A data analysis and early warning method for human health monitoring

Abstract

The application discloses a data analysis and early warning method for human health monitoring, and belongs to the technical field of electroencephalogram monitoring. First, the application processes the electroencephalogram time domain signals in the sleep and wake states by using a sliding window, obtains frequency domain signals through Fourier transform, and divides the frequency domain signals into three frequency band signals. Then, the deviations of the three frequency band signals in the frequency domain gravity and energy in the two states are calculated respectively to obtain corresponding frequency difference sequences and energy difference sequences. Subsequently, features are extracted from the sequences and a feature matrix is constructed, and the feature matrix is processed through a multi-channel multi-scale neural network to obtain an electroencephalogram abnormality evaluation value. When the evaluation value exceeds a threshold value, the system will trigger a risk warning, so that accurate evaluation and timely warning of electroencephalogram abnormalities can be realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electroencephalogram monitoring, and particularly relates to a data analysis and early warning method for human health monitoring. BACKGROUND

[0002] As one of the important physiological signals, electroencephalogram (EEG) is widely used in human health monitoring, especially for the early diagnosis and risk warning of neurological diseases and sleep disorders. Through the analysis of electroencephalogram, the neural activity of individuals can be effectively monitored, and potential health risks such as seizures, anxiety, depression, etc. can be identified. Therefore, the development of efficient electroencephalogram analysis technology will greatly improve the ability of health monitoring and early intervention.

[0003] Existing electroencephalogram analysis technology has many limitations. Most methods only analyze electroencephalogram under a single state (such as wakefulness or sleep), ignoring the relevance and differences of electroencephalogram features under different states, making it difficult to fully reflect the dynamic changes of brain function and fully exploit the complex health information contained in electroencephalogram, resulting in low accuracy of electroencephalogram abnormality evaluation. SUMMARY

[0004] In view of the above problems in the prior art, the present application provides a data analysis and early warning method for human health monitoring, which solves the problem of low accuracy of electroencephalogram abnormality evaluation in the prior art.

[0005] In order to achieve the above-mentioned application purpose, the technical scheme adopted by the present application is as follows: a data analysis and early warning method for human health monitoring, comprising the following steps:

[0006] Sliding windows are used to slide on the electroencephalogram time domain signals in sleep state and wakefulness state respectively, and after each sliding, the signal segment under the sliding window is subjected to Fourier transform to obtain electroencephalogram frequency domain signals;

[0007] The electroencephalogram frequency domain signals corresponding to each sliding window are divided into first frequency band signals, second frequency band signals and third frequency band signals;

[0008] According to the deviation of the two states in the frequency domain centers of the first frequency band signals, the second frequency band signals and the third frequency band signals, first frequency difference sequences, second frequency difference sequences and third frequency difference sequences are obtained;

[0009] According to the deviation of the two states in the energy of the first frequency band signals, the second frequency band signals and the third frequency band signals, first energy difference sequences, second energy difference sequences and third energy difference sequences are obtained;

[0010] Frequency difference features are extracted from each frequency difference sequence to construct three frequency difference feature matrices, and energy difference features are extracted from each energy difference sequence to construct three energy difference feature matrices;

[0011] The three frequency difference feature matrices and the three energy difference feature matrices are processed by a multi-channel multi-scale neural network to obtain an electroencephalogram abnormality evaluation value, and when the electroencephalogram abnormality evaluation value exceeds a threshold value, a risk warning is performed.

[0012] Further, the frequency range of the first frequency band signal is 0-10 Hz, the frequency range of the second frequency band signal is 10-20 Hz, and the frequency range of the third frequency band signal is 20-40 Hz.

