Extraction method of cross-frequency coupling features of cortical EEG signals

Abstract

The present invention belongs to the field of brain-computer interface technology, specifically a method for extracting cross-frequency coupling features of cortical EEG signals. The present invention includes: signal preprocessing (filtering, signal abnormality segment detection, noise suppression) to reduce various interferences on cortical EEG signals; extracting features, that is, calculating the cross-frequency coupling strength between each frequency band as a feature, specifically using bandpass filtering to extract signals of different frequency bands respectively, and calculating the cross-frequency coupling strength between signals of different frequency bands, including phase-amplitude coupling strength, phase-phase coupling strength and amplitude-amplitude coupling strength; and introducing a sliding time window to calculate the maximum value of the coupling strength over time; and improving the frequency discrimination of features by subdividing low-frequency components. The present invention can improve the signal-to-noise ratio, optimize signal quality, and have a stronger detection capability for potential symptoms associated with frequency differences.

Description

Technical Field

[0001] The present invention belongs to the technical field of brain-computer interfaces, and in particular relates to a method for extracting cross-frequency coupling features of cortical electroencephalogram (EEG) signals. Background Art

[0002] Cortical EEG signals are acquired by placing electrodes directly on the exposed surface of the cerebral cortex. They reflect the electrical activity of neurons in the nearby cortex. Compared to scalp EEG signals, they have higher amplitude, spatial resolution, and signal-to-noise ratio, and can more accurately reflect certain physiological states of the brain. As an invasive acquisition method, cortical EEG signals are currently primarily used during brain surgery for intraoperative monitoring and localization of the affected area, or in brain-computer interfaces to enable patients with neurological conduction dysfunction to perform actions. Because EEG signals are non-stationary and subject to numerous unavoidable interferences during the acquisition process, this poses significant challenges to EEG signal processing and feature extraction methods. Currently, the clinical application of cortical EEG relies primarily on physicians' empirical observation of the signal's amplitude, waveform, and primary frequency components. However, limited research has examined the correlation between different frequency components.

[0003] EEG signals are composed of components in different frequency bands, which are not mutually independent. The correlation between different frequency components is called cross-frequency coupling. Cross-frequency coupling can be applied to the analysis of continuous electrophysiological signals such as scalp EEG signals and local field potentials. Previous studies have found that it is useful for predicting epileptic seizures, locating epileptogenic zones, and analyzing cognitive function. It is also associated with certain diseases such as Parkinson's disease, schizophrenia, and social anxiety disorder. For patients undergoing craniotomy or implantable brain-computer interfaces, analysis of cortical EEG signals is crucial for prognosis, but effective signal feature extraction methods for analysis are still lacking. Summary of the Invention

[0004] The present invention aims to propose a method for extracting cross-frequency coupling features of cortical EEG signals in order to obtain the correlation between different frequency components of the signal.

[0005] The present invention provides a method for extracting cross-frequency coupling features of cortical EEG signals, which specifically comprises the following steps: signal preprocessing to reduce various interferences to the cortical EEG signals; and feature extraction, i.e., calculating the cross-frequency coupling strength between each frequency band as a feature.

[0006] (1) Signal preprocessing to reduce various interferences to cortical EEG signals, including:

[0007] Filter out power frequency interference and high-frequency noise;

[0008] Detect abnormal signal segments. The specific method is to add several sliding windows to the signal and judge whether the signal has poor acquisition conditions or large external interference based on the amplitude and change of the signal within the window. If so, the signal segment is marked and distinguished in subsequent processing.

[0009] Noise suppression uses statistical methods to set a signal power threshold and remove signal segments whose power exceeds the threshold to eliminate interference such as artifacts and environmental noise.

[0010] (2) Extracting features, i.e., calculating the cross-frequency coupling strength between frequency bands as features; including:

[0011] Bandpass filtering is used to extract signals in different frequency bands separately, and the cross-frequency coupling strength between signals in different frequency bands is calculated, including three types of coupling strength: phase-amplitude coupling strength, phase-phase coupling strength, and amplitude-amplitude coupling strength. A sliding time window is introduced to calculate the maximum value of the cross-frequency coupling strength over time. The low-frequency components are subdivided to improve the frequency discrimination of the features.

