Fully automatic peak frequency estimation of the posterior dominant rhythm in a large retrospective hospital EEG cohort.

OBJECTIVE: To develop and test a fully automated method for estimation of the peak frequency of the posterior dominant rhythm (PDR) in a large retrospective EEG cohort.
METHODS: Thresholding was used to select suitable EEG data segments for spectral estimation for electrode O1 and O2. A random sample of 100 peak frequency estimates were blindly rated by two independent raters to validate the results of the automatic PDR peak frequency estimates. We investigated the relationship with age, sex and binary EEG classification.
RESULTS: There were 9197 eligible EEGs which resulted in a total of 6104 PDR peak frequency estimates. The relationship between automatic estimates and age was found to be consistent with the literature. The correlation between human ratings and automatic scoring was very high, rho = 0.94-0.95. There was a sex difference of d = 0.33 emerging at puberty with females having a faster PDR peak frequency than males.
CONCLUSIONS: Fully automatic PDR peak frequency estimation not dependent on annotated EEG produced results that are very close to human ratings. SIGNIFICANCE: PDR peak frequency can be automatically estimated. A compiled version of the algorithm is included as an app for independent use.

Introduction

An approximately 10 Hz oscillation posteriorly on the scalp was one of the earliest observations following the invention of EEG (Berger, 1929). In healthy adults there are three rhythms in the alpha band: the midtemporal and mu rhythms and the posterior dominant rhythm (PDR), maximally visible above the occipital areas (Niedermeyer, 1997), best seen with the subjects eyes closed and during relative mental inactivity (Kane et al., 2017). The maturation of the PDR with age has been well studied, emerging from less differentiated rhythms at 4 Hz at 4 months of age, increasing to 6 Hz at 6 months, 8 Hz at 3 years and stabilizing at near 10 Hz at 10 years (Eeg-Olofsson, 1971a, Eeg-Olofsson, 1971b, Lindsley, 1939, Niedermeyer, 2011) and declines in senescence (Aurlien et al., 2004). In a study of 500 normal adults the mean PDR peak frequency was found to be 10.5 ± 0.9 Hz (Brazier and Finesinger, 1944). The PDR peak frequency is stable in adulthood but exhibits short-term intra-individual variability from being influenced by attention and fatigue. The between subject variability is much higher. Intraindividual alpha frequency variation has been reported to be 0.67 ± 0.46 Hz in a study of 120 patients with focal and generalized epilepsy and 40 healthy controls with a mean age of 33.8 (±13.6) years (Khan et al., 2018).

The PDR can be slowed by various pathologies such as Parkinson’s disease (Cozac et al., 2016), Alzheimer’s disease and depression (Prinz and Vitiell, 1989). Anxiety may be associated with faster PDR. Clinically, the PDR peak frequency is evaluated by visually inspection. Frequently utilizing a frequency analysis tool that is integrated in the EEG viewer program. The maximum frequency resolution from this tool may vary between programs. Automatic characterizations may be a way to save time, improve objectivity and to more precisely determine the peak frequency.

Lodder and van Putten used an interative curve-fitting method on 1215 normal eyes-closed resting-state EEG from individuals aged 4 months to 96 years detecting a peak for 1160 of the EEGs with 92.5% of the automatic PDR estimates within 1.2 Hz of the manual PDR estimates (Lodder and van Putten, 2011).

This approach required following a specific protocol to secure two minutes of eyes-closed activity (Lodder and van Putten, 2011). Our aim was to develop a method to automatically estimate the peak frequency of the PDR without requiring eyes-closed annotations and test this algorithm on a large dataset from a hospital population spanning all ages and to reproduce the age-related changes known from the literature (Lodder and van Putten, 2011).