[0013] Further, the process of obtaining the first frequency difference sequence, the second frequency difference sequence and the third frequency difference sequence comprises:

[0014] The spectral frequency center of the first frequency band signal in the wake state is subtracted from the spectral frequency center of the first frequency band signal in the sleep state, and the spectral frequency center difference is normalized to obtain a first frequency difference coefficient;

[0015] The spectral frequency center of the second frequency band signal in the wake state is subtracted from the spectral frequency center of the second frequency band signal in the sleep state, and the spectral frequency center difference is normalized to obtain a second frequency difference coefficient;

[0016] The spectral frequency center of the third frequency band signal in the wake state is subtracted from the spectral frequency center of the third frequency band signal in the sleep state, and the spectral frequency center difference is normalized to obtain a third frequency difference coefficient;

[0017] The first frequency difference coefficients are arranged in chronological order to obtain a first frequency difference sequence;

[0018] The second frequency difference coefficients are arranged in chronological order to obtain a second frequency difference sequence;

[0019] The third frequency difference coefficients are arranged in chronological order to obtain a third frequency difference sequence.

[0020] Further, the process of obtaining the first energy difference sequence, the second energy difference sequence and the third energy difference sequence comprises:

[0021] The energy of the first frequency band signal in the wake state is subtracted from the energy of the first frequency band signal in the sleep state, and the energy difference is normalized to obtain a first energy difference coefficient;

[0022] The energy of the second frequency band signal in the wake state is subtracted from the energy of the second frequency band signal in the sleep state, and the energy difference is normalized to obtain a second energy difference coefficient;

[0023] The energy of the third frequency band signal in the wake state is subtracted from the energy of the third frequency band signal in the sleep state, and the energy difference is normalized to obtain a third energy difference coefficient;

[0024] arranging each first energy difference coefficient in time sequence to obtain a first energy difference sequence;

[0025] arranging each second energy difference coefficient in time sequence to obtain a second energy difference sequence;

[0026] arranging each third energy difference coefficient in time sequence to obtain a third energy difference sequence.

[0027] Further, the process of constructing the three frequency difference feature matrices and the three energy difference feature matrices comprises:

[0028] dividing the frequency difference sequence and the energy difference sequence into N parts to obtain a plurality of frequency difference subsequences and a plurality of energy difference subsequences, wherein N is a positive integer;

[0029] extracting a frequency difference mean value, a frequency difference fluctuation value and a frequency difference discrete span for the frequency difference subsequence;

[0030] extracting an energy difference mean value, an energy difference fluctuation value and an energy difference discrete span for the energy difference subsequence;

[0031] constructing an N*3 frequency difference feature matrix by taking the frequency difference mean value, the frequency difference fluctuation value and the frequency difference discrete span of each frequency difference subsequence of the same frequency difference sequence as elements;

[0032] constructing an N*3 energy difference feature matrix by taking the energy difference mean value, the energy difference fluctuation value and the energy difference discrete span of each energy difference subsequence of the same energy difference sequence as elements.

[0033] Further, the frequency difference mean value is the mean value of each frequency difference coefficient in the frequency difference subsequence, the frequency difference fluctuation value is the variance of each frequency difference coefficient in the frequency difference subsequence, and the frequency difference discrete span is the difference between the maximum frequency difference coefficient and the minimum frequency difference coefficient in the frequency difference subsequence;

[0034] the energy difference mean value is the mean value of each energy difference coefficient in the energy difference subsequence, the energy difference fluctuation value is the variance of each energy difference coefficient in the energy difference subsequence, and the energy difference discrete span is the difference between the maximum energy difference coefficient and the minimum energy difference coefficient in the energy difference subsequence.

[0035] Further, the multi-channel multi-scale neural network comprises a first feature fusion enhancement channel, a second feature fusion enhancement channel, a third feature fusion enhancement channel, a multi-scale feature extraction unit, a splicing layer and a full connection layer;

[0036] The first input end of the first feature fusion enhancement channel is configured to input the first frequency difference feature matrix, and the second input end thereof is configured to input the first energy difference feature matrix; the first input end of the second feature fusion enhancement channel is configured to input the second frequency difference feature matrix, and the second input end thereof is configured to input the second energy difference feature matrix; and the first input end of the third feature fusion enhancement channel is configured to input the third frequency difference feature matrix, and the second input end thereof is configured to input the third energy difference feature matrix.

