Brain wave acquisition and analysis method for sleep staging

Abstract

The invention relates to the technical field of intelligent brain waves, and particularly provides a brain wave acquisition and analysis method for sleep staging, which comprises the following steps: S1, acquiring brain wave initial data, and sequentially carrying out downsampling processing, scaling processing, Gaussian filtering processing, Gaussian smoothing processing and data overlapping filling processing on the initial data; s2, carrying out labeling operation on the processed brain wave data; s3, collecting a human body sleep image, and judging the human body sleep time in combination with the brain wave data; s4, acquiring a time point of deep sleep of the human body; s5, building a sleep neural network to judge the sleep state of the human body; according to the method, data collection and processing and label creation are completed through the steps S1-S5, and then a sleep neural network of a user is trained by using a data set and the label, so that the judgment of the sleep stage of the human body by using a machine is completed, the accuracy of sleep stage division is improved, effective sleep information is identified, and the monitoring of the whole sleep process is realized.

Description

Technical Field

[0001] The present invention relates to the field of intelligent brain wave technology, and in particular to a brain wave collection and analysis method for sleep staging. Background Art

[0002] As we all know, sleep is an indispensable physiological phenomenon for human beings. About 1 / 3 of a person's life is spent in sleep. Good sleep can eliminate human fatigue, allow the organs of the human body to get enough rest, and prepare for the physiological activities of the human body when awake. However, the accelerated pace of life and increased pressure of life of contemporary people have caused more and more people to have sleep problems, which seriously affects people's daily life and physical health. The sleep process can be divided into different sleep stages according to the different representations of physiological signals. The AASM sleep staging criteria divide the sleep cycle into wakefulness, N1, N2, N3 and REM. EEG signals are important physiological signals that reflect human brain activity and have rich information content. By collecting and analyzing EEG signals, we can understand the emotional state or cognitive function of the subject, providing valuable reference for mental health, education, entertainment and other fields.

[0003] Traditional EEG signal data classification and recognition methods are usually based on artificially extracted feature parameters, such as power spectrum, coherence, entropy, etc. These feature parameters often cannot fully express the inherent laws of EEG signals, resulting in unsatisfactory data classification and recognition effects and low utilization rate. Summary of the invention

[0004] In view of the above-mentioned shortcomings of the prior art, an object of the present invention is to provide a method for collecting and analyzing brain waves for sleep staging, so as to solve the problem of low utilization rate of brain wave signal data composed of artificially extracted feature parameters in the prior art.

[0005] To achieve the above-mentioned and other related purposes, the present invention provides a method for collecting and analyzing brain waves for sleep staging, the method comprising the following steps:

[0006] S1, collecting brain wave data, the collected initial data is 512Hz, and the initial data is sequentially subjected to downsampling processing, scaling processing, Gaussian filtering processing, Gaussian smoothing processing, and data overlap filling processing;

[0007] S2, labeling the processed EEG data;

[0008] S3, collecting human sleep images and combining brain wave data to determine the time when the human falls asleep;

[0009] S4, obtaining the time point of deep sleep of the human body;

[0010] S5 builds a sleep neural network to determine the sleep state of a human body, wherein the construction of the sleep neural network includes convolution operation, normalization, activation function, pooling, regularization, and feature splicing.

[0011] In one embodiment of the present invention, in step S1, the formula used in the downsampling process is:

[0012] x 100Hz [n] = x[m·r]

[0013] in

[0014] x is the original data, x 100Hz This is the data after downsampling the EEG data from the original 512Hz to 100Hz.

[0015] In one embodiment of the present invention, in step S1, the formula used in the scaling process is:

[0016]

[0017] y is the scaled data.

[0018] In one embodiment of the present invention, in step S1, the formula used for the Gaussian filter is:

[0019] f min =0.5Hz,f max =100Hz

[0020] m={y i |y i ≥f min ∧y i ≤f max ,i∈[1,N]}

[0021] z={y i |m i =1}

[0022] Where m is a Boolean array identifying data points within the frequency range;

[0023] y i It is the data for judging the completion of scaling;

[0024] f min is the minimum frequency threshold when extracting data;

[0025] f max The maximum frequency threshold when extracting data;

[0026] z is the valid data that meets the screening conditions.