Methods

EEG data

EEG records from region Zealand in Denmark (population approximately 820,000) between 2006 and 2019 were used. EEG exams were from in- and outpatients of all ages. Typical referral diagnoses were related to epilepsy, sleep and impaired consciousness. Recordings varied in duration from only a few minutes to approximately 30 min for standard EEGs, to several hours for continuous EEG and polysomnography-EEG. Standard EEGs followed a semi-standardized protocol with repetitive eye blinks, photic stimulation and hyperventilation. Records shorter than 10 min were excluded to avoid exams aborted for technical and or patient compliance reasons. Exams longer than 160 min were excluded to avoid data from the epilepsy monitoring unit, polysomnography studies and cEEGs from critically ill patients.

Personal ID (a unique identifier that every person in Denmark has) was used to ensure that each individual could only contribute with one exam to the analysis. In case of multiple, the first one was used. The personal ID was only accessed programmatically to identify date of birth and sex and then substituted with a pseudo-ID to protect anonymity. Study was approved by the regional committee for data protection and ethics (19-000079/001).

The EEGs have been recorded on Nicolet amplifiers (CareFusion 209 Inc. USA). Sampling rate is either 250, 256 or 1024 Hz. Channel numbers are typically either 19 or 25 (including the inferior temporal row) in the 10–20 system.

Algorithm

The analysis was done using MATLAB (version 2018b. Natick, MA). The EEG files were stored in Nicolet’s proprietary e-format. We used the function NicoletFile to access header information and import the EEG data as a channel-by-data matrix. The algorithm follows several steps, pre-processing, channel rejection, segmentation, spectral estimation and peak detection explained below (Fig. 1). Examples of the segmentation and peak detection results are provided in the Supplementary Material (Supplementary 1).

Fig. 1

Flowchart of data selection. There was a total of 14,807 EEG records in the archive. Duration must be within 10 and 160 min otherwise the record was excluded. Each individual can only contribute with one record in this study so any additional exams from the same individuals were excluded. Records without age and sex information were excluded. This leaves 9197 records for the main analysis. An overview of the algorithm is shown to the right. All event lists and spectrograms of the results were manually reviewed, resulting in 2002 excluded exams. Among the 7195 EEGs in the final analysis, different number of peak detection for O1 and O2 electrodes are shown below.

Preprocessing

Data was re-referenced to common average. To reject bad electrodes, we used root mean square (RMS) as a measure of high-frequency noise frequently elicited by movements or poor electrode contact. RMS was calculated pr. electrode for the recording. The RMS results for all electrodes were converted to z-scores and electrodes outside ±2 standard deviations were rejected. This works when the average amount of noise is low, therefore, electrodes with a RMS above an empirically chosen threshold of 1000 were also rejected.

If O1 or O2 were excluded or if the remaining number of channels were reduced to below 15, the EEG record was excluded. Data during photic stimulation was ignored.

Segmentation

The PDR cannot be assumed to be uniformly visible throughout a recording. Even if the procedure is to have the patient close his or her eyes during standard exams, this behavior cannot be expected to be constant. During segmentation we aim at selecting relatively noise-free intervals with a visible and relatively stationary PDR.

A finite impulse response (FIR) bandpass filter between 0.5 and 70 Hz was applied. To isolate relatively noise free intervals for the next steps in the analysis, we used a combination of methods: To reject intervals with too much average noise across channels, we used a 0.1 Hz high-passed RMS-envelope using a 1 s sliding window of the global mean field power (GMFP). To deselect noise particularly affecting O1 and O2, RMS-envelopes were also calculated for these channels. A threshold based on the mean of the maximum values of each envelope is set. If this threshold is lower than 60, it is increased by 10%. If it is below the mean of either envelope, it is increased by 50%. These settings were adjusted by testing on a small subset of the EEGs. Next, continuous intervals longer than 10 s where the maximum values sample-by-sample of both envelopes fall below the threshold were identified. These intervals were passed on to the next step.

If the combined duration of the accepted segments was shorter than 2 min, an alternative selection mode was used. An interval from 60 s into the file until 60 s before the start of photic stimulation was selected. In the absence of photic stimulation, 10 min was used. If photic stimulation commenced early, then the 10 min was taken 1 min after the end of the photic stimulation interval. Refer to the Supplementary Material 1 (example 5, plot 2) for an example of alternative data selection.

