- Algorithms based on multilayer feedforward and recurrent neural networks were employed to automatically detect epileptic discharges (spikes) in unprocessed electroencephalograms (EEG). Results were compared to two separate benchmarks: analysis provided by an expert and output of algorithm based on mathematical morphological filters [3]. Feedforward neural network was able to on average detect ~95% of spikes detected by the expert and ~60% of spikes detected by an aforementioned algorithm. Recurrent neural network was able to on average detect ~93% and 80% spikes respectively. While in some cases artificial neural network was able to outperform algorithm based on morphological filter in terms of detected spikes, the main issue remains the high number of false positives and in particular low-amplitude spike-like waveforms that cannot be utilised for diagnostic purposes.
Early and accurate diagnostics of epilepsy is one of the key factors in order to prescribe proper treatment and improve the quality of life for patients and caretakers alike. Electroencephalograms (EEG) remain one of the main tools used for diagnostic purposes and can be accurately analysed by professionals; however, such process is time consuming and inefficient. While many algorithms are employed to support manual EEG analysis, issues like classification quality or inflexibility still limit practical application.
Despite that currently there are many algorithms designed
for automatic EEG analysis, in many cases it’s still done
manually by visually inspecting recording [
The main purpose of this research is to analyse how Copyright held by the author(s). machine learning and in particular artificial neural networks can be employed for the automatic EEG analysis. The main motivation behind selection of this method is an opportunity to develop an algorithm that can potentially solve problems mentioned earlier and improve identification of various EEG elements. Nonetheless, it would be much easier to adjust such algorithm for diagnostic of other medical conditions. Case study is done with spikes common to benign childhood epilepsy with centrotemporal spikes (referred as Rolandic epilepsy from now on), yet described methods can potentially be applied for diagnostic purposes of other types of epilepsy and brain injuries alike.
The paper is organized into three main parts: first of all we will discuss the current state of research in an automatic EEG analysis. Secondly we will describe data collection and preparation procedures as well as main research methods. Finally we will present how multilayer feedforward neural network and recurrent neural network was able to deal with aforementioned research goals.
Idea to develop algorithm for automatic EEG analysis dates
back to 1970s. In 1972 Carrie JR. [
While aforementioned algorithms present good results in certain cases they all face several issues: 1) Sensitivity to noise. In most cases results are highly affected by noise, which highly depends not only on how recording was performed but on device itself
application for other purposes remains limited and in many cases – impossible.
Machine learning algorithms (and especially artificial
neural networks) are in many cases employed to address
similar problems. Main advantage of such method is that it’s
not necessary to specify spike duration and morphology [8]
therefore neural network can be quickly adjusted (trained) to
analyse signals recorded with different electroencephalographs
and employed to detect spikes with different durations and
morphologies. Research related to application of neural
networks for EEG analysis can mostly fall within two distinct
categories: pre-processed EEG data (see. Gabor and Seyal,
1992 [
Cheng-We and Hsiao-Wen [
While it‘s rather difficult to analyse raw EEG data, much
better results were achieved while working with pre-processed
datasets. Nigam, and Graupe [
III. CHARACTERISTICS OF ROLANDIC EPILEPSY EEG SPIKES AND DATA COLLECTED FOR THIS REASERCH
Rolandic epilepsy is one of the most common types of
childhood epilepsy with ~23% of early school age (mean age
~7 years) children being affected [
Spikes can be classified into two main categories based on
duration [
Both can be described as short term elements, clearly distinguishable from normal brain activity with the main component having negative phase in most cases. Both (spike and sharp waves) have similar initial (or elevation) stage; however, sharp waves have longer demotion stage. In this paper spikes are not distinguished from sharp waves and are analysed together.
Second distinct characteristic can be defined as a
combination of three key measurements: amplitude, duration
and sharpness. Frost J.D. [
For this research EEG readings of 20 patients was used. Among those data of 3 patients was analysed by expert. In each of the files aforementioned expert (a paediatric neurologist) marked spikes and elements that visually resemble spikes but are definitely not. In total coordinates of 499 spikes were collected.