[0037] The first input end of the multi-scale feature extraction unit is connected with the output end of the first feature fusion enhancement channel, the second input end thereof is connected with the output end of the second feature fusion enhancement channel, and the third input end thereof is connected with the output end of the third feature fusion enhancement channel; the input end of the concatenation layer is connected with the first output end and the second output end of the multi-scale feature extraction unit respectively, and the output end thereof is connected with the input end of the fully connected layer; and the output end of the fully connected layer is the output end of the multi-channel multi-scale neural network.

[0038] Further, the first feature fusion enhancement channel, the second feature fusion enhancement channel and the third feature fusion enhancement channel each comprise a feature fusion module and a feature enhancement module.

[0039] The feature fusion module is configured to extract features from the frequency difference feature matrix and the energy difference feature matrix, and perform feature fusion to obtain fused features; and the feature enhancement module is configured to perform feature enhancement on the fused features to obtain enhanced features.

[0040] Further, the feature fusion module comprises a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer and an adder A1.

[0041] The input end of the first convolutional layer is the first input end of the feature fusion module, and the output end thereof is connected with the input end of the second convolutional layer; the input end of the third convolutional layer is the second input end of the feature fusion module, and the output end thereof is connected with the input end of the fourth convolutional layer; the input ends of the adder A1 are respectively connected with the output end of the second convolutional layer and the output end of the fourth convolutional layer, and the output end thereof is the output end of the feature fusion module.

[0042] Further, the multi-scale feature extraction unit comprises a fifth convolutional layer, a sixth convolutional layer, a seventh convolutional layer, an eighth convolutional layer, a ninth convolutional layer, a tenth convolutional layer, an adder A2 and an adder A3.

[0043] The input end of the fifth convolutional layer is connected with the input end of the sixth convolutional layer, and is the first input end of the multi-scale feature extraction unit; the input end of the seventh convolutional layer is connected with the input end of the eighth convolutional layer, and is the second input end of the multi-scale feature extraction unit; and the input end of the ninth convolutional layer is connected with the input end of the tenth convolutional layer, and is the third input end of the multi-scale feature extraction unit.

[0044] The input end of the adder A2 is connected with the output end of the fifth convolutional layer, the output end of the seventh convolutional layer and the output end of the ninth convolutional layer respectively, and the output end thereof is as a first output end of the multi-scale feature extraction unit;

[0045] The input end of the adder A3 is connected with the output end of the sixth convolutional layer, the output end of the eighth convolutional layer and the output end of the tenth convolutional layer respectively, and the output end thereof is as a second output end of the multi-scale feature extraction unit.

[0046] The present application has the following beneficial effects:

[0047] 1、The present application simultaneously focuses on the electroencephalogram signals in the sleep and wake states, calculates the frequency domain barycenter deviation and energy deviation of different states in each frequency band, fully utilizes the correlation and difference of the electroencephalogram features in the two states, breaks through the limitation of traditional single state analysis, can more comprehensively reflect the dynamic changes of the brain neural activities, and thus effectively improves the accuracy of electroencephalogram abnormality recognition.

[0048] 2、The present application performs Fourier transform on the electroencephalogram time domain signals in each sliding window to obtain electroencephalogram frequency domain signals, divides the electroencephalogram frequency domain signals into three segments, analyzes the frequency domain barycenter deviation in different states, can reflect the activity degree and potential functional disorder of the brain in each frequency band, and the energy change reflects the dynamic activity change of the nerve, and comprehensively reflects the dynamic change of the brain function.

[0049] 3、The present application extracts features from the frequency difference sub-sequences and the energy difference sub-sequences corresponding to the three frequency bands, constructs three frequency difference feature matrices and three energy difference feature matrices, can accurately capture the specificity difference of each frequency band in the sleep and wake states. The deviation of the electroencephalogram signals in different frequency bands in the frequency domain barycenter and the energy reflects the abnormal degree of the neural activities corresponding to the frequency band, and separate processing can avoid the mutual interference of the features of different frequency bands, and ensures that the feature information of each frequency band can be fully extracted.