Spectral estimation

Welch's method for spectral estimation (Matlab function pwelch) was used on the accepted intervals for O1 and O2. A 4 s hamming window with 50% overlap and zero-padding to 4 times the window length was used. Frequencies were trimmed to the range 2–16 Hz. Power was converted to decibel. A Savitzky-Golay filter was used to smooth the spectrum to reduce random fluctuations (Corcoran et al., 2018). We empirically determined that a third order filter with 25 frames worked well for the length of the obtained PSD functions.

Spectral peaks were found using the Matlab function findpeaks with a minimum distance between peaks of 0.5 Hz and a prominence of at least 5 dB. For each frequency, the area under the curve (AUC) spanning the width around the peak was calculated and saved along with the amplitude, frequency location, width and prominences.

Manual rejection

Artifacts, seizures and paroxysmal activity can produce spurious spectral peaks at different frequencies that do not represent an endogenous posterior dominant rhythm. Event lists extracted from the EEG files were screened for key words indicating decreased level of consciousness or sedation, seizures or significant artifacts and records were excluded if such indicators were found. Spectrograms were also inspected for signs of artifacts and seizure activity and records were excluded if such were found. Examples of the type of event list comments and spectrographic appearance that prompted manual rejection is provided in the Supplementary Material (Supplementary 2). These were generally found to be cEEGs from critically ill and sedated patients.

EEG classification from text scan

For some of the EEG records there was an associated word document containing the EEG report and a conclusion that would use the specific phrase “nothing abnormal” in the absence of abnormalities. We used a script to search for the phrase “nothing abnormal” appearing after “conclusion”. If this was found, the result was coded normal, otherwise abnormal. If there was no associated word-text EEG report it was classified as unknown.

Comparison with manual PDR frequency estimates

To assess the accuracy of the automatic PDR peak frequency estimates, a random sample of 100 of these were evaluated by two raters, one board-certified in neurophysiology, who were blinded to the automatic estimates.

Statistics

Means are indicated with ± standard deviations, if non-normal judged from visual inspections of histograms, medians and ranges are provided and kernel density plots of the distributions. Correlations use Pearson’s method. Interrater statistics are based on intra-class coefficient ICC(A,1). Groups differences are converted to cohen’s d. The smoothing line on the PDR by age relationship scatter plot use a generalized additive model (GAM). Bland-Altman plots were used to evaluate pairwise comparisons of asymmetric PDR frequency estimates. The two-sample Komolgorov-Smirnov test was used to test if the validation set was representative of the full set of EEGs. Confidence intervals are 95% unless stated otherwise. Significance tests are two-sided, and alpha is 0.05 unless stated otherwise.

Statistics were done in R version 3.6.2 (R Core Team (2019)) using packages “mgcv”, “ggplot2”, ”ggridges” and “irr”.

Results

Peak frequencies

There were 9197 EEGs available for the analysis. Of these, 2002 were rejected as described in section 2.3, leaving 7195 EEG records. 6104 either had a unilateral peak at O1/2 or, more commonly, bilaterally detectable peaks (94.09%). In case of bilateral peaks, we used the mean of the O1 and O2 peak frequency as the PDR peak frequency for that individual as the peak frequencies were generally very close to each other (refer to Supplementary Material 1, the examples). This is further discussed in Section 3.3.

Fig. 2 shows a scatterplot of all 6104 peak frequency estimates. As expected, we see a rapid increase in frequency from early age reaching a maximum during adolescence. There are more observations in this range as can be seen from the density plots (upper margin of Fig. 2), showing a bimodal distribution with a peak among children and second wider one from approximately age 50 to 80. Frequency remains stable at approximately 10 Hz until age 40 when a gradual decline on the group level sets in. In higher age, there is greater dispersion, which most likely reflects differences in underlying disease and prevalence of abnormal exam results also increase with age.

