogram examples available here.

The Hypnogram is an essential representation of the sleep stages as a function of time. By itself the Hypnogram cannot be used to verify spectral results, but is critical when combined with the Density Spectra, EEG Power Density Spectrogram and SWA log10 Density. The Hypnogram can be used during adjudication to aid in:

1. Verify periods of stage 3 sleep where SWA is expected and rule out erroneous SWA.

2. Verify the presence and length of any scored REM periods. Studies with very short REM periods will produce skewed Density Spectra results, but can represent otherwise clean data. Studies with no scored REM will not render REM-related results.

3. Verify sleep/wake periods in order to aid in reviewing studies with very busy frequency spectrograms and numerous artifacts caused by extremely fragmented sleep. All periods scored as wake on the Hypnogram will be removed from analysis and the resulting Density Spectra and bin density summaries, but will be left in the EEG Power Density Spectrogram.

SWA & Hypnogram examples available here.

The above Spectral Adjudication Summary Sheet shows an example of relatively clean SpectralTrainFig results. Using the data in these plots it can be inferred that this pt’s PSG contained:

• Clean EEG data denoted by a smooth curve with no sharp spikes on the Density Spectra and very few delta or beta artifact markers on their SWA log10 Density (1).

• Relatively healthy (although fragmented) sleep that follows the expected sleep stage progression and includes stages W, 1, 2, 3 and REM without any epochs scored as movement time or “unsure” (2).

• No significant fast artifact that would cause a density increase in the >15Hz range on the Density Spectra (3).

• No significant slow artifact due to a clean SWA plot with few values >3.5 uV²/Hz and rounded, non-attenuated increases in the Density Spectra in the theta and sigma ranges.

• No significant ECG artifact due to the lack of jagged spikes on the Density Spectra and a lack "striping” across the EEG Power Density Spectrogram (4).

• Good spindle activity marked by an increase in sigma power density that attenuates during REM (5).

• One period of sustained slow wave sleep corroborated by SWA >3 uV²/Hz during a period scored as stage 3 sleep on the hypnogram. Other spikes in the SWA spectrum appear to be correctly marked as artifact and are most likely due to lead popping (ex. very high, wide spike at epoch 485) (6).

During adjudication the most common causes for poor spectral analysis results fell into three categories: slow artifacts, fast artifacts and ECG artifacts. The following examples can be used to corroborate suspected poor results.

The most common reason to see artificial slow wave activity in results is lead popping (1). The occasional electrode pop will be caught by automatic artifact detection, but periods of sustained popping and loose leads can be included in results. "Sweat sway” artifact can also manifest as erroneous SWA but is much less common than lead popping. This may be attributed to sweat sway typically affecting shorter periods of the recording with amplitudes closer to true SWA.

• Slow wave artifact is difficult to discern from the Density Spectra alone as it typically blends in smoothly with other results after the spectrum’s scaling incorporates this SWA. However, studies with an excessive amount of extremely slow (or fast) artifact will tend to have a smoother overall curve missing the usual increases in density seen in the theta and sigma bands (2).

• The most obvious place in the results to locate slow artifact is in the SWA (0.5-4.5 Hz) plot. Generally, physiological SWA’s log10 density will fall in the 2.5-4.0 uV²/Hz range with artifact rising above it and plateauing in the 4.0-7.0 uV²/Hz range. Having multiple leads' SWA (0.5-4.5 Hz) plots and hypnogram can help identify these plateaus and confirm them as artifact when they do not line up with periods of scored stage 3 sleep, especially when isolated to one lead. In this example this extremely pronounced SWA is occuring (erroneously) during scored REM sleep (3).

• In the example above it is also evident that slow artifacts are present by viewing the spectrogram. When truly physiological the yellow/orange/red sections denoting increases in density that occur <2Hz on the spectrogram will occur in plateaus that line up with scored stage 3 sleep on the hypnogram. Scattered, tall spikes like above can be indicative of a chronically loose EEG lead (4).

Muscle artifact and electrical interference both require a lead to become loose or broken that then records activity at a much higher frequency than physiological EEG activity. On the pt’s raw EEG these artifacts typically resemble 50/60Hz or alpha activity due to a very consistent rate, amplitude, and waveform (1). A clean 60Hz signal would not be visible in the Adjudication Summary Sheet because it occurs outside the range of interest (0.5-20 Hz), but will be observable in results when the electrical signal is the combination of several slower component frequencies called harmonics.

• Harmonic electrical artifact is very obvious and easily identifiable on a pt’s Density Spectra. This example presents as three large, sharp spikes at 8, 16 and 24Hz (2) and can occur in both REM and/or NREM depending on when the lead becomes loose or broken. In the example above no harmonic artifact occurred during REM. Although the C3 lead came loose at epoch 620 it was temporarily re-attached during the next REM period (3). These same spikes at 8, 16 and 24 Hz are easily viewable on the spectrogram (4).

• Non-harmonic, high-speed artifact can be identified by looking for an unusually smooth curve on the Density Spectra with spiking or rising in the sigma+ ranges (5). Normal density results will slope gradually from left to right with the beta range being the lowest values, but channels with excessive arousal, wake time or muscle artifact can either curve up after the sigma bands or simply display an unusually high density (> -0.5 Hz) in the beta+ bins.

ECG artifact is very common in overnight, and especially unattended PSGs and occurs with varying intensities. Due to its chronic, repetitive nature ECG artifact is not removed through automatic artifact detection. For many recordings with ECG artifact flagging every epoch containing this artifact for exclusion is impractical and could render the entire study flagged for removal. While minor ECG artifact will blend in cleanly with most results there are still visual clues that denote extreme ECG bleed.

• Most ECG artifact can immediately be seen on a channel’s EEG spectrogram as evenly spaced horizontal stripes which stick out against the overall background (1). This "striping" can occur in studies where ECG artifact is significant enough to affect results or not, but in general the more this striping sticks out against the spectrogram’s background the more problematic the artifact.

• Pronounced ECG artifact can be seen on the Density Spectra as clusters of sharp spikes that stick out against the overall curve. Pts with more variation in their BPM will have multiple clusters of ECG artifact at intervals across their Density Spectra (2). This effect is very similar to studies containing both ECG and harmonic artifacts, which can be identified by looking for additional "striping" caused the harmonics which appears more dense and at fewer frequencies.

• Subtle ECG artifact can present as a gently wavering or jagged density spectra with low amplitude spikes that can be present in any or all bins of the spectra. Due to the prevalence of ECG artifact and its varying degrees of intensity each investigator should establish their own cutoff criteria for thess artifacts based the hypothesis under investigation.

• The first Density Spectra below shows pronounced ECG artifact with well defined theta and beta+ components. ECG artifact can be confined to either end of the spectrum depending on the length of the pt’s QRS interval with wide complexes skewing results in the delta & theta spectrums (1) while narrower complexes can affect sigma and beta results (2).

• ECG artifact can also affect the entire Density Spectra when pts have large variation in their BPM and QRS intervals throughout the night (3).

• ECG artifact occurring only during REM sleep is a common variation affecting by lead placement and ppt physiology. Due to the loss of muscle tone and (generally) lower amplitude EEG background ECG spikes that did not affect the spectrum during NREM can become noticeable during REM results, especially in studies with limited REM time (4).