[0050] 4、The present application adopts a multi-channel multi-scale neural network to process the three frequency difference feature matrices and the three energy difference feature matrices, fully excavates the deep association information in the feature matrices, combines the frequency difference features and the energy difference features to predict the electroencephalogram abnormality evaluation value, and improves the evaluation accuracy of the electroencephalogram abnormality evaluation value. BRIEF DESCRIPTION OF DRAWINGS

[0051] Figure 1 It is a flow chart of a human health monitoring data analysis and early warning method;

[0052] Figure 2 It is a structural schematic diagram of a multi-channel multi-scale neural network;

[0053] Figure 3A structural schematic diagram of a first feature fusion enhancement channel, a second feature fusion enhancement channel and a third feature fusion enhancement channel;

[0054] Figure 4 A structural schematic diagram of a feature fusion module;

[0055] Figure 5 A structural schematic diagram of a multi-scale feature extraction unit;

[0056] Figure 6 A structural schematic diagram of a feature enhancement module. DETAILED DESCRIPTION

[0057] The specific embodiments of the present application are described below to facilitate the understanding of the present application for those skilled in the art, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all the inventions utilizing the concept of the present application are within the scope of protection.

[0058] As shown in Figure 1 A data analysis and early warning method for human health monitoring, comprising the following steps:

[0059] Sliding windows are used to slide on the sleep state and the wake state electroencephalogram time domain signals respectively, and after each sliding, Fourier transform is performed on the signal segment under the sliding window to obtain the electroencephalogram frequency domain signal;

[0060] The electroencephalogram frequency domain signal corresponding to each sliding window is divided into a first frequency band signal, a second frequency band signal and a third frequency band signal;

[0061] According to the deviation of the two states in the frequency domain centers of the first frequency band signal, the second frequency band signal and the third frequency band signal, a first frequency difference sequence, a second frequency difference sequence and a third frequency difference sequence are obtained;

[0062] According to the deviation of the two states in the energy of the first frequency band signal, the second frequency band signal and the third frequency band signal, a first energy difference sequence, a second energy difference sequence and a third energy difference sequence are obtained;

[0063] Frequency difference features are extracted from each frequency difference sequence respectively to construct three frequency difference feature matrices, and energy difference features are extracted from each energy difference sequence respectively to construct three energy difference feature matrices;

[0064] A multi-channel multi-scale neural network is used to process the three frequency difference feature matrices and the three energy difference feature matrices to obtain an electroencephalogram abnormality evaluation value, and when the electroencephalogram abnormality evaluation value exceeds a threshold value, a risk warning is performed.

[0065] In the embodiment, the window length is 1 second, the sampling rate is 256 Hz, the number of window points is 256, the sliding step is 0.25 second, the window overlap rate is 75%, and the signal length of the brain electrical time domain signals in the sleep state and the wakeful state is 30 seconds.

[0066] In the embodiment, the frequency range of the first frequency band signal is 0-10 Hz, the frequency range of the second frequency band signal is 10-20 Hz, and the frequency range of the third frequency band signal is 20-40 Hz. The brain electrical frequency domain signal belonging to the frequency range of 0-10 Hz is classified as the first frequency band signal, the brain electrical frequency domain signal belonging to the frequency range of 10-20 Hz is classified as the second frequency band signal, and the brain electrical frequency domain signal belonging to the frequency range of 20-40 Hz is classified as the third frequency band signal.

[0067] In the sleep state, the brain activity is weakened, the 0-10 Hz frequency band signal is dominant, and the signal is characterized by large amplitude and slow frequency; and in the wakeful state, the signal in the frequency band changes with the relaxation / concentration state.

[0068] In the sleep-to-wake transition, the 10-20 Hz signal gradually increases; in the wakeful state, the frequency band is related to attention and mild thinking activity; and in the sleep state, the signal in the frequency band is significantly weakened or even suppressed by the low-frequency signal.