Fig. 2

All posterior dominant rhythm (PDR) peak frequency estimates plotted as a function of age. The blue line is a generalized additive model of the mean PDR with confidence regions indicated as a grey band. PDR increases from infancy to stabilize at approximately 10 Hz in adolescence. On the group level, there is an incipient decline starting at approximately age 40. Marginal density plots show the age distributing grouped by sex (top) and PDR frequency by EEG classification (right). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 3 shows the relationship between mean PDR peak frequency estimates grouped by sex. The sexes follow each other closely until puberty after which mean PDR peak frequency is higher for females than males and remains so even as it declines for both sexes due to age-related factors.

Fig. 3

Left) Posterior dominant rhythm (PDR) as a function of age grouped by sex. The lines overlap until early teens whereupon the female mean increases more than the male mean and remains higher as PDR declines for both sexes. Right) Age-stratified kernel density plots of the PDR grouped by sex.

Table 1 shows the group summary statistics from the density plots. In the group classified as normal, the youngest aged 0 to 9 have a median peak frequency is 7.9 Hz. For age 10–39 the median is 9.7–9.9 Hz and then it declines for every decade reaching a median of 8.17 Hz for the above 80 group. In the abnormal group the median follows the same trajectory but is slightly displaced downwards.

Table 1

Posterior dominant rhythm peak frequency by age group and EEG classification.

	Posterior dominant rhythm frequency
Normal EEG (n = 2839)
Age group	Mean	sd	Median	Range	n
0 to 9	7.25	1.97	7.94	[2.19, 10.97]	613
10 to 19	9.66	1.01	9.7	[3.06 12.48]	760
20 to 29	9.91	1.00	9.97	[5.56, 12.48]	281
30 to 39	9.87	1.08	9.94	[7.00, 12.38]	165
40 to 49	9.66	1.08	9.62	[5.64, 12.59]	249
50 to 59	9.36	1.02	9.37	[5.75, 12.00]	245
60 to 69	9.06	1.05	9.09	[2.50, 12.59]	286
70 to 79	8.78	1.14	8.88	[5.50, 12.73]	174
80 and up	8.24	1.07	8.17	[5.91, 11.91]	66
Abnormal EEG (n = 1571)
Age group	Mean	sd	Median	Range	n
0 to 9	7.26	2.13	7.88	[2.06, 12.64]	277
10 to 19	9.29	1.25	9.43	[3.19, 11.97]	314
20 to 29	9.28	1.54	9.53	[3.90, 11.72]	108
30 to 39	9.25	1.68	9.53	[2.38, 12.25]	73
40 to 49	9.23	1.18	9.30	[6.22, 11.38]	102
50 to 59	8.5	1.54	8.82	[3.47, 12.91]	138
60 to 69	8.34	1.41	8.47	[2.81, 12.44]	226
70 to 79	8.15	1.28	8.26	[3.31, 11.97]	232
80 and up	7.71	1.22	7.72	[2.43, 10.99]	101

For 4410 EEGs with both a PDR estimate and a classification we can plot PDR by age grouped by abnormal and normal EEG classification (Fig. 4). The overall shape of the relationship remains the same, but the trajectory for the abnormal group is displaced towards lower PDR peak frequency. There are 2166 EEGs from patients above age 11 with a normal classification. The effect size difference between the sexes in this group is d = 0.33, CI95 [0.25, 0.42].

Fig. 4

Scatterplots of peak frequency estimates of the posterior dominant rhythm (PDR) grouped by sex (left) and EEG classification (right). PDR is higher in females than in males. With advancing age, the proportion of abnormal exams increase, which can be seen from clustering of red observations for age above 50 and PDR tends to be lower as well. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Validation

Fig. 5 is a correlation matrix showing the output of the algorithm against rater 1 and 2. From the pairwise comparisons we see a correlation of 0.94 and 0.95 between the algorithm and raters. Between human raters the correlation is 0.93. The two-sample Komolgorov-Smirnov test was used to evaluate the representativeness of the validation set compared the full data set. There was no significant difference between the distributions of PDR peak frequency estimates between rater 1 and the algorithm (p = 0.56), rater 2 and the algorithm (p = 0.34). The age distribution in the test set did also not differ significantly from the full set (p = 0.85).