[0012] Further:

[0013] The specific operation process of the signal preprocessing is:

[0014] (1) Filtering: A 50 Hz notch filter is used to filter out power frequency interference and its higher harmonics from the original cortical EEG signal, and a 0.5-140 Hz bandpass filter is then used to filter out high-frequency noise in the signal.

[0015] (2) Abnormal segment removal: During the acquisition process, the acquisition signal is often abnormal due to interference from other devices, accidental touch, or poor electrode adhesion. This is usually manifested as a signal amplitude that is too large or even reaches the upper limit of the acquisition device, or the signal has almost no change for a period of time. A non-overlapping sliding window of 0.1s is used to traverse the signal. If the average absolute value of the signal in the window is greater than 400μV or the maximum value is greater than 500μV, or the sum of the absolute values ​​of the change values ​​of two adjacent points in this window is less than 5μV, then this signal segment is removed.

[0016] (3) Noise suppression: After removing the power frequency and high-frequency noise, the cortical EEG signal still contains electrooculography, electromyography, and other noises brought by the external environment. The superposition of noise will increase the signal energy. Therefore, the window energy of the filtered signal x(n) in a short period of time is calculated and a threshold is set to detect noise. The window energy is:

[0017]

[0018] Where ω[n] is the rectangular window and N is the number of sample points.

[0019] The average value of each channel window energy plus the standard deviation of the window energy value sequence is used as the upper threshold, and windows with energy greater than the threshold are removed. After the above preprocessing process is completed, if the signal of one channel exceeds 70% and is removed, this channel is judged as a bad track and the channel data is discarded.

[0020] The operational process of the feature extraction is:

[0021] Use a bandpass filter to extract sub-signals x of different frequency bands of signal x(n) f (n), and obtain their envelope amplitude sequences through standard Hilbert transform and phase sequence ω f (n), calculate the following features:

[0022] (1) Phase-amplitude coupling strength:

[0023] The f1 frequency band sub-signal x f1 The phase sequence ω of (n) f1 (n) and f2 band signal x f2 Amplitude sequence of (n) According to n, the one-to-one correspondence is obtained, and the function of the amplitude of the sub-signal of the f2 frequency band relative to the phase of the sub-signal of the f1 frequency band is recorded as The range of phase ω [-π,π] is evenly divided into J segments, denoted as d j (j=1,2,…,J), statistics The phase range of each segment d j The amplitude of the f2 frequency band sub-signal distributed in the middle, find its mean, recorded as pa f1-f2 (j) Calculate the normalized distribution PA of the amplitude of the f2 frequency band sub-signal along with the phase of the f1 frequency band sub-signal f1-f2 (j):

[0024]

[0025] In the present invention, J=40 is taken to obtain the normalized distribution of the amplitude of the f2 frequency band sub-signal along with the phase of the f1 frequency band sub-signal. The process example is shown in the attached figure. Figure 2 As shown; find the entropy of this distribution:

[0026]

[0027] The phase-amplitude coupling strength is expressed as the modulation index MI:

[0028]

[0029] (2) Phase-phase coupling strength:

[0030] For two sub-signals x in different frequency bands f1(n) and x f2 The phase sequence ω of (n) f1 (n) and ω f2 (n), phase difference:

[0031] Δω(n)=ω f1 (n)-ω f2 (n), (5)

[0032] The phase-phase coupling strength is expressed using the phase-locked value PLV:

[0033]

[0034] (3) Amplitude-amplitude coupling strength:

[0035] Suppose there are two sub-signals x in different frequency bands f1 (n) and x f2 The power spectra of (n) are P f1 (n) and P f2 (n), with a mean of μ f1 and μ f2 , and the standard deviations are σ f1 and σ f2 , the amplitude-amplitude coupling strength is expressed using the Pearson correlation coefficient:

[0036]

[0037] Compared with the prior art, the present invention has the following advantages:

[0038] (1) The present invention proposes a relatively complete set of preprocessing procedures for cortical EEG signals, adds methods to deal with problems that may be encountered during the acquisition process, removes signal artifacts and channels with poor signal quality, improves the signal-to-noise ratio, and optimizes signal quality;

[0039] (2) The present invention proposes to apply three types of cross-frequency coupling features to the analysis of cortical EEG signals, which can reflect the interaction between different components within the cortical EEG. Compared with other common EEG signal features such as amplitude, power spectrum, burst suppression duty cycle, etc., it focuses more on the interaction between signals of different frequency components and has a stronger detection ability for potential symptoms associated with frequency differences. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is the process of extracting cross-frequency coupling features of cortical EEG signals proposed by the present invention.

