Detecting Spike Wave Discharges (SWD) using 1-dimensional Residual UNet

TL;DR

It is found that a 1D UNet performs best for labeling SWDs in this dataset, and the 1D UNet with data augmentation, AugUNet1D, showed superior performance and detected events with more similar features to the SWDs labeled manually.

Abstract

The manual labeling of events in electroencephalography (EEG) records is time-consuming. This is especially true when EEG recordings are taken continuously over weeks to months. Therefore, a method to automatically label pertinent EEG events reduces the manual workload. Spike wave discharges (SWD), which are the electrographic hallmark of absence seizures, are EEG events that are often labeled manually. While some previous studies have utilized machine learning to automatically segment and classify EEG signals like SWDs, they can be improved. Here we compare the performance of 14 machine learning classifiers on our own manually annotated dataset of 961 hours of EEG recordings from C3H/HeJ mice, including 22,637 labeled SWDs. We find that a 1D UNet performs best for labeling SWDs in this dataset. We also improve the 1D UNet by augmenting our training data and determine that scaling showed the greatest benefit of all augmentation procedures applied. We then compare the 1D UNet with data augmentation, AugUNet1D, against a recently published time- and frequency-based algorithmic approach called "Twin Peaks". AugUNet1D showed superior performance and detected events with more similar features to the SWDs labeled manually. AugUNet1D, pretrained on our manually annotated data or untrained, is made public for others users.

Figures (6)

• Figure 1: Schematic of EEG data preprocessing, augmentation, and the architecture of the AugUNet1D network.
• Figure 2: Example of a seizure in an EEG trace. Note the high amplitude, rhythmic events that define SWD occurring during the correctly segmented orange portion of the trace.
• Figure 3: Training dataset statistics and spectral features. A) Distributions of seizure durations (in seconds) shown separately for each mouse. B) Each portion of the whole corresponds to a single mouse. The size of the portion and the number correspond to the number of SWDs in a given mouse. C) Top: Raw EEG data showing example SWDs in three different mice. Middle: Corresponding spectrograms showing power of the signals at different frequencies across time during the events. Bottom: Power spectral density plots showing the average power at different frequencies across the entire events. Vertical red lines indicate the peak frequency of the corresponding event. D) The average (± SD) of the peak frequencies of SWDs in the 8 mice included in the training dataset
• Figure 4: False positives, negatives, confusion matrices and F1 scores for AugUNet1D and Twin Peaks. A) Example of false positive in the dotted blue box. B) Example of false negative in the dotted red box. Blue lines correspond to events detected by AugUNet1D. Red lines correspond to manually labeled events in A and B. C) Confusion matrices and F1 scores for AugUNet1D and Twin Peaks prediction approaches for each mouse in the test dataset. Proportions in colored panels are as follows: Top left, true positives to all positives. Top right, false positives to all positives. Bottom left, false negatives to all negatives. Bottom right, true negatives to all negatives. These values were calculated on a per event basis, meaning that, for example, a predicted SWD was considered true if it had any overlap with an manually labeled SWD and false otherwise.
• Figure 5: Test dataset statistics, spectral features, and comparison across detection methods. A) Peak frequency of events detected manually, with the modified UNet, or using Twin Peaks across all ten mice in the test dataset. B) Same as A but showing event durations. C) Bars show average durations across all ten mice using three different detection methods. Open circles show the median event durations for each mice. D) Same as in C but showing event peak frequencies. The Twin Peaks event detection method shows a bias for shorter events with higher peak frequencies. Error bars show ± standard deviation