Universal neural signal feature extraction method

Abstract

The invention discloses a universal neural signal feature extraction method. According to the method, a difference order formula constructed based on a sampling frequency and a target center frequency is provided. The multi-order difference is carried out on the pre-processed neural signal by using the difference order obtained by the formula, clutters except the target frequency can be filtered ideally, better decoding performance can be obtained, and meanwhile, the calculation amount can be greatly reduced.

Description

Technical Field

[0001] The present invention relates to the field of brain-computer interface, and in particular to a universal neural signal feature extraction method. Background Art

[0002] Neural signal feature extraction plays a pivotal role in brain-computer interface (BCI) technology and is the core link in the signal processing process. This process is dedicated to accurately extracting the feature information that is crucial to solving specific problems from the complex raw EEG signals. These features are like keys that can unlock the door to subsequent recognition, classification, and control signal conversion.

[0003] One of the core missions of feature extraction is to achieve dimensionality reduction and compression of data. This means that we can remove redundancy and noise while retaining key information, thereby representing the original data with a more streamlined feature subset. This transformation has greatly promoted the efficient use of machine learning and artificial intelligence algorithms in classification and diagnosis tasks, allowing the algorithms to focus more on the essence of the problem.

[0004] More importantly, through the carefully designed feature extraction strategy, the classification accuracy of the BCI system has been significantly improved. This progress is of immeasurable value for comprehensively optimizing the performance of the BCI system and promoting its wide application in medical rehabilitation, human-computer interaction and other fields. Therefore, neural signal feature extraction is not only a technical challenge, but also the key to promoting BCI technology to a higher level.

[0005] Commonly used feature extraction methods include: time domain features such as average amplitude, line length, etc. These methods are fast in processing, but may not always produce the most relevant and robust features; frequency domain features involve spectral analysis of the signal to extract frequency-related features, such as power spectral density (PSD), etc.; time-frequency domain methods: combining time domain and frequency domain features, common techniques include short-time Fourier transform (STFT) and S-transform, etc.; decomposition domain feature extraction: such as wavelet transform and empirical mode decomposition (EMD), these methods allow the signal to be filtered and decomposed simultaneously to extract useful features. These methods can be used alone or in combination to improve the performance and accuracy of BCI systems.

[0006] In recent years, a variety of new feature extraction methods have been proposed, achieving better neural decoding results. The literature Nason et al., 2020, Nature Biomedical Engineering disclosed the spiking-band power (SBP), which is generated by local neuronal discharge activity and can decode brain signals more accurately with low energy consumption, thereby improving the accuracy and efficiency of control. The literature Ahmadi et al., 2021, Journal of Engineering disclosed the entire spiking activity (ESA), which achieved higher signal decoding performance than any other input. This method can maintain high decoding performance even when the spike is removed from the original signal, a different number of channels is used, and less training data is used.

[0007] However, the above method has a large amount of calculation and is not suitable for implementation on hardware. In addition, the above method is only suitable for processing high-frequency intracortical neural signals and cannot solve the feature extraction of all neural signals. In view of this, the present invention aims to provide a universal neural signal feature extraction method, which can significantly reduce the amount of calculation while maintaining the computing performance to adapt to different application scenarios and data types. Summary of the invention

[0008] The present invention provides a universal neural signal feature extraction method, which can significantly reduce the amount of calculation while maintaining good decoding performance.

[0009] The present invention provides a general neural signal feature extraction method, comprising:

[0010] removing common noise signals from the sampled multi-channel neural signals to obtain preprocessed neural signals;

[0011] Perform k-order difference on the preprocessed neural signal to obtain a differential signal, and perform full-wave rectification on the differential signal to obtain a target characteristic signal;

[0012] The difference order k is:

[0013] k=round(f s / 2f d )

[0014] Among them, round(·) is the rounding function, f s is the sampling frequency, f d is the center frequency of the target feature.

[0015] Preferably, the method for obtaining the preprocessed neural signal includes: subtracting the median signal of all channels at each moment from the neural signal of the current channel. The above method can preliminarily remove noise.

[0016] Preferably, the differential signal s k (t) is

[0017] s k (t) = s(t) - s(tk)

[0018] Where t is the sampling time point in the neural signal.

[0019] Preferably, the center frequency f of the target feature d ≤f s / 2. The neural signal features at the center frequency in the first half of the sampling frequency can be extracted better.

[0020] Preferably, the target characteristic signal s f (t) is:

[0021] s f (t) = abs(s k (t))

[0022] Where abs(·) is the absolute value function and t is the sampling time point in the neural signal.

[0023] Preferably, the obtained target characteristic signal is further smoothed by windowed averaging.

[0024] Preferably, the window length is 50-300 ms and the step length is 10-100 ms.