[0027] In one embodiment of the present invention, in step S1, the formula for Gaussian smoothing is:

[0028]

[0029]

[0030] k is the size of the convolution kernel;

[0031] σ is the standard deviation of the convolution kernel;

[0032] K is a Gaussian kernel generated according to the specified convolution kernel size and standard deviation of the convolution kernel;

[0033] i represents the position in the convolution kernel;

[0034] F is the data after applying Gaussian filtering, which represents the result after smoothing.

[0035] In one embodiment of the present invention, in step S1, the data overlapping filling process includes: dividing every 3000 data into one block, and then overlapping the blocks to achieve data expansion;

[0036] N is the total length of F after data processing;

[0037] B is the block data;

[0038] O is the overlapping data

[0039] M is the number of blocks;

[0040] (1) Calculate the number of blocks M:

[0041]

[0042] (2) Calculate the starting S for each block i, where i = 0, 1, ..., M-1 i and end index E i :

[0043] S i =i·(BO),E i =S i +B

[0044] (3) Merge all blocks into one expanded data

[0045] E=min(B+S i ,N)

[0046] (4) Merge data

[0047]

[0048] di is the data of the i-th block, and X is the merged data.

[0049] In one embodiment of the present invention, in step S2, the labeling operation on the processed EEG data includes:

[0050] Define a label for every 3,000 data points and calculate the number of labels n required for a piece of data;

[0051]

[0052] Where N is the total amount of data.

[0053] In one embodiment of the present invention, in step S3, a video is recorded while a person is sleeping, and the state of muscle relaxation and heart rate change of the person are observed by combining the video and the collected brain wave data to determine the time when the person falls asleep;

[0054] Define the time to fall asleep as t1 seconds, and The label is recorded as sober; the rest This label is temporarily recorded as light sleep.

[0055] In one embodiment of the present invention, in step S4, each 30 seconds of data is analyzed to find the point with a higher average spectrum amplitude as the time point of deep sleep;

[0056] (1) Calculate FFT:

[0057]

[0058] (2) Calculation frequency:

[0059]

[0060] (3) Calculate the amplitude spectrum:

[0061] m=|F|

[0062] (4) Perform frequency range screening, i.e., screening the range of delta waves (0.5 Hz to 4 Hz):

[0063] M=(f≥0.5)∧(f≤4)

[0064]

[0065] Where N is the data length, r is the sampling rate, A is the data sample, F is the fft result, f is the frequency, m is the amplitude spectrum, and m a is the average amplitude value, M is the frequency mask; by comparing and observing the spectrum amplitude, the higher point is recorded as the time point of deep sleep.

[0066] In one embodiment of the present invention, in step S5, (1) the convolution operation includes the following:

[0067] The input signal is recorded as x, the convolution kernel is recorded as w, and the formula used for the convolution operation is:

[0068]

[0069] Where K is the size of the convolution kernel, y[n] is the output of the convolution, n represents the position of the output signal, and m represents the index of the current element used by the convolution kernel when performing calculations;

[0070] (2) Let the input signal be x and the output signal be The formula used for the normalization is:

[0071]

[0072]

[0073]

[0074] The final output normalization result is:

[0075]

[0076] Where ε is a small positive number, γ and β are learnable parameters;

[0077] (3) The activation function adopts the formula:

[0078] f(x)=max(0,x)

[0079] The activation function keeps positive values ​​unchanged and changes negative values ​​to 0, which is used to introduce nonlinear factors to avoid overfitting of the model:

[0080] (4) The formula used for the pooling is:

[0081]

[0082] Where x is the input signal and k is the pooling window size parameter;

[0083] (5) The regularization randomly discards neurons with probability p, A is the discarded neuron, and the formula used is:

[0084]

[0085] (6) The feature concatenation adopts a fully connected layer, and the matrix formula is input into the fully connected layer:

[0086] y=Wx+b

[0087] In the formula, x is the input vector, W is the weight matrix, and b is the bias; the value of the matrix y is: y = [abc], where a, b, and c represent the scores of the three categories respectively, then the following formula is obtained:

[0088]

[0089] Among them, 0 represents wakefulness, 1 represents light sleep, and 2 represents deep sleep. The y matrix is ​​used to judge the human body's sleep state.