[0069] In the wakeful state, especially when complex thinking or active activity is performed, the 20-40 Hz high-frequency signal is significantly enhanced; and in the sleep state (especially in the non-rapid eye movement sleep), the signal in the frequency band is almost suppressed, and only slightly rises in the rapid eye movement sleep (quasi-wakeful state), but the intensity is much lower than that in the wakeful state.

[0070] In the embodiment, the process of obtaining the first frequency difference sequence, the second frequency difference sequence and the third frequency difference sequence comprises:

[0071] The spectral frequency centroid of the first frequency band signal in the wakeful state is subtracted from the spectral frequency centroid of the first frequency band signal in the sleep state, and the spectral frequency centroid difference is normalized to obtain the first frequency difference coefficient;

[0072] The spectral frequency centroid of the second frequency band signal in the wakeful state is subtracted from the spectral frequency centroid of the second frequency band signal in the sleep state, and the spectral frequency centroid difference is normalized to obtain the second frequency difference coefficient;

[0073] The spectral frequency centroid of the third frequency band signal in the wakeful state is subtracted from the spectral frequency centroid of the third frequency band signal in the sleep state, and the spectral frequency centroid difference is normalized to obtain the third frequency difference coefficient;

[0074] The first frequency difference coefficients are arranged in chronological order to obtain the first frequency difference sequence;

[0075] arranging each second frequency difference coefficient in time sequence to obtain a second frequency difference sequence;

[0076] arranging each third frequency difference coefficient in time sequence to obtain a third frequency difference sequence.

[0077] The spectral frequency center (a frequency point where spectral energy is concentrated) is a core index reflecting the dominant frequency of a signal in a frequency band: in sleep and wake states, the dominant frequency of the same frequency band will have characteristic shifts (for example, the 0~10Hz frequency band is biased to low-frequency delta waves in sleep and to high-frequency alpha waves in wake).

[0078] The three frequency bands of the present application correspond to the frequency characteristics of different activity intensities of the brain (low frequency reflects the basic state, medium and high frequency reflect the transition state, and high frequency reflects the active state). The present application can accurately quantify the differences in the spectral frequency center of the signals in the three frequency bands in sleep and wake states by obtaining the first, second and third frequency difference sequences through the above process.

[0079] In the present embodiment, the calculation formula of the spectral frequency center is: wherein Z is the spectral frequency center, f k is the kth frequency value in a frequency band signal, y k is the amplitude corresponding to the kth frequency value in a frequency band signal, and K is the number of frequency values in a frequency band signal.

[0080] In the present embodiment, the formula for normalizing the spectral frequency center difference is: wherein μ z is the frequency difference coefficient, Z d is the spectral frequency center difference, Z max is the maximum spectral frequency center difference.

[0081] In the present embodiment, the process of obtaining the first, second and third energy difference sequences includes:

[0082] Subtracting the energy of the first frequency band signal in the sleep state from the energy of the first frequency band signal in the wake state, and normalizing the energy difference to obtain the first energy difference coefficient;

[0083] Subtracting the energy of the second frequency band signal in the sleep state from the energy of the second frequency band signal in the wake state, and normalizing the energy difference to obtain the second energy difference coefficient;

[0084] Subtracting the energy of the third frequency band signal in the sleep state from the energy of the third frequency band signal in the wake state, and normalizing the energy difference to obtain the third energy difference coefficient;

[0085] arranging each first energy difference coefficient in time sequence to obtain a first energy difference sequence;

[0086] arranging each second energy difference coefficient in time sequence to obtain a second energy difference sequence;

[0087] arranging each third energy difference coefficient in time sequence to obtain a third energy difference sequence.

[0088] The first, second and third energy difference sequences can accurately capture the energy differences of the three frequency band signals in the sleep and wake states, and the energy difference coefficients arranged in time sequence can record the dynamic evolution information of the energy differences in the state conversion process, and the three energy difference sequences can comprehensively present the energy distribution differences of the sleep and wake states from the low, medium and high frequency spectrum dimensions.

[0089] In the embodiment, the calculation formula of the energy of the frequency band signal is: wherein E is the energy of the frequency band signal.