[0041] Figure 2 A block diagram of the cross-frequency coupling strength calculation process (taking θ-γ phase-radiation coupling as an example). DETAILED DESCRIPTION

[0042] The present invention is further described below by means of specific embodiments in conjunction with the accompanying drawings:

[0043] The present invention is based on a retrospective study of 100 cases of craniotomy surgery with intraoperative cortical EEG signals recorded by monitoring centers, with a signal sampling rate of 500 Hz.

[0044] (1) Use a 50Hz notch filter to filter out power frequency interference and its higher harmonics, and then use a 0.5-140Hz bandpass filter to filter out high-frequency noise in the signal.

[0045] (2) Use a 0.1s non-overlapping sliding window to traverse the signal. If the mean of the absolute value of the signal in the window is greater than 400μV or the maximum value is greater than 500μV, or the sum of the absolute values ​​of the change values ​​of two adjacent points in this window is less than 5μV, then remove this signal segment.

[0046] (3) Calculate the window energy of the filtered signal in a short period of time and set a threshold to detect noise. Remove windows with energy greater than the threshold. If more than 70% of the signal in one channel is removed, the channel is judged as a bad channel and the channel data is discarded.

[0047] (4) Using a bandpass filter to extract the sub-signals of different frequency bands of the signal, the envelope amplitude sequence and phase sequence are obtained respectively through the standard Hilbert transform. For the sub-signals of all frequency bands, the cross-frequency coupling characteristics between them are calculated, namely the phase-amplitude coupling strength, phase-phase coupling strength and amplitude-amplitude coupling strength. Specifically, the six frequency bands of the sub-signals used are δ (1-4 Hz), θ (4-8 Hz), α (8-14 Hz), β (14-30 Hz), γ (30-60 Hz), and broadband γ (60-140 Hz); five subdivided low-frequency bands (0.5-1.5 Hz, 1.5-2.5 Hz, 2.5-3.5 Hz, 3.5-4.5 Hz, and 4.5-5.5 Hz) are also calculated. For the six frequency bands and the subdivided low-frequency bands, the three cross-frequency coupling strengths between them are calculated.

[0048] (5) The obtained features were tested for performance in the application of distinguishing epileptic seizures. Compared with the method of using other cortical EEG signal features or using a certain coupling strength as a feature, the method proposed in the present invention has a better effect in distinguishing whether an epileptic seizure occurs. 20 cases of epileptic seizures were selected from 100 EEG data sets, and 20 cases without epileptic seizures were randomly selected. The features were extracted using the method proposed in the present invention to distinguish the possibility of epileptic seizures. The results without preprocessing, the results of using only some features, and the results obtained by using traditional cortical EEG signal features (amplitude, power spectrum, burst suppression duty cycle) were compared. The comparison results are shown in Table 1. The results show that the preprocessing steps proposed in the present invention can effectively improve the influence of signal quality on feature extraction, and the three types of cross-frequency coupling features used have better accuracy when using cortical EEG signals for distinguishing epileptic seizures.

[0049] Table 1. Results of different feature sets and preprocessing methods for the discrimination of epileptic seizure possibility

[0050] Feature Set Accuracy (%) Sensitivity (%) Specificity (%) Cross-frequency coupling characteristics 77.5 75.0 80.0 Cross-frequency coupling characteristics (without preprocessing) 60.0 50.0 70.0 Only phase-phase coupling strength is used 70.0 70.0 70.0 Using only the amplitude-amplitude coupling strength 62.5 60.0 65.0 Using only the phase-amplitude coupling strength 72.5 75.0 70.0 Using traditional cortical EEG signal features 70.0 75.0 65.0 .