Considering spike characteristics discussed previously, also given that sharp waves and spikes can be both associated with Rolandic epilepsy – a fixed spike length of 200 ms (maximum length of sharp wave) will be used in this research. Nonetheless sub-elements will not be separated in any way and will be treated as one typical epileptic discharge associated with Rolandic epilepsy. EEG readings used in this paper was presented in .edf (European data format) type files recorded in 256Hz sample rate (1 element in time series is equal to 3.90625 ms). Further data analysis was done in fixed-duration sections. In order to prepare such sections, 200 ms. (maximum spike duration used in this research) was approximated to 50 elements in time series or roughly 196 ms. Each section with spike was prepared by choosing 25 elements backwards from spike centre and 25 elements forward, thus ensuring that spike centre (peak) and centre of section will align.
While expert tried to mark spikes at the exact centre it was proven to be difficult to point exact coordinates, therefore data was further post-processed by shifting positions of spike centres to lowest voltage value in the section. While theoretically spike peaks can have positive phase it was not the case for this data therefore no further adjustments was done. Example is presented in figure 1
FIG 1. EXAMPLE OF INITIAL (SOLID VERTICAL LINE) AND ADJUSTED (DASSHED VERTICAL LINE) SPIKE CENTERS
Waveforms that morphologically resemble spikes, however
have lower amplitude, were not included to initial sample.
Those can be considered a non-typical spikes. While those
sometimes can be associated with Rolandic epilepsy, however,
as discussed with expert, final diagnosis is never based on such
“non-typical“ elements. Summary of amplitude characteristics
of all spikes is presented in table I where it is evident that
average amplitude is approximately 12% lower when
compared to research of Frost J.D. [
As discussed in previous chapter data pre-processing can
potentially improve detection quality, however we decided to
keep it to a minimum, mostly because EEG data pre-processing
methods used for feature extraction or denoising are in most
cases designed to work with specific data (eg.
Electrooculogram (EOG) induced noise can be successfully
removed in pre-procesing stage by applying regression based
methods [
As discussed in paragraph III data from 3 patients and 4 EEG channels were used for initial training of neural networks. Given that main purpose of this analysis is to be able to identify spikes in full EEG channel data was prepared in following manner. Each channel was divided into sections each consisting of 50 elements based on rolling window i.e. each section starts with element n and ends at n+50 while next section starts at n+1 and ends at n+1+50. Under such rules each channel is represented by number of sections equal to number of elements in channel.
In the next step segments was classified into two main categories with corresponding numeric values: 0.
Data of full EEG channel was used for training. Each section was classified as spike, if centre of that section falls within interval:
[v – Section size * 0,425; v + section size* 0, 425]
Where v – centre of actual spike, section size – section size or in this case 200 ms.
Coefficient 0,425 represents 42,5 % and is calculated under main assumption that at least ¾ of spike must be visible in order to correctly classify segment it as spike. Therefore (assuming that all elements of specific segment falls within interval [0 ms.; 200 ms.] with spike centre being at 100 ms.) the earliest point when ¾ of shortest possible spike (20ms) can be visible in this segment is when x coordinate of spikes’ centre is equal to 15 ms. (3/4 * 20 ms. = 15 ms.) and latest point is 185 ms. (200 ms. – 15 ms.). This gives interval of [15 ms.; 185 ms.] or ~21 element. Summary of data used for training is presented in TABLE II.