[0090] In the embodiment, the formula for normalizing the energy difference is: wherein μ E is the energy difference coefficient, E d is the energy difference, and E max is the maximum energy difference.

[0091] In the embodiment, the process of constructing the three frequency difference feature matrices and the three energy difference feature matrices includes:

[0092] The frequency difference sequence and the energy difference sequence are divided into N parts to obtain a plurality of frequency difference subsequences and a plurality of energy difference subsequences, wherein N is a positive integer;

[0093] The frequency difference mean, the frequency difference fluctuation value and the frequency difference discrete span are extracted from the frequency difference subsequence;

[0094] The energy difference mean, the energy difference fluctuation value and the energy difference discrete span are extracted from the energy difference subsequence;

[0095] The frequency difference mean, the frequency difference fluctuation value and the frequency difference discrete span of each frequency difference subsequence of the same frequency difference sequence are taken as elements to construct a frequency difference feature matrix of N*3;

[0096] The energy difference mean, the energy difference fluctuation value and the energy difference discrete span of each energy difference subsequence of the same energy difference sequence are taken as elements to construct an energy difference feature matrix of N*3.

[0097] The frequency difference feature matrix obtained from the first frequency difference sequence is a first frequency difference feature matrix, the frequency difference feature matrix obtained from the second frequency difference sequence is a second frequency difference feature matrix, and the frequency difference feature matrix obtained from the third frequency difference sequence is a third frequency difference feature matrix.

[0098] The energy difference feature matrix obtained from the first energy difference sequence is a first energy difference feature matrix, the energy difference feature matrix obtained from the second energy difference sequence is a second energy difference feature matrix, and the energy difference feature matrix obtained from the third energy difference sequence is a third energy difference feature matrix.

[0099] In the embodiment, the frequency difference mean value is the mean value of each frequency difference coefficient in the frequency difference subsequence, the frequency difference fluctuation value is the variance of each frequency difference coefficient in the frequency difference subsequence, and the frequency difference discrete span is the difference between the maximum frequency difference coefficient and the minimum frequency difference coefficient in the frequency difference subsequence.

[0100] The energy difference mean value is the mean value of each energy difference coefficient in the energy difference subsequence, the energy difference fluctuation value is the variance of each energy difference coefficient in the energy difference subsequence, and the energy difference discrete span is the difference between the maximum energy difference coefficient and the minimum energy difference coefficient in the energy difference subsequence.

[0101] The present application divides the frequency difference sequence and the energy difference sequence into N subsequences, and then extracts the mean value, the fluctuation value and the discrete span of each subsequence. The mean value reflects the overall level of the features in the subsequence, the fluctuation value reflects the stability of the features, and the discrete span reflects the distribution range of the features.

[0102] In the embodiment, N is a positive integer greater than or equal to 5. A frequency difference subsequence corresponds to a frequency difference mean value, a frequency difference fluctuation value and a frequency difference discrete span as elements of each row, obtaining a frequency difference feature matrix of N*3; the energy difference mean value, the energy difference fluctuation value and the energy difference discrete span of an energy difference subsequence as elements of each row, obtaining an energy difference feature matrix of N*3.

[0103] In the embodiment, as shown in Figure 2 The multi-channel multi-scale neural network comprises a first feature fusion enhancement channel, a second feature fusion enhancement channel, a third feature fusion enhancement channel, a multi-scale feature extraction unit, a splicing layer and a full connection layer.

[0104] The first input end of the first feature fusion enhancement channel is used for inputting the first frequency difference feature matrix, and the second input end thereof is used for inputting the first energy difference feature matrix; the first input end of the second feature fusion enhancement channel is used for inputting the second frequency difference feature matrix, and the second input end thereof is used for inputting the second energy difference feature matrix; the first input end of the third feature fusion enhancement channel is used for inputting the third frequency difference feature matrix, and the second input end thereof is used for inputting the third energy difference feature matrix.

