RU2747712C1 - Method for detecting epileptiform discharges in long-term eeg recording - Google Patents

Abstract

FIELD: computing technology.

SUBSTANCE: invention relates to medicine and computing technology. The method comprises the stages wherein at least one EEG signal is preprocessed on a computing device, where local extrema with a window of 0.02 seconds are identified in at least one EEG signal, wherein each extremum is marked in accordance with the available markup; indicators are calculated for each extremum based on the original EEG signal and the wavelet transformation of the original EEG signal; a sample of indicators and marked marks is received by the input of the classifier designed for assessment of the probability of presence of epileptiform patterns in a signal segment; to reduce the number of false negatives, the probability threshold is determined as equal to 5%; post-processing of the obtained results of the classifier is conducted, close extrema marked as epileptiform are combined into one epileptiform discharge.

EFFECT: ensured accurate detection of generalised epileptiform discharges, reduced number of skipped discharges and reduced time for detection of generalised epileptiform discharges.

2 cl, 3 dwg, 3 tbl

Description

FIELD OF TECHNOLOGY

This technical solution relates to the fields of medicine and computing, in particular to a computer-implemented method for automated detection of generalized epileptiform discharges (hereinafter referred to as GER) in long-term recording of an electroencephalogram (hereinafter referred to as EEG).

LEVEL OF TECHNOLOGY

The most visible manifestations of epilepsy - seizures - are poorly suited for regular monitoring of the course of the disease due to their unpredictability and health risks. However, in most patients on the EEG recording, characteristic patterns of epileptiform activity - epileptiform discharges (ER) - can be observed, assessing the frequency of which it is possible to conclude about the intensity and dynamics of the disease. The presence of ER on the EEG is highly likely to indicate the presence of epilepsy.

ER is usually determined by specialist epileptologists by visual examination of long-term EEG recordings (the recording should include a signal taken during sleep to improve the accuracy of the conclusion). This is a very monotonous and tedious process, as it requires a person to review and manually mark off approximately 10 or more hours of recording. The specialist must mark each ER, and then, on the basis of the obtained markup, determine the localization in the case of focal epileptiform discharges (hereinafter - FER) and calculate the indices of epileptiform activity to assess the condition and write a conclusion. This work takes 3-4 hours and has the following number of problems.

The large time spent on marking leads to the fact that this work is given to not the most qualified employees. An experienced specialist only undertakes the function of primary diagnostics of the discharge type and validation of the result. And, as a rule, less experienced specialist takes over the marking and counting of indices. This process creates forced markup errors.

With a sufficiently large flow of patients, specialists are forced to start marking not the entire record, but separate short intervals (usually 10 minutes each), evaluating the total number of ERs based on them. Since ERs can occur unevenly in time, this leads to uncontrollable errors in the calculation of indices.

Experiments show that the markup of ER by different specialists can differ significantly due to the subjectivity of the definition of ER by a person due to the fact that ER patterns can differ significantly from each other and, at the same time, resemble artifacts or some normal patterns.

The prior art discloses a source of information US 8,972,001 B2 03.03.2015 disclosing a method and a data display system in which a plurality of epochs are stitched together with an overlapping section to represent a continuous EEG recording. Artifact reduction is done in epochs and then epochs are merged together with overlapping sections, preferably between two and four seconds. In this case, the initial EEG signal is generated from a machine containing a plurality of electrodes and a processor; split the original EEG signal into many epochs, and each of the many epochs has an epoch duration and an overlap increment; performing artifact reduction for a plurality of epochs to generate a plurality of artifact-reduced epochs; and also combine multiple reduced artifact epochs to generate a continuous processed EEG recording, with each of the multiple reduced artifact epochs having an epoch duration of two seconds in one second increments, with each of the multiple reduced artifact epochs overlapping the reduced epoch of adjacent artifacts to produce a continuous processed EEG recording without discontinuities in a continuous processed EEG recording, with each of the plurality of reduced artifact epochs and adjacent reduced artifact epochs being combined using a weighted average, with the weight proportional to the ratio of the distances to the centers of the epochs, where the value from each of the set of reduced artifact epochs is weighted higher for the overlapping portion closer to the center of each of the set of reduced artifact epochs, and the value from the adjacent reduced artifact epoch is weighted higher for the overlapping portion closer to the center of the neighboring reduced artifact epoch Tom.

The proposed solution is aimed at automatic marking of focal epileptiform discharges of the EEG signal.