Claims

1. A method for extracting cross-frequency coupling features of cortical EEG signals, characterized in that: The specific steps are: (1) Signal preprocessing to reduce various interferences to cortical EEG signals, including: Filter out power frequency interference and high-frequency noise; Signal abnormality detection: Add several sliding windows to the signal and judge whether the signal has poor acquisition conditions or large external interference based on the amplitude and change of the signal within the window. If so, mark the signal and distinguish it in subsequent processing; Noise suppression: Use statistical methods to set the signal power threshold and remove signal segments whose power exceeds the power threshold to eliminate artifacts and environmental noise interference; (2) Extracting features, i.e., calculating the cross-frequency coupling strength between frequency bands as features; including: extracting signals from different frequency bands separately using bandpass filtering, calculating the cross-frequency coupling strength between signals from different frequency bands, including three types of coupling strength: phase-amplitude coupling strength, phase-phase coupling strength, and amplitude-amplitude coupling strength; and introducing a sliding time window to calculate the maximum value of the cross-frequency coupling strength over time; wherein the frequency discrimination of the features is improved by subdividing the low-frequency components; The operation process of extracting features is as follows; Use bandpass filter to extract signal sub-x of different frequency bands of signal x(n) f (n), and obtain their envelope amplitude sequences through standard Hilbert transform and phase sequence ω f (n), calculate the following features: (1) Phase-amplitude coupling strength: The f1 frequency band sub-signal x f1 The phase sequence ω of (n) f1 (n) and f2 frequency band sub-signal x f2 Amplitude sequence of (n) According to n, the one-to-one correspondence is obtained, and the function of the amplitude of the sub-signal of the f2 frequency band relative to the phase of the sub-signal of the f1 frequency band is recorded as The range of phase ω [-π,π] is evenly divided into J segments, denoted as d j (j=1,2,…,J), statistics The phase range of each segment d j The amplitude of the f2 frequency band sub-signal distributed in the middle, find its mean, recorded as pa f1-f2 (j) Calculate the normalized distribution PA of the amplitude of the f2 frequency band sub-signal along with the phase of the f1 frequency band sub-signal f1-f2 (j): Get the normalized distribution of the amplitude of the f2 frequency band sub-signal along with the phase of the f1 frequency band sub-signal, and calculate the entropy of the normalized distribution: The phase-amplitude coupling strength is expressed as the modulation index MI: (2) Phase-phase coupling strength: For two sub-signals x in different frequency bands f1 (n) and x f2 The phase sequence ω of (n) f1 (n) and ω f2 (n), phase difference: Give(n)=ω f1 (n)-ω f2 (n) (5) The phase-phase coupling strength is expressed using the phase-locked value PLV: (3) Amplitude-amplitude coupling strength: Suppose there are two sub-signals x in different frequency bands f1 (n) and x f2 The power spectra of (n) are P f1 (n) and P f2 (n), with a mean of μ f1 and μ f2 , and the standard deviations are σ f1 and σ f2 , the amplitude-amplitude coupling strength is expressed using the Pearson correlation coefficient:

2. The method for extracting cross-frequency coupling features of cortical EEG signals according to claim 1, characterized in that: The operational flow of signal preprocessing is as follows; (1) Filtering: Use a 50 Hz notch filter to filter out power frequency interference and its higher harmonics, and then use a 0.5-140 Hz bandpass filter to filter out high-frequency noise in the original cortical EEG signal; (2) Abnormal segment removal: Use a 0.1s non-overlapping sliding window to traverse the signal. If the mean absolute value of the signal within the window is greater than 400μV or the maximum value is greater than 500μV, or the sum of the absolute values ​​of the change values ​​of two adjacent points within the window is less than 5μV, then remove the signal segment; (3) Noise suppression: Calculate the window energy of the filtered signal x(n) in a short period of time and set a threshold to detect noise. The window energy is: Where ω[n] is a rectangular window and N is the number of sample points; The average value of each channel window energy plus the standard deviation of the window energy value sequence is used as the threshold, and the windows with energy greater than the threshold are removed; After the above pre-processing process is completed, if more than 70% of the signal of one channel is removed, the channel is determined to be a bad track and the data of the channel is discarded.

Images