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

Figures

• 

  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.

• 

  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. 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.

• 

  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.)

• 

  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.

• 

  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.)

• 

  Fig. 8. Graphical interface for the app that provides the analysis pipeline described in this paper.