Fig. 5

Correlation matrix of pairwise comparisons between the posterior dominant rhythm estimating algorithm (Algo) and rater 1 (R1) and 2 (R2). Histograms on the diagonal show the distribution of frequencies. Scatterplots off-diagonal show the comparison indicated by the labels on the edges. Number in top-left corner of each scatterplot is the Pearson correlation coefficient.

Absolute agreement assessed using intra-class agreement (ICC) is 0.93, CI95 [0.91, 0.95], which is excellent. The performance of the algorithm is interchangeable with any of the two raters, which are also interchangeable with each other.

Agreement does not appear to depend on age. Linear regression predicting the differences between any rater from age was insignificant (see Supplementary Material 3). Compared to the raters, the algorithm appears to be slightly biased towards lower peak PDR frequency estimation, one sample t-test shows a mean difference of 0.19 Hz, p = 0.00042 for rater 1 and 0.16 Hz, p = 0.0006 for rater 2 (Supplementary Material 3). The human raters compared to each other do not exhibit such a bias, p = 0.6.

Asymmetry

For most cases, there is little difference between PDR peak frequency estimates from O1 and O2 (Fig. 6). In 52% of cases the O1 and O2 estimates are within 0.1 Hz of each other and 94.5% of cases are within 1.5 Hz of each other. From the scatterplot we can see that most of the larger discrepancies cluster around age 5–15 and are in the theta range. Because of the 1/f spectral power distribution, much briefer segments containing slower activity will generate comparatively large spectral peaks that might be identified by the algorithm as the dominant rhythmicity. A Bland-Altman plot (Fig. 7) shows this tendency of asymmetry to become larger at slower frequencies. The limits of agreement are −1.7 to 1.7 Hz. The mean difference is −0.035 and a one-sample t-test for bias is significant (p = 0.002) suggesting a tendency for faster frequencies for O2, but the size of this differences falls below the frequency precision of the spectral estimation (0.1 Hz) and only becomes significant because of the large sample size.

Fig. 6

Left) Scatterplot of the frequency estimates for the posterior dominant rhythm as a function of age. Color indicates the difference between O1 and O2 frequency estimates (red means O2 is higher, blue means O1 is higher). There is a cluster among children with low frequency estimates where there are comparatively more asymmetric results. Right) Distribution of O1 and O2 differences. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 7

Bland-Altman plot of O1 and O2 peak frequency estimates differences plotted against their mean. Limits of agreement indicated with dashed horizontal red lines. Observations within the limits are blue and those outside are red. We observe increasing differences in both directions as frequencies go towards 6 Hz. The mean difference is −0.035 Hz and a one-sample t-test for bias indicate that this small bias towards higher O2 frequency estimates is statistically significant, p = 0.0021 (top right box). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

The EEG data comes from patients so PDR asymmetry could reflect underlying structural brain disease and not error. To test this, we selected a random sample of 25 EEGs where the algorithm had found a difference between O1 and O2 larger than 1 Hz. Rater 1 evaluated the same EEGs without knowledge of the automatic estimate. We found that in only two cases the human rater could confirm a significant PDR asymmetry (Supplementary Material 4). In the remaining cases the human expert judged that there was no significant asymmetry. We therefore conclude that the algorithm cannot be relied upon to identify true cases of asymmetric PDR.

PDR peak frequency estimation app

The pipeline from EEG file to PDR peak frequency estimate described in this paper has been compiled into a distributable app that can be downloaded as part of the Supplementary Material (Supplementary Material 5) along with instructions for installation and use. It comes with a graphical interface as shown in Fig. 8.

Fig. 8

Graphical interface for the app that provides the analysis pipeline described in this paper.

Depending on CPU speed it can analyze a 30 min EEG file in less than half a minute, which is time efficient compared to expert visual analysis.