At this point it‘s important to establish criteria to evaluate algorithm performance. Typical binary classification accuracy is not suitable here for several reasons:
First of all, given that in typical 60 second EEG channel (that is consisted of 15360 segments), only less than 1% of those segments can be typically classified as spikes, simple binary accuracy will provide relatively good results despite that spikes was not necessarily found. Nonetheless main purpose of such algorithms is spike detection. Also when using binary classification accuracy, a case where multiple sections was detected next a single spike, is treated in a same manner as when multiple sections were detected next to different spikes. Yet again – given that main purpose is to detect all spikes this would not represent desired performance.
expert (true spikes)
Given that exact coordinates of spikes were well-known following methodology was employed: 1)
All algorithms were design in a way that only sections that were classified as spikes will be outputted. If section was classified as spike it‘s coordinate is adjusted in following manner:
xn = xo + section size / 2 where xn – new coordinate of segment; xo – old coordinate of segment (number of first element); section size – 200 ms 2) Each xn is then compared to known spike coordinates.
Whenever it falls within aforementioned interval of: [v – section size * 0,425; v + section size* 0, 425]
(where v – centre of actual spike) that section is considered correctly classified and spike is considered detected.
According to our evaluation methodology - each spike can be only detected once, therefore more sections (classified as spikes) around single spike will not inflate result. All sections outside interval are classified as non-spikes. Those sections were further grouped in order to assess number of false positives where each group was composed of up to 42 sections starting from first false positive section, with last one being not further than 42 elements while ignoring true positives.
In order to design optimal structure of neural network (number of neurons in hidden layers) a k-fold cross validation was employed (with k = 3). Due to non-standard measurement, data was not shuffled but instead neural network was trained with data from two patients, while tested on data from third one. As in typical cross validation such technique was repeated 3 times. In order to have a starting point several researches were taken into consideration:
Based on research by Cheng-Wen Ko and Hsiao-Wen Chu
[
Cheng-wen et. al. [
Weng and Khorasani [
Mirchandani and Cao [
H = log2 M
Where H – number of neurons in hidden layer, M – largest number of linear separable regions in input data (in this M = 50 – 1 = 49).
Using aforementioned researches as a starting point and employing k-fold cross validation optimal feedforward network architecture was found to be 4 layers with following distribution of neurons in each corresponding layer: 50, 25, 90, 1. Averaged results of cross validation is presented in table III (TP and FP refers to True Positives and False Positives respectively).
While algorithm was able to detect ~95% spikes detected by expert, there were ~30% of segments (not spikes) falsely classified as spikes. Nonetheless ~48% difference of FP and TP segments in channels indicated that results are highly dependent upon channel. Nonetheless at this point analysis was only performed on 3 EEG files therefore results are not good representation of algorithm’s performance.
In order to expand analysis results from single channel were further processed. In order to ensure that each false positive segment will represent a new spike results were grouped. First of all true positives were properly identified and matched to corresponding spikes. Then each false positive is grouped so that each group was filled with segments that are no further from the false positive (that is not part of any other group) than 50 elements. E.g. whenever segment is classified as false positive a new group is formed where that segment is a first group member and all other segments that are distanced no more than 50 elements from the first one are assigned to the same group. One distant segment can have its’ own group.
T3 channel of patient 2 was further processed using aforementioned technique thus resulting in 53 false positive groups that will be referred as 53 false positives from now on. Further analysis of false positives revealed that such elements can be classified into 3 main categories:
pre-defined range.
Technically 1st group is classified correctly, however given that spike peak is on the edge of the segment (at least 3/4 of spike is not visible) it would be impossible to tell without any further references that this is actually a spike therefore such elements won‘t be included or reclassified. See fig.2 for example.
FIG 2. BARELY VISIBLE SPIKE CLASSIFIED AS FALSE POSITIVE
Despite all efforts – due to low amount of high quality data
it is difficult to came up to conclusions therefore data sample
(of 3 patients) was increased by 17 more EEG files without
spike coordinates. To prepare more results for validation an
algorithm based on morphological filters were employed [
Summary of results are presented in table IV. All numbers (True positives or false positives) are in relation to results of algorithm based on morphological filters. Also 3 channels, where aforementioned algorithm identified less than 10 spikes, were removed from final results. max 99% 100%
On average feedforward neural network was able to detect 60% of spikes, detected by algorithm based on mathematical morphological filters, however a lot of elements were falsely identified as spikes (average of ~78%). Nonetheless in 11 out of 17 (with 3 channels removed from initial sample) cases neural network was able to detect more than 50% spikes with 86% detected spikes on average.