[0025] Preferably, when extracting features of neural signals in the cortex, the sampling frequency is 30 kHz, the center frequency is 1 kHz, and the difference order is 15.

[0026] Compared with the prior art, the present invention has the following beneficial effects:

[0027] The present invention proposes for the first time a differential order formula constructed based on sampling frequency and center frequency. The differential order obtained by the formula is used to differentiate the preprocessed neural signal, which can filter out noise more clearly, achieve better decoding performance, and significantly reduce the amount of calculation. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0029] Figure 1 A schematic flow chart of a general neural signal feature extraction method provided in a specific embodiment of the present invention;

[0030] Figure 2 A comparison diagram of the computational time consumption of a general neural signal feature extraction method provided in a specific embodiment of the present invention and full spike signal (ESA) and spike band power (SBP) features;

[0031] Figure 3 A comparison diagram of the decoding performance of the general neural signal feature extraction method provided in a specific embodiment of the present invention and the full spike signal (ESA) and spike band power (SBP) features;

[0032] Figure 4 A frequency response diagram provided by a specific embodiment of the present invention. DETAILED DESCRIPTION

[0033] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0034] It should be noted that, in the absence of conflict, the features in the following embodiments and implementations may be combined with each other.

[0035] The specific embodiment of the present invention provides a general neural signal feature extraction method, such as Figure 1 As shown, including:

[0036] (1) Preprocessing the pulse neural signal: removing the common noise signal from the sampled multi-channel neural signal to obtain the preprocessed neural signal.

[0037] Specifically, the common noise signal is removed from the sampled multi-channel neural signal to obtain the preprocessed neural signal 。

[0038] (2) The required differential order is extracted based on the center frequency of the target feature. The differential order k is:

[0039] k=round(f s / 2f d )

[0040] Among them, round(·) is the rounding function, f s is the sampling frequency, f dis the center frequency of the target feature, the center frequency of the target feature f d ≤f s / 2, using the differential order obtained by the formula for differentiation can efficiently obtain the target characteristic signal of the required center frequency. The formula is equivalent to realizing bandpass filtering. The differential order provided by the present invention can better realize filtering, thereby achieving better decoding performance.

[0041] (3) In a specific embodiment of the present invention, a k-order difference is performed on the preprocessed neural signal to obtain a differential signal, and a full-wave rectification is performed on the differential signal to obtain a target characteristic signal, thereby completing the feature extraction of the neural pulse signal.

[0042] Specifically, the differential signal s provided in the specific embodiment of the present invention is k (t) is

[0043] s k (t) = s(t) - s(tk)

[0044] Where t is the sampling time point in the neural signal.

[0045] Specifically, the target characteristic signal s provided in the specific embodiment of the present invention is f (t) is:

[0046] s f (t) = abs(s k (t))

[0047] Where abs(·) is the absolute value function and t is the sampling time point in the neural signal.

[0048] Specifically, the target characteristic signal obtained by further smoothing the windowed average is further smoothed in a specific embodiment of the present invention, and the parameters of the windowed average are a window length of 50-300 ms and a step length of 10-100 ms.

[0049] In a specific embodiment, taking a high-density neural signal collected in the cortex as an example, the neural signal feature extraction method in the embodiment of the present invention may include the following steps:

[0050] 1. Preprocessing of neural pulse signals. For the original neural pulse signals, calculate the median sequence of all signal channels in the time dimension, then reduce each signal channel to the median sequence, remove the common noise signal between channels, and complete the preprocessing;

[0051] 2. Calculate the differential order. The sampling frequency f of the neural pulse signal s =30kHz, while the frequency distribution of the effective component of the nerve impulse signal, the spike potential, is around 1kHz. Therefore, the center frequency f d =1kHz, calculate the required differential order k=round(fs / 2f d )=15;

[0052] 3. Calculate the k-order difference value s of the neural signal k (t). Perform 15th-order difference on the neural pulse signal, the specific form is:

[0053] s 15 (t) = s(t) - s(t-15)

[0054] Among them, s(t) refers to the preprocessed neural pulse signal, s 15 (t) refers to the signal after 15th order difference;

[0055] 4. Full-wave rectification. Then the differential signal s 15 (t) Full-wave rectification is performed, the specific form is:

[0056] s f (t) = abs(s 15 (t))

[0057] Here, abs(·) refers to the absolute value function.

[0058] The universal neural signal feature extraction method proposed in the present invention is verified in an actual clinical brain-computer interface dataset.

[0059] The subjects of this clinical trial experienced a C4 spinal cord injury, which resulted in a complete loss of sensation and movement below the shoulders. In the experiment, a computer screen was placed in front of the subjects, and the cursor displayed on the screen wrote in the stroke order of Chinese characters. The subjects were asked to simulate writing the Chinese characters in their minds according to the movement trajectory of the cursor. While the subjects were imagining writing, the researchers recorded the subjects' original nerve impulse signals in real time, and simultaneously recorded the speed at which the cursor wrote the Chinese characters and whether the cursor was in the stroke writing state.