Discussion

Our automatic PDR peak frequency estimates from 6104 patients exhibit a similar relationship with age as has been found in other studies (Aurlien et al., 2004, Lodder and van Putten, 2011), however, unlike (Lodder and van Putten, 2011) our automatic approach is not dependent on pre-selected 2 min interval with eyes closed but also automates visible PDR interval selection. We have also compiled the algorithm scripts into a simple app that anybody can use. The negative trade-off is that sometimes the algorithm will select inappropriate data rendering the output misleading. The app produces also produces spectrograms of the data (similar to what is shown in the Supplementary Material 1 and 2). From these a human rater can easily discard unrealistic estimates. As such, the numeric output and information about asymmetry from the app should be appraised judiciously.

PDR maturation follows a pattern of appearing at around 4 months of age at 4 Hz and then increasing in frequency, entering the alpha band range typically at age 3 and center around 10 Hz in maturity (Niedermeyer, 1997). Marcuse et al. examined 79 normal adolescents at 15 years of age and found a mean PDR at 9.9 Hz ± 0.5 Hz (range 8.7–11.0 Hz) increasing to mean 10.0 Hz when reexamining the same group a year later (Marcuse et al., 2008). This is similar to a study of 185 normal subjects aged 16–21 where the mean PDR frequency was 10.2 Hz ± 0.9 Hz (Eeg-Olofsson, 1971a, Eeg-Olofsson, 1971b). Eeg-Olofson found a mean of 9 Hz at age 7 and 10 Hz at age 15 (Eeg-Olofsson, 1971a, Eeg-Olofsson, 1971b). From a magnetoencephalography (MEG) study of 51 normal adults found a mean alpha peak frequency of 10.3 Hz ± 0.9 Hz and a between-subject variability of 2.8 Hz (Haegens et al., 2014).

Twin studies suggest that alpha peak frequency has a mean heritability of 0.85 and that the heritability is lower in advanced age (Smit et al., 2005). This could reflect random effects of age-related diseases affecting one monozygotic twin but not the other, reducing heritability with age.

We identified a similar sex difference in peak PDR frequency as (Aurlien et al., 2004) and that it emerges at puberty. This is consistent with other studies: In a study of 160 patients with focal and generalized epilepsy (40 controls) there was a sex differences with 0.32 Hz faster PDR in females than males (Stephani et al., 2019). Eeg-Olofsen found more slow frequency activity in girls than boys until age 8 but a reversal of this pattern at age 14–15 and also noted higher peak frequencies in girls (Eeg-Olofsson, 1971a, Eeg-Olofsson, 1971b). The difference might be larger depending on the menstrual cycle as PDR peak frequency is higher in the luteal than in the follicular phase (Brötzner et al., 2014). Estradiol increases neuronal excitability and progesterone decreases it (Finocchi and Ferrari, 2011) and the effects on the PDR peak frequency appear negatively correlated at 0.4 (Brötzner et al., 2014). Women on oral-contraceptives had a PDR peak frequency more similar to women in the follicular phase than in the luteal phase consistent with this (Brötzner et al., 2014). Differences in body temperature has been suggested as an explanation (Stephani et al., 2019) since EEG alpha frequencies become faster with higher temperature (Deboer, 1998) and there is evidence from large cohorts that women have slightly higher body temperature than men (Protsiv et al., 2020). The relationship with temperature is unlikely to be causally driving the difference in PDR. The underlying processes that cause secondary sexual differentiation at puberty could have pleiotropic effects that causes both differences in PDR and body temperature even though PDR may also be influenced by body temperature. We intend to study the relationship between PDR and menstrual cycle in a study of 365 days of continuous EEG.

Changes in anti-epileptic drugs (AED) can also affect peak frequency, AED slows alpha frequency and discontinuing AED removes this effect (Clemens et al., 2006, Salinsky et al., 2003). Affects may vary depending on the type of AED one study finding a slowing effect of carbamazepine and a possible accelerating effect of lamotrigine (Clemens et al., 2006).