[0090] As described above, the brain wave collection and analysis method for sleep staging of the present invention has the following beneficial effects:

[0091] 1. After the brain wave data is collected, the present invention sequentially performs downsampling processing, scaling processing, Gaussian filtering processing, Gaussian smoothing processing, and data overlapping filling processing. The downsampling processing can reduce the collected 512Hz original data to 100Hz, so as to avoid the excessive amount of data per second affecting the subsequent neural network training process, which is beneficial to improving the neural network training effect and reducing the training time; the scaling processing can reduce the change of brain wave data and improve the stability of brain wave data; the Gaussian filtering can effectively reduce the noise of the brain wave data; the Gaussian smoothing processing can process the brain wave data to make it smoother as a whole, reduce the influence of noise, and improve the data quality; the data overlapping filling processing can fill the original data by overlapping data, so that the number of original data is complete; after the S1 step processing, the brain wave data that is easy to analyze can be obtained, which improves the effective utilization rate of the collected brain wave data and reduces the difference of individual data.

[0092] 2. The present invention labels the processed EEG data and combines it with human sleep images, which can clearly divide the sleep stages, improve the accuracy of sleep staging, and facilitate subsequent judgment of sleep cycles.

[0093] 3. The sleep neural network constructed by the present invention has convolution operations that can extract features, normalization that can stabilize the training process and accelerate convergence, activation functions that can introduce nonlinear factors to avoid model overfitting, pooling that can reduce the spatial size of feature maps during feature extraction while retaining important feature information, regularization that can improve the robustness of the network, simplify the model, and enhance training efficiency, and feature splicing that can merge large and small convolution branches into a fully connected layer to achieve a feature fusion effect.

[0094] 4. The present invention completes data collection and processing, label creation through steps S1-S5, and then uses the data set and labels to train our own sleep neural network, thereby completing the use of the machine to judge the human sleep stage, improving the accuracy of sleep staging, identifying effective sleep information, and realizing the monitoring of the entire sleep process. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] Figure 1 Shown is a flow chart of the EEG acquisition and analysis method for sleep staging disclosed in the present invention.

[0096] Figure 2 Shown is a sleep stage diagram obtained over time using the present invention.

[0097] Figure 3 The display shows the average amplitude value of 0.5-4 Hz obtained by the present invention over time. DETAILED DESCRIPTION

[0098] The following is a description of the implementation of the present invention by means of specific embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification.

[0099] See also Figure 1-3 It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings in this specification are only used to match the contents disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any structural modification, change in proportion or adjustment of size should still fall within the scope of the technical contents disclosed in the present invention without affecting the effects and purposes that can be achieved by the present invention.

[0100] See also Figure 1 The present invention provides a method for collecting and analyzing brain waves for sleep staging, the method comprising the following steps:

[0101] S1, collecting brain wave data, the collected initial data is 512Hz, and the initial data is sequentially subjected to downsampling processing, scaling processing, Gaussian filtering processing, Gaussian smoothing processing, and data overlap filling processing;

[0102] The formula used in the downsampling process is:

[0103] x 100Hz [n] = x[m·r]

[0104] in

[0105] x is the original data, x 100Hz This is the data after downsampling the EEG data from the original 512Hz to 100Hz.

[0106] The formula used in the scaling process is:

[0107]

[0108] y is the scaled data.