The closest analogue is the Persyst 13 software package, in which only 60% of the spikes are marked by an automatic procedure. This tells the doctor that when using the package, 40% of the digits will remain unmarked and means the need to carefully review the entire original signal to make sure that nothing important has been missed. The need to carefully review the entire signal increases the signal processing time.

SUMMARY OF THE INVENTION

The technical problem to be solved by the claimed technical solution is the creation of a computer-implemented method for detecting generalized epileptiform discharges in a long-term EEG recording, which is described in an independent claim. Additional embodiments of the present invention are presented in the dependent claims.

The technical result consists in providing accurate detection of generalized epileptiform discharges, reducing the number of missed discharges and reducing the time for detecting generalized epileptiform discharges.

The claimed result is achieved through the implementation of a computer-implemented method for automated detection of generalized epileptiform discharges in a long-term EEG recording, containing the stages at which:

preprocessing of at least one EEG signal is carried out on the computing device, where local extrema with a window of 0.02 seconds are highlighted in at least one EEG signal, and each extremum is marked in accordance with the existing markup;

for each extremum, signs are calculated based on the original EEG signal and the wavelet transform of the original EEG signal;

a sample of features and marked marks is fed to the input of the classifier;

to reduce the number of false negatives, set the probability threshold equal to 5%;

post-processing of the obtained classifier results is carried out; close extrema marked as epileptiform are combined into one epileptiform discharge.

In a particular embodiment of the proposed method, at least the following features are calculated for the extremum:

- amplitude modulus of the filtered signal at the point of local extremum

- time distance to the next and / or after the next local extrema;

- the maximum amplitude in the window is 0.05 seconds around the extremum.

- root-mean-square deviation of the signal gradient in a window of 0.05 seconds around the extremum.

- the modulus of the difference of the averaged signal gradient to the right and to the left of the extremum in the window 0.05 seconds.

DESCRIPTION OF DRAWINGS

The implementation of the invention will be described in the following in accordance with the accompanying drawings, which are presented to explain the essence of the invention and in no way limit the scope of the invention. The following drawings are attached to the application:

Figure 1 illustrates the characteristics of generalized epileptiform discharges.

Figure 2 illustrates the general scheme of the proposed method.

FIG. 3 illustrates an example of a general arrangement of a computing device.

DETAILED DESCRIPTION OF THE INVENTION

In the following detailed description of an implementation of the invention, numerous implementation details are set forth in order to provide a thorough understanding of the present invention. However, it will be obvious to a person skilled in the art how the present invention can be used, both with and without these implementation details. In other instances, well-known techniques, procedures, and components have not been described in detail so as not to obscure the details of the present invention.

In addition, from the above presentation it will be clear that the invention is not limited to the above implementation. Numerous possible modifications, changes, variations and substitutions, while retaining the spirit and form of the present invention, will be apparent to those skilled in the art.

From the point of view of localization of the source, in patients with epilepsy it is possible to distinguish

two types of epileptiform activity outside of seizures: generalized epileptiform discharges (GER) or focal epileptiform discharges (FED). The number of interictal discharges at a time is used as one of the main indicators of the dynamics of the patient's condition by epileptologists.

The present invention is directed to the implementation of a computer-implemented method for detecting generalized epileptiform discharges in a long-term EEG recording.

In the proposed method, an algorithm has been developed, on the basis of which a predictive model has been created for marking the GER on a long-term EEG recording based on ensembles of decision trees, which allows obtaining an accuracy sufficient for a conclusion to be made by a specialist and having a sufficiently fast operating speed. Marking a ten-hour record using the proposed method takes up to 5 minutes, while with manual marking, a specialist spends about three hours on processing such a record.

The accuracy of the proposed method on the provided control examples averaged 5.8% of missed GERs on cross-validation for patients, which is 39.6% better than the base model with 9.6% of missed discharges.

ERs can have various shapes, however, as a rule, they contain a common element - a spike or a sharp wave. These patterns differ only in duration (20-70 ms for a spike, 70-200 ms for a sharp wave) and do not carry any other qualitative difference, therefore, for simplicity, the term spike will be used everywhere below, except when it is required to explicitly separate these concepts. A spike is a burst of activity in the form of a short sharp peak clearly visible on the signal. A spike is often followed by a slow wave - a high-amplitude hilly disturbance with a duration of 150-350 ms at a spectral frequency of the order of 3.5-4 Hz. The combination of these patterns is called the spike-slow wave complex.