[0105] The first input end of the multi-scale feature extraction unit is connected with the output end of the first feature fusion enhancement channel, the second input end is connected with the output end of the second feature fusion enhancement channel, and the third input end is connected with the output end of the third feature fusion enhancement channel; the input end of the splicing layer is connected with the first output end and the second output end of the multi-scale feature extraction unit respectively, and the output end is connected with the input end of the full connection layer; and the output end of the full connection layer is the output end of the multi-channel multi-scale neural network.

[0106] The application adopts each feature fusion enhancement channel to process a pair of frequency difference feature matrices and energy difference feature matrices, the frequency difference focuses on the time sequence change of the frequency gravity center difference, the energy difference reflects the dynamic law of the energy difference, and the fusion of the two can make the channel capture the information in the frequency and energy dimensions at the same time, and then the multi-scale feature extraction unit extracts the multi-scale features from the fused enhanced features, thereby improving the classification accuracy of the full connection layer.

[0107] As shown in Figure 3 The first feature fusion enhancement channel, the second feature fusion enhancement channel and the third feature fusion enhancement channel each include a feature fusion module and a feature enhancement module.

[0108] The feature fusion module is used for extracting features from the frequency difference feature matrix and the energy difference feature matrix and performing feature fusion to obtain fused features; and the feature enhancement module is used for performing feature enhancement on the fused features to obtain enhanced features.

[0109] As shown in Figure 4 The feature fusion module includes a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer and an adder A1.

[0110] The input end of the first convolutional layer is the first input end of the feature fusion module, and the output end is connected with the input end of the second convolutional layer; the input end of the third convolutional layer is the second input end of the feature fusion module, and the output end is connected with the input end of the fourth convolutional layer; the input ends of the adder A1 are respectively connected with the output end of the second convolutional layer and the output end of the fourth convolutional layer, and the output end is the output end of the feature fusion module.

[0111] In the embodiment, the convolution kernel size of the first convolutional layer and the third convolutional layer is 1x1, and the convolution kernel size of the second convolutional layer and the fourth convolutional layer is 3x3.

[0112] The application extracts features from the frequency difference feature matrix through the first convolutional layer and the second convolutional layer, extracts features from the energy difference feature matrix through the third convolutional layer and the fourth convolutional layer, and then performs bit-by-bit addition through the adder A1 to obtain the fused features.

[0113] As shown in Figure 5As shown, the multi-scale feature extraction unit includes: a fifth convolutional layer, a sixth convolutional layer, a seventh convolutional layer, an eighth convolutional layer, a ninth convolutional layer, a tenth convolutional layer, adder A2, and adder A3;

[0114] The input of the fifth convolutional layer is connected to the input of the sixth convolutional layer and serves as the first input of the multi-scale feature extraction unit; the input of the seventh convolutional layer is connected to the input of the eighth convolutional layer and serves as the second input of the multi-scale feature extraction unit; the input of the ninth convolutional layer is connected to the input of the tenth convolutional layer and serves as the third input of the multi-scale feature extraction unit.

[0115] The input of adder A2 is connected to the output of the fifth convolutional layer, the output of the seventh convolutional layer, and the output of the ninth convolutional layer, respectively, and its output serves as the first output of the multi-scale feature extraction unit.

[0116] The input of adder A3 is connected to the outputs of the sixth, eighth and tenth convolutional layers, respectively, and its output serves as the second output of the multi-scale feature extraction unit.

[0117] In this embodiment, the kernel size of the fifth, seventh, and ninth convolutional layers is 1×1, and the kernel size of the sixth, eighth, and tenth convolutional layers is 1×3.

[0118] This invention extracts features at different scales using convolutional kernels of different sizes, and adds features at the same scale bit by bit using an adder to achieve the fusion of features at the same scale. At the splicing layer, the two fused scale features are spliced ​​together to improve the classification accuracy of the fully connected layer.

[0119] like Figure 6 As shown, the feature enhancement module includes: a max pooling layer, an average pooling layer, an eleventh convolutional layer, a twelfth convolutional layer, an adder A4, a sigmoid layer, and a multiplier M1;

[0120] The input of the max pooling layer is connected to the input of the average pooling layer and the first input of the multiplier M1, respectively, and serves as the input of the feature enhancement module.