[0109] The formula used for the Gaussian filter is:

[0110] f min =0.5Hz,f max =100Hz

[0111] m={y i |y i ≥f min ∧y i ≤f max ,i∈[1,N]}

[0112] z={y i |m i =1}

[0113] Where m is a Boolean array identifying data points within the frequency range;

[0114] y i It is the data for judging the completion of scaling;

[0115] f min is the minimum frequency threshold when extracting data;

[0116] f max The maximum frequency threshold when extracting data;

[0117] z is the valid data that meets the screening conditions;

[0118] The formula for the Gaussian smoothing process is:

[0119]

[0120]

[0121] k is the size of the convolution kernel;

[0122] σ is the standard deviation of the convolution kernel;

[0123] K is a Gaussian kernel generated according to the specified convolution kernel size and standard deviation of the convolution kernel;

[0124] i represents the position in the convolution kernel;

[0125] F is the data after applying Gaussian filtering, which means the result after smoothing;

[0126] The data overlapping filling process includes: dividing every 3000 data into one block, and then overlapping the blocks to achieve data expansion;

[0127] N is the total length of F after data processing;

[0128] B is the block data;

[0129] O is the overlapping data

[0130] M is the number of blocks;

[0131] (1) Calculate the number of blocks M:

[0132]

[0133] (2) Calculate the starting S for each block i, where i = 0, 1, ..., M-1 i and end index E i :

[0134] S i =i·(BO),E i =S i +B

[0135] To ensure that the end index does not exceed the index range;

[0136] (3) Merge all blocks into one expanded data

[0137] E=min(B+S i ,N)

[0138] (4) Merge data

[0139]

[0140] d i is the data of the i-th block, and X is the merged data.

[0141] In this step, after the EEG data is collected, downsampling, scaling, Gaussian filtering, Gaussian smoothing, and data overlapping and filling are performed in sequence. The downsampling process can reduce the collected 512Hz original data to 100Hz, avoiding the excessive amount of data per second affecting the subsequent neural network training process, which is beneficial to improving the neural network training effect and reducing the training time; the scaling process can reduce the change of EEG data and improve the stability of EEG data; the Gaussian filtering can effectively reduce the noise of the EEG data; the Gaussian smoothing process can process the EEG data to make it smoother as a whole, reduce the impact of noise, and improve data quality; the data overlapping and filling process can fill the original data by overlapping data to make the number of original data complete; after processing in step S1, easy-to-analyze EEG data can be obtained, which improves the effective utilization rate of collected EEG data and reduces the difference in individual data.

[0142] S2, labeling the processed EEG data;

[0143] The labeling operation of the processed EEG data includes:

[0144] Define a label for every 3,000 data points and calculate the number of labels n required for a piece of data;

[0145] Where N is the total amount of data.

[0146] S3, collecting human sleep images, combining brain wave data to determine the time when the human falls asleep; recording the human body while sleeping, combining the recording and the collected brain wave data to observe the state of muscle relaxation and heart rate changes of the human body, and determining the time when the human body falls asleep, wherein the recording of the human body includes the process from wakefulness to sleep, and the recording ends when the human body just wakes up after sleep. The recording and brain wave collection are carried out simultaneously to ensure the correspondence between the data and sleep;

[0147] Define the time to fall asleep as t1 seconds, and The label is recorded as sober; the rest This label is temporarily recorded as light sleep.

[0148] S4. Obtain the time point of deep sleep of the human body; since the delta wave is more active during deep sleep, and the traditional spectrum analysis always analyzes all the data at one time, which is different from our purpose, here we analyze the data of every 30 seconds and find the point with higher average spectrum amplitude as the time point of deep sleep;

[0149] (1) Calculate FFT:

[0150]

[0151] (2) Calculation frequency:

[0152]

[0153] (3) Calculate the amplitude spectrum:

[0154] m=|F|

[0155] (4) Perform frequency range screening, i.e., screening the range of delta waves (0.5 Hz to 4 Hz):

[0156] M=(f≥0.5)∧(f≤4)

[0157]

[0158] Where N is the data length, r is the sampling rate, A is the data sample, F is the fft result, f is the frequency, m is the amplitude spectrum, and m a is the average amplitude value, M is the frequency mask; by comparing and observing the spectrum amplitude, the higher point is recorded as the time point of deep sleep.

[0159] Steps S2-S4 label the processed EEG data and combine them with human sleep images, which can clearly divide the sleep stages, improve the accuracy of sleep staging, and facilitate subsequent judgment of sleep cycles.

[0160] See also Figure 2-3 S5 builds a sleep neural network to determine the sleep state of a human body, and the construction of the sleep neural network includes convolution operation, normalization, activation function, pooling, regularization, and feature splicing.