GER can have a rather diverse form, usually the following types of discharge forms are distinguished (their form is shown in Fig. 1):

A - spike

B - sharp wave

C - spike-wave complexes

D - complex acute-slow wave

E - slow-spike complex and slow wave

F - polyspike complex - wave

G - complex multiple sharp and slow waves

H - chain hoist complex

I & J - multiple sharp waves

It can be seen that all types of GER patterns that the specialist identifies contain a pronounced spike. Spike-like shapes (potential spikes) can be found quickly and reliably in a signal using a set of simple rules:

- it should be an extreme of a certain width and sufficient sharpness. This observation makes it possible, when constructing a chain, to consider as candidates for discharges

- only parts of the signal containing potential spikes, and not the entire signal, which leads to significant time savings when processing long recordings.

Many GER patterns consist of more than one spike and can be quite long (up to several seconds), therefore, the method should take into account that several spikes belong to the same discharge and combine them into one label. In addition, such a combination increases the proportion of marked GERs, since it can compensate for the omission of one spike in a long-term discharge if spikes are marked next to it.

The task of the proposed method, in automatic mode, is to display the GER markup and the indices of epileptiform activity calculated from the markup to a specialist. The indices of epileptiform activity are understood as a numerical indicator in the conclusion, on the basis of which an assessment is made of the severity and dynamics of the condition, as well as the effectiveness of therapy and, accordingly, the final result of the proposed method. The accuracy of the index determination depends on the number of missing digits and the number of false positives not corrected manually. The solution of the problem is carried out in the construction of an algorithm that, based on the training sample (a set of X-Y pairs), using the solution of the unbalanced classification problem, builds a model that can distinguish between benign and epileptiform EEG patterns in patients that were not included in the training set.

The specialist has preliminary information that the recording shows an encephalogram of a person who, upon preliminary examination, found activity similar to GER. Each EEG recording is accompanied by an expert markup indicating the locations and durations detected by the GER. Also, a conclusion with information about the person's condition and the values of the indices is attached to each patient. The markup and conclusion are drawn up in accordance with applicable standards. The specified data is in the database. The specialist downloads to the computing device a file with a long-term EEG recording (7-12 hours) made according to the 10-20 system and starts the proposed method for marking the GER. The 10-20 system refers to the standardized scalp electrode placement system recommended by the International Federation of Clinical Neurophysiology (IFCN). According to this system, the location of the electrodes is determined by measuring the head between the external landmarks and taking 10 or 20% of these measurements.

At step 101 of the proposed solution, the computing device preprocesses at least one EEG signal from the database, where local extrema with a window of 0.02 seconds are highlighted in at least one EEG signal, and each extremum is marked in accordance with the existing markup.

As preprocessing, local extrema in the signal are identified, which were considered “candidates” in the ER, in addition, the numerical derivative of the signal with respect to time is estimated.

To form a training sample, local extrema with a window of 0.02 seconds were identified in the initial EEG signal. It is assumed that this window is small enough not to miss ERs, the duration of which, in accordance with the standard, cannot be less than 0.02 seconds. Each extremum was labeled as epileptiform or normal, according to the available expert markup. Then, around each extremum, a signal window was highlighted, in which the features were calculated. The resulting sample of features and labeled marks constituted a training sample for the predictive model.

At step 102 of the attached solution, features are calculated for each extremum based on the original EEG signal and the wavelet transform of the original EEG signal.

Based on the known factors associated with epileptiform discharges and analysis of the literature, to solve the classification problem, two main groups of signs were used, calculated from the EEG signal. The first group consisted of various statistics over the baseline signal. The second group of features contained statistics based on the wavelet transform of the signal.

The first group has the following common features:

A_fltr_abs - amplitude modulus of the filtered (4 - 70 Hz) signal at the local extremum point. Such filtering reduces the amplitude of the slow wave, practically without affecting the amplitude of the spike, which allows this feature to distinguish them.

next, next2 - time distance to the next / after the next local extrema. Potentially, they can highlight areas of signal monotony, which means a slow wave.

max_A_w0.05 - maximum amplitude in the 0.05 second window around the extreme.

std_grad_w0.05 - standard deviation of the signal gradient in a window of 0.05 seconds around the extremum.

delta_grad_w0.05 - the modulus of the difference between the averaged signal gradient to the right and to the left of the extremum in the 0.05 second window ("sharpness").

As a result, we got 6 signs:

'A_fltr_abs', 'next', 'next2', 'max_A_w0.05', 'delta_grad_w0.05', 'std_grad_w0.05'.