[0121] The input of the eleventh convolutional layer is connected to the output of the max pooling layer; the input of the twelfth convolutional layer is connected to the output of the average pooling layer; the input of adder A4 is connected to the outputs of the eleventh and twelfth convolutional layers respectively, and its output is connected to the input of the sigmoid layer; the output of the sigmoid layer is connected to the second input of multiplier M1; the output of multiplier M1 serves as the output of the feature enhancement module.

[0122] In the embodiment, the convolution kernel size of the eleventh convolution layer and the twelfth convolution layer is 1*1.

[0123] The maximum pooling layer focuses on significant features, and retains key abnormal or state mutation information; the average pooling layer captures global trends, and both are processed in parallel, so that the network simultaneously obtains "local prominent features" and "mean features", the "local prominent features" and "mean features" are fused in the adder A4 after being processed by the eleventh convolution layer and the twelfth convolution layer, and attention is generated through the Sigmoid layer to enhance the output features of the feature fusion module and adaptively focus on important features.

[0124] In the embodiment, the feature enhancement module can also be replaced by an existing channel attention module.

[0125] In the embodiment, the range of the electroencephalogram abnormality evaluation value can be set to 0-10 points, (0 points for completely normal and 10 points for extreme abnormality), and the threshold is set to 5 points (i.e., more than 5 points triggers an early warning). For example, 1.5 points and 2.3 points indicate that the electroencephalogram rhythm is normal and there is no abnormal risk.

[0126] The application simultaneously focuses on the electroencephalogram signals in the sleep and wake states, calculates the frequency domain barycenter deviation and energy deviation in each frequency band in different states, fully utilizes the correlation and difference of the electroencephalogram features in the two states, breaks through the limitation of traditional single-state analysis, and can more comprehensively reflect the dynamic changes of the brain neural activities, thereby effectively improving the accuracy of electroencephalogram abnormality recognition.

[0127] The application performs Fourier transform on the electroencephalogram time domain signals in each sliding window to obtain electroencephalogram frequency domain signals, divides the electroencephalogram frequency domain signals into three segments, and can reflect the activity degree and potential functional disorder of the brain in each frequency band by analyzing the frequency domain barycenter deviation in different states, and the change of energy reflects the dynamic activity change of the nerve, and comprehensively reflects the dynamic change of the brain function.

[0128] The application extracts features from the frequency difference sub-sequences and energy difference sub-sequences corresponding to the three frequency bands, constructs three frequency difference feature matrices and three energy difference feature matrices, and can accurately capture the specificity difference of each frequency band in the sleep and wake states. The deviation of the electroencephalogram signals in different frequency bands in the frequency domain barycenter and energy reflects the abnormal degree of the neural activities corresponding to the frequency band, and separate processing can avoid the mutual interference of the features of different frequency bands, and ensure that the feature information of each frequency band can be fully extracted.

[0129] The application adopts a multi-channel multi-scale neural network to process the three frequency difference feature matrices and the three energy difference feature matrices, fully excavates the deep correlation information in the feature matrices, combines the frequency difference features and the energy difference features to predict the electroencephalogram abnormality evaluation value, and improves the evaluation accuracy of the electroencephalogram abnormality evaluation value.

[0130] The above merely provides the preferred embodiments of the present application, and is not intended to limit the present application. The present application can have various modifications and changes for those skilled in the art. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall fall into the scope of protection of the present application.

Claims

1. A data analysis and early warning method for human health monitoring, characterized in that, Includes the following steps: A sliding window was used to slide across the EEG time-domain signals in both sleep and wake states. After each slide, a Fourier transform was performed on the signal segment under the sliding window to obtain the EEG frequency-domain signal. The EEG frequency domain signal corresponding to each sliding window is divided into the first frequency band signal, the second frequency band signal, and the third frequency band signal; Based on the deviations of the two states at the centroids of the frequency domain of the first, second, and third frequency band signals, the first frequency difference sequence, the second frequency difference sequence