[0161] (1) The convolution operation, i.e., feature extraction, includes the following:

[0162] The input signal is recorded as x, the convolution kernel is recorded as w, and the formula used for the convolution operation is:

[0163]

[0164] Where K is the size of the convolution kernel, y[n] is the output of the convolution, n represents the position of the output signal, and m represents the index of the current element used by the convolution kernel when performing calculations;

[0165] (2) Let the input signal be x and the output signal be The purpose of the normalization is to stabilize the training process and accelerate convergence. The formula used for normalization is:

[0166]

[0167]

[0168]

[0169] The final output normalization result is:

[0170]

[0171] Where ε is a small positive number, γ and β are learnable parameters;

[0172] (3) The activation function adopts the formula:

[0173] f(x)=max(0,x)

[0174] The activation function keeps positive values ​​unchanged and changes negative values ​​to 0, which is used to introduce nonlinear factors to avoid overfitting of the model:

[0175] (4) The formula used for the pooling is:

[0176]

[0177] Where x is the input signal and k is the pooling window size parameter;

[0178] The purpose of pooling is to reduce the spatial size of the feature map during feature extraction while retaining important feature information.

[0179] (5) The regularization randomly discards neurons with probability p, A is the discarded neuron, and the formula used is:

[0180]

[0181] Regularization improves the robustness of the network, simplifies the model, and enhances training efficiency

[0182] (6) The feature concatenation adopts a fully connected layer, and the matrix formula is input into the fully connected layer:

[0183] y=Wx+b

[0184] In the formula, x is the input vector, W is the weight matrix, and b is the bias; the value of the matrix y is: y = [abc], where a, b, and c represent the scores of the three categories respectively, then the following formula is obtained:

[0185]

[0186] Among them, 0 represents wakefulness, 1 represents light sleep, and 2 represents deep sleep. The y matrix is ​​used to judge the human body's sleep state.

[0187] The sleep neural network built in this step has convolution operations that can extract features, normalization that can stabilize the training process and accelerate convergence, activation functions that can introduce nonlinear factors to avoid model overfitting, pooling that can reduce the spatial size of feature maps during feature extraction while retaining important feature information, regularization that can improve the robustness of the network, simplify the model, and enhance training efficiency, and feature concatenation that can merge large and small convolution branches into the fully connected layer to achieve feature fusion effects.

[0188] In summary, the present invention completes data collection and processing, label creation through steps S1-S5, and then uses the data set and labels to train our own sleep neural network, thereby completing the judgment of human sleep stages using a machine, improving the accuracy of sleep staging, identifying effective sleep information, and realizing monitoring of the entire sleep process. Therefore, the present invention effectively overcomes the various shortcomings of the prior art and has a high industrial utilization value.

[0189] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical concept disclosed by the present invention shall still be covered by the claims of the present invention.

Claims

1. A method for collecting and analyzing brain waves for sleep staging, characterized in that: The steps include: S1, collecting brain wave data, the collected initial data is 512Hz, and the initial data is sequentially subjected to downsampling processing, scaling processing, Gaussian filtering processing, Gaussian smoothing processing, and data overlap filling processing; S2, labeling the processed EEG data; S3, collecting human sleep images and combining brain wave data to determine the time when the human falls asleep; S4, obtaining the time point of deep sleep of the human body; S5 builds a sleep neural network to determine the sleep state of a human body, wherein the construction of the sleep neural network includes convolution operation, normalization, activation function, pooling, regularization, and feature splicing.

2. The method for collecting and analyzing brain waves for sleep staging according to claim 1, characterized in that: In step S1, the formula used in the downsampling process is: x 100Hz [n]=x[m·r] in x is the original data, x 100Hz This is the data after downsampling the EEG data from the original 512Hz to 100Hz.

3. The method for collecting and analyzing brain waves for sleep staging according to claim 2, characterized in that: In step S1, the formula used in the scaling process is: y is the scaled data.