The second group - wavelet signs - consisted of statistics on the wavelet transform of the initial EEG signal. The four-level Daubechies wavelet transform (db4) in a window of 0.57 seconds was taken as a basis. This transformation gives five sets of coefficients corresponding to different frequency components of the signal: A4, D4, D3, D2, D1. For each set, 6 statistics were calculated, including minimum, maximum, mean, standard deviation, mean over local minima and mean and local maxima of the coefficients.

As a result, we got 31 signs:

'Db4_A4_max', 'db4_A4_min', 'db4_A4_mean', 'db4_A4_std', 'db4_A4_mean_relmin', 'db4_A4_mean_relmax', 'db4_D4_max', 'db4_D4_min', 'db4_D4_mean', 'db4_D4_std', 'db4_D4_mean_relmin', 'db4_D4_mean_relmax', 'db4_D3_max ',' db4_D3_min ',' db4_D3_mean ',' db4_D3_std ',' db4_D3_mean_relmin ',' db4_D3_mean_relmax ',' db4_D2_max ',' db4_D2_min ',' db4_D2_mean ',' db4_D2_std ',' db4_D2_mean_relmin ',' db4_D2_mean_relmax ',' db4_D1_max ', 'db4_D1_min', 'db4_D1_mean', 'db4_D1_std', 'db4_D1_mean_relmin', 'db4_D1_mean_relmax', 'db4_D4_mean_abs', attribute 'db4_D4_mean_abs'.

Due to the ability to separate the frequency components of the signal, the wavelet transform can be used, in particular, to find areas of fast and slow signal variability, which is useful in the selection of spikes and slow waves.

The next step (103) of the proposed method is the submission of a sample of features and marked marks to the input of the classifier.

In the proposed method, the problem of teaching with a teacher is considered.

Algorithm input: data sample Σ = {Xi, Yi} i = 1, n

Where:

n is the sample size;

Xi - a set of variables characterizing various pre-calculated characteristics of a small segment of the EEG signal;

Yi is a binary variable characterizing the presence of epileptiform patterns during this signal segment, which is the mark of a specialist. This variable is taken from the database.

Algorithm output: a predictive model that takes Xi as input and estimates the probability of the presence of epileptiform patterns in the corresponding signal segment (real number in the range [0, 1]).

The proposed solution uses the XGBoost (EXtreme Gradient Boosting) library that implements the gradient boosting model. The method consists in creating an ensemble of successively refining decision trees.

A typical way to account for class imbalance in many methods, including XGBoost, is to introduce a scaling factor that increases the contribution of each of the elements of the positive class to the overall optimized error function.

As a predictive model, we used the XGBoost model with the optimization of 3 gimer parameters: max_depth went over the grid of values [3, 4, 5,6], the total number of trees in the ensemble n_estimators - over the grid [50, 100, 300, 1000], the scaling factor the contribution of the minor class to the scale_pos_weight error function - along the grid [1, 5000, 10000].

To reduce the number of false negatives, at step 104, a probability threshold of 5% is set.

Since the presence of false negative responses is critical, the probability threshold was additionally shifted, on the basis of which the model makes a decision up to 5%. By default, the classifier considers as epileptiform those events for which its estimated probability exceeds 50%.

At step 105, post-processing of the obtained results of the classifier is carried out, close extrema marked as epileptiform are combined into one epileptiform discharge.

Combine predicted events that are closer than 0.5 seconds.

Display the result of detection on the screen.

The proposed method was compared with the solution offered by the Persyst 13 software package. When Persyst 13 is launched, it simultaneously tries to mark both focal and generalized discharges, therefore, for completeness of comparison, the values of the metrics of the proposed method were compared with two Persyst markings: with a markup containing only predicted generalized discharges , and markup containing all bit types.

This was due to the fact that Persyst was found to mark as focal some discharges, which were marked by specialists as generalized. The results of the comparison of algorithms - the number of missed (uncovered) ERs and the total number of predicted ERs - are presented in Tables 1, 2.

Table 1

Patient code The proposed method,% Persyst (all),% Persyst (gen.),% Patient 1 0 3 17 Patient 2 7 21 38 Patient 3 four 6 17 Patient 4 6 12 25 Patient 5 12 6 17 Average 5.8 9.6 22.8

Table 2.