[0060] The original neural pulse signal recorded in this embodiment uses the general neural signal feature extraction method provided in the specific embodiment of the present invention to obtain the mean feature of the absolute value of the difference, and at the same time, the sliding average is performed through a sliding average window with a window length of 200ms and a step length of 50ms to align the speed labels of the corresponding time to construct a clinical brain-computer interface data set. The effective duration of the data set signal is 3705 seconds, including 96 channels of neuronal pulse signal pathways in the motor cortex brain area.

[0061] The experimental results obtained by the general neural signal feature extraction method provided in this embodiment are compared with the full spike signal (ESA) and spike band power (SBP) features. The feature extraction process of the full spike signal (ESA) is: 1. 300Hz first-order Butterworth high-pass filtering; 2. Full-wave rectification; 3. 12Hz first-order Butterworth low-pass filtering; 4. Downsampling. The feature extraction process of the spike band power (SBP) is: 1. 300-1000Hz second-order Butterworth band-pass filtering; 2. Full-wave rectification; 3. Downsampling. In order to maintain the same frequency as the kinematic label, both features are averaged through a sliding average window with a window length of 200ms and a step size of 50ms.

[0062] Figure 2 This is a comparison chart of the computational time of this method and two signal features, where the x-axis is the sequence length and the y-axis is the computational time in seconds. The experiment was repeated 100 times for each point. The average computational time of the full spike signal (ESA) feature is 4.51 times the mean of the absolute value of the difference, and the average computational time of the spike band power (SBP) feature is 3.20 times the mean of the absolute value of the difference.

[0063] Figure 3 This is a comparison chart of the decoding performance of this method and two signal features, where the mean square error is a commonly used metric for decoding performance, ns means that there is no statistical difference between the mean of the absolute difference and the decoding performance of the full spike signal (ESA) feature, and *** means that the decoding performance of the mean of the absolute difference is statistically significantly better than the spike band power (SBP) feature. The above results show that compared with traditional high-frequency continuous neural signal features, this method can maintain consistent decoding performance in the context of using neural signal decoding kinematics, but the computational time is greatly reduced.

[0064] Figure 4 It is the frequency response diagram corresponding to the feature extraction method of the present invention. It can be seen that the equivalent effect of the feature extraction method of the present invention is a multi-passband bandpass filter, in which the first passband is at a frequency f d (i.e. f s / 2k) as the center, [f s / 3k+f s / k,2f s / 3k+f s / k] is within the range of -3dB bandwidth. Since the amount of information of the neural signal in other passbands is already very small, the latter passbands can be ignored. Therefore, the present invention can obtain the frequency of interest f through simple multi-order difference calculation d Nearby neural signal features are used to perform efficient feature extraction.

[0065] The specific embodiment of the present invention provides a universal calculation formula for the differential order, which is used to extract signal features in a specific frequency band; it also extracts features through a multi-order differential method, which greatly simplifies the amount of calculation; and the universal neural signal feature extraction method provided by the specific embodiment of the present invention is suitable for feature extraction of any time series signal including neural signals.

[0066] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A general neural signal feature extraction method, characterized in that: include: removing common noise signals from the sampled multi-channel neural signals to obtain preprocessed neural signals; Perform k-order difference on the preprocessed neural signal to obtain a differential signal, and perform full-wave rectification on the differential signal to obtain a target characteristic signal; The difference order k is: k=round(f s / 2f d ) Among them, round(·) is the rounding function, f s is the sampling frequency, f d is the center frequency of the target feature.

2. The general neural signal feature extraction method according to claim 1, characterized in that: The method for obtaining a preprocessed neural signal includes: subtracting the median signal of all channels at each moment from the neural signal of the current channel.

3. The general neural signal feature extraction method according to claim 1, characterized in that: The differential signal s k (t) is s k (t)=s(t)-s(t-k) Where t is the sampling time point in the neural signal.

4. The general neural signal feature extraction method according to claim 1, characterized in that: The center frequency f of the target feature d ≤f s / 2.

5. The general neural signal feature extraction method according to claim 1, characterized in that: The target characteristic signal s f (t) is: s f (t)=abs(s k (t)) Where abs(·) is the absolute value function and t is the sampling time point in the neural signal.

6. The general neural signal feature extraction method according to claim 1, characterized in that: The target feature signal is further smoothed by windowed averaging.

7. The general neural signal feature extraction method according to claim 6, characterized in that: The window length is 50-300ms and the step length is 10-100ms.

8. The general neural signal feature extraction method according to claim 1, characterized in that: When extracting features from neural signals in the cortex, the sampling frequency is 30 kHz, the center frequency is 1 kHz, and the difference order is 15.

Images