4. The method for collecting and analyzing brain waves for sleep staging according to claim 3, characterized in that: In step S1, the formula used for the Gaussian filter is: f min =0.5Hz,f max =100Hz m={y i |and i ≥f min ∧y i ≤f max ,i∈[1,N]} z={y i |m i =1} Where m is a Boolean array identifying data points within the frequency range; y i It is the data for judging the completion of scaling; f min is the minimum frequency threshold when extracting data; f max The maximum frequency threshold when extracting data; z is the valid data that meets the screening conditions.

5. The method for collecting and analyzing brain waves for sleep staging according to claim 4, characterized in that: In step S1, the formula for Gaussian smoothing is: k is the size of the convolution kernel; σ is the standard deviation of the convolution kernel; K is a Gaussian kernel generated according to the specified convolution kernel size and standard deviation of the convolution kernel; i represents the position in the convolution kernel; F is the data after applying Gaussian filtering, which represents the result after smoothing.

6. The method for collecting and analyzing brain waves for sleep staging according to claim 5, characterized in that: In step S1, the data overlapping filling process includes: dividing every 3000 data into one block, and then overlapping the blocks to achieve data expansion; N is the total length of F after data processing; B is block data; O is the overlapping data M is the number of blocks; (1) Calculate the number of blocks M: (2) Calculate the starting S for each block i, where i = 0, 1, ..., M-1 i and end index E i : S i =i·(B-O),E i =S i +B (3) Merge all blocks into one expanded data E=min(B+S i ,N) (4) Merge data d i is the data of the i-th block, and X is the merged data.

7. The method for collecting and analyzing brain waves for sleep staging according to claim 1, characterized in that: In step S2, the labeling operation of the processed EEG data includes: Define a label for every 3,000 data points and calculate the number of labels n required for a piece of data; Where N is the total amount of data.

8. The method for collecting and analyzing brain waves for sleep staging according to claim 7, characterized in that: In step S3, a video is recorded while the person is sleeping, and the state of muscle relaxation and heart rate change of the person are observed by combining the video and the collected brain wave data to determine the time when the person falls asleep; Define the time to fall asleep as t1 seconds, and The label is recorded as sober; the rest This label is temporarily recorded as light sleep.

9. The method for collecting and analyzing brain waves for sleep staging according to claim 8, characterized in that: In step S4, each 30 seconds of data is analyzed to find the point with higher average spectrum amplitude as the time point of deep sleep; (1) Calculate FFT: (2) Calculation frequency: (3) Calculate the amplitude spectrum: m=|F| (4) Perform frequency range screening, i.e., screening the range of delta waves (0.5 Hz to 4 Hz): M=(f≥0.5)∧(f≤4) Where N is the data length, r is the sampling rate, A is the data sample, F is the fft result, f is the frequency, m is the amplitude spectrum, and m a is the average amplitude value, M is the frequency mask; by comparing and observing the spectrum amplitude, the higher point is recorded as the time point of deep sleep.

10. The method for collecting and analyzing brain waves for sleep staging according to claim 9, characterized in that: In step S5, (1) the convolution operation includes the following: The input signal is recorded as x, the convolution kernel is recorded as w, and the formula used for the convolution operation is: Where K is the size of the convolution kernel, y[n] is the output of the convolution, n represents the position of the output signal, and m represents the index of the current element used by the convolution kernel when performing calculations; (2) Let the input signal be x and the output signal be The formula used for the normalization is: The final output normalization result is: Where ε is a small positive number, γ and β are learnable parameters; (3) The activation function adopts the formula: f(x)=max(0,x) The activation function maintains positive values ​​unchanged and changes negative values ​​to 0, which is used to introduce nonlinear factors to avoid overfitting of the model: (4) The formula used for the pooling is: Where x is the input signal and k is the pooling window size parameter; (5) The regularization randomly discards neurons with probability p, A is the discarded neuron, and the formula used is: (6) The feature concatenation adopts a fully connected layer, and the matrix formula is input into the fully connected layer: y=Wx+b In the formula, x is the input vector, W is the weight matrix, and b is the bias; the value of the matrix y is: y = [abc], where a, b, and c represent the scores of the three categories respectively, then the following formula is obtained: Among them, 0 represents wakefulness, 1 represents light sleep, and 2 represents deep sleep. The y matrix is ​​used to judge the human body's sleep state.

Images