Some subjects did not have a detectable peak. This could be a true negative result, which is the case for some of the excluded subjects. Some were unconscious and should not be expected to have PDR activity in the alpha range. In case of alpha-coma, there could be alpha band activity, but such subjects have also been excluded. Most negative results are most likely false negative, where the automatic method could not detect the PDR frequency. From visual inspection of spectrograms, especially in older subjects, there were lower amplitudes, but a transversal band was visible. For younger patients, the problem was more often movement related artifacts.

Clinical utility and future research

The mature PDR peak frequency is stable within-person in the absence of disease (Samson-Dollfus et al., 1991) and can thus be conceived as a marker of brain health. Conditions such as mild cognitive impairment and Alzheimer’s disease cause slowing of occipital alpha band activity (Garcés et al., 2013, Prinz and Vitiell, 1989). More accurate PDR peak frequency estimation could therefore help in earlier diagnosis. In visual EEG analysis, PDR frequency is often rounded to the nearest 0.5 Hz, but this reduces precision if earlier detection of PDR slowing is the objective. Our automatic method increases PDR frequency resolution to 0.1 Hz. In longitudinal studies, it is therefore likely, that automatic methods will be superior in early detection of PDR slowing, whereas in traditional visual EEG inspection, slowing might not become apparent before PDR frequency has dropped at least 1 Hz. However, PDR peak frequency estimation for this purpose needs to take intra-individual PDR frequency variation (Haegens et al., 2014) into account. If this degree of precision is required, it is not optimal to complicate the task with having the algorithm select parts of the data for analysis as is the case for our method.

Limitations

For this study we have no way to access the diagnostic information associated with the EEGs or what medication the patients are on if any. We also do not have access to information about whether the patient is sleep deprived. These unaccounted variables add to the variability in PDR peak frequency estimates on the group level.

Even though the algorithm can produce true negative results i.e. that no alpha peak is visible it was not designed to classify between EEG from awake and aware subjects and patients in coma. In the absence of visible theta-alpha range peaks and in the presence of variable noise, the algorithm may produce unrealistic results. From the spectrograms, however, such cases can be manually rejected. In cases of low voltage EEGs, no peaks may rise above the threshold and the algorithm will fail to detect an alpha peak under those settings. The threshold may be manually readjusted.

Text scanning produces a rough dichotomization in normal and abnormal based on the presence of certain phrases in the EEG report. In the results we observe some very low peak frequency estimates but in some cases the EEG was classified as not abnormal and this is untrue by definition.

Alpha band brain oscillations can be described in terms of a mixture of components originating from distributed processes (Barzegaran et al., 2017). Because of volume conduction, spectral analysis on individual channels may produce spectra containing multiple peaks representing different alpha-rhythms (Barzegaran et al., 2017). During maturation, there are reliable age specific waveforms that the algorithm was not designed to take into account, such as posterior slow waves of youth and other slow 2–5 Hz frequencies occurring in infants up to 36 months of age (Niedermeyer, 2011). These could have contributed to the cluster of delta-range estimates found in the youngest patients (refer to Fig. 2). Non-PDR alpha rhythms can have different cortical origins reflecting different underlying processes (Nunez et al., 2001). Source separation techniques could help identify one rhythm as a more likely candidate for the PDR than others based on its topographical characteristics. Our analyses were limited to the common average montage.

Conclusions

Fully automated PDR estimation provides estimates that are strongly correlated with expert human ratings (ICC among raters including the algorithm was 0.93, CI95 [0.91, 0.95]), and the PDR by age relationship obtained from a sample of 6104 EEGs is consistent with the literature. PDR peak frequency increases with age to about 10 Hz in adolescence. We also detect a small sex difference (d = 0.33) emerging at puberty that females have faster PDR frequencies than males.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. IZ is funded from the Innovation Fund Denmark, 6152-00005A and Læge Sofus Carl Emil Friis og Hustru Olga Doris Friis’ Legat. TK is a full-time employee of Region Zealand.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.