At this point false several examples of positives were sent to expert for detailed analysis. Due to limited time and resources expert was able to analyse only 12 false positives however this potentially reveals certain trend. Among 12 results were distributed in following order:
FIG 2. EXAMPLE OF SPIKE (CICRLED ON THE LEFT SIDE) AND LOW AMPLITUDE SPIKE-LIKE WAVEFORM (ON THE RIGHT)
Therefore we conclude that the main problem remains low amplitude or low sharpness spike-like waveforms that are constantly identified as spikes although are unwanted for diagnostic purposes.
To sum up we conclude that multilayer feedforward neural network can potentially be used for primary analysis of EEG with a purpose to identify spikes associated with Rolandic epilepsy and can detect ~60% of spikes on average when comparing with algorithm based on morphological operations or ~95% when comparing to data prepared by expert. However main problem remains low amplitude, spike-like waveforms identified as spikes that can potentially be removed by certain filters. Finally, algorithm discussed in this paragraph did not detected spikes in channels where there potentially were no spikes (or low number of spikes, eg. 1).
Given that too many low-amplitude spike-like waveforms were falsely classified as spikes, a recurrent neural network was employed to address this issue. Main rationale for this type of neural network is that it captures information from previous elements therefore it‘s actually possible to estimate spike amplitude relative to the background brain activity.
Jezusefovich et. al [
While working with the same data, first step was to identify optimal network architecture. Same three-fold cross validation was applied as previously discussed with feedforward neural network. Optimal network depth (layers) was proven to be 1 input layer, 2 LSTM layers and 1 output layer (feedforward neural network). Each LSTM layers had 90 units (also referred as cells). Results of cross validation (when algorithm output was compared to analysis performed by expert) are presented in Table V. Only results of network with aforementioned optimal structure is included. It is evident that while recurrent neural network performed slightly worse in comparison to feedforward neural network (eg.93% detected spikes versus 95%), deviation is minor and given low number of test and training samples in each case we conclude that there are no major differences in performance when testing on similar type of data.
Furthermore, research was repeated in a same manner as
previously (compared to results of algorithm based on
morphological filters [
At this point recurrent neural network was able to detect mores spikes on average, however at the expense of false positives with a small increase from 78% to 81%. After visual analysis of false positives it was clear that same low amplitude waveforms are dominant in this group as well therefore it is evident that shift from feedforward to recurrent neural network alone can’t significantly improve results.
Two algorithms aimed to detect spikes in unprocessed EEG data were analysed in this research.
The algorithm based on recurrent neural network was able to on average detect 78% of spikes (see table VI) when comparing results to an algorithm based on mathematical morphological filters. Also, it outperformed algorithm based on feedforward neural network by ~10% (see table V) on average in terms of detected spikes, but increased number of false positives by 3% on average. While neither algorithm developed in this paper is sufficient on its’ own for diagnostic purposes, neural networks can be used for data preparation or primary analysis.
While neural networks can effectively classify EEG spikes in some cases, the main issue is false positives and in particular - low amplitude spike-like waveforms (see figure 2). While in some cases such waveforms can be classified as spikes, those are not used for diagnostic purposes thus are undesirable and should be removed from the result set. It can possibly be achieved by applying amplitude based filters thus making such combined algorithm an effective tool for EEG analysis, however, this is beyond scope of this paper and is recommended for future research.
We deeply acknowledge paediatric neurologist MD Rūta Samaitienė for her input and expertise related to epilepsy and data collection efforts. Also, I would like to thank MSc Andrius Vytautas Misiukas Misiūnas for support and methodological guidance.