Patient code The number of ER marked by a specialist The proposed method,% Persyst (all),% Persyst (gen.),% Patient 1 231 388 727 400 Patient 2 29 42 65 38 Patient 3 144 184 421 245 Patient 4 52 155 528 200 Patient 5 149 215 790 392

It can be seen from the tables that both Persyst markup options pass on average 39% more ER and generate more false positives than the proposed scheme.

The estimation of the processing time of a long EEG recording manually and using the considered solutions is given in Table 3. Note that the required time is calculated on the assumption that the specialist manually corrects false positives on the algorithm markup. If we assume that there are no manual edits, then the marking time is reduced to the pure running time of the algorithm.

Table 3.

Markup type Required time, min Net time, min Manually 182 - Persyst 78.8 45 The proposed method 16.25 four

FIG. 3, a general diagram of a computing device (300) will be presented below, which provides data processing necessary for the implementation of the claimed solution.

In general, the device (300) contains such components as: one or more processors (301), at least one memory (302), data storage means (303), input / output interfaces (304), I / O means ( 305), networking tools (306).

The processor (301) of the device performs the basic computational operations necessary for the operation of the device (300) or the functionality of one or more of its components. The processor (301) executes the necessary computer readable instructions contained in the main memory (302).

Memory (302), as a rule, is made in the form of RAM and contains the necessary program logic that provides the required functionality.

The data storage medium (303) can be performed in the form of HDD, SSD disks, raid array, network storage, flash memory, optical information storage devices (CD, DVD, MD, Blue-Ray disks), etc. The means (303) allows performing long-term storage of various types of information, for example, the aforementioned files with user data sets, a database containing records of time intervals measured for each user, user identifiers, etc.

Interfaces (304) are standard means for connecting and working with the server side, for example, USB, RS232, RJ45, LPT, COM, HDMI, PS / 2, Lightning, FireWire, etc.

The choice of interfaces (304) depends on the specific implementation of the device (300), which can be a personal computer, mainframe, server cluster, thin client, smartphone, laptop, etc.

As means of I / O data (305) in any embodiment of the system that implements the described method, a keyboard should be used. The hardware design of the keyboard can be any known: it can be either a built-in keyboard used on a laptop or netbook, or a stand-alone device connected to a desktop computer, server or other computer device. In this case, the connection can be either wired, in which the connecting cable of the keyboard is connected to the PS / 2 or USB port located on the system unit of the desktop computer, or wireless, in which the keyboard exchanges data via a wireless communication channel, for example, a radio channel, with base station, which, in turn, is directly connected to the system unit, for example, to one of the USB ports. In addition to the keyboard, I / O data can also include: joystick, display (touch screen), projector, touchpad, mouse, trackball, light pen, speakers, microphone, etc.

Networking means (306) are selected from a device that provides network reception and transmission of data, for example, Ethernet card, WLAN / Wi-Fi module, Bluetooth module, BLE module, NFC module, IrDa, RFID module, GSM modem, etc. The means (305) provide the organization of data exchange via a wired or wireless data transmission channel, for example, WAN, PAN, LAN, Intranet, Internet, WLAN, WMAN or GSM.

The components of the device (300) are interconnected via a common data bus (310).

In the present application materials, the preferred disclosure of the implementation of the claimed technical solution has been presented, which should not be used as limiting other, particular embodiments of its implementation, which do not go beyond the scope of the claimed scope of legal protection and are obvious to specialists in the relevant field of technology.

Claims (12)

1. A computer-implemented method for the automated detection of generalized epileptiform discharges in a long-term EEG recording, containing the stages at which: the computing device carries out preprocessing of at least one EEG signal, where local extrema with a window of 0.02 seconds are highlighted in at least one EEG signal, and each extremum is marked in accordance with the existing markup; for each extremum, signs are calculated based on the original EEG signal and the wavelet transform of the original EEG signal; a sample of features and marked marks is fed to the input of the classifier designed to assess the probability of the presence of epileptiform patterns in the signal segment; to reduce the number of false negatives, set the probability threshold equal to 5%; post-processing of the obtained results of the classifier is carried out, close extrema marked as epileptiform are combined into one epileptiform discharge. 2. A method according to claim 1, characterized in that at least the following features are calculated for the extremum: - amplitude modulus of the filtered signal at the point of local extremum; - time distance to the next and / or after the next local extrema; - the maximum amplitude in the window is 0.05 seconds around the extremum; - root-mean-square deviation of the signal gradient in the window of 0.05 seconds around the extremum; - the modulus of the difference of the averaged signal gradient to the right and to the left of the extremum in the window 0.05 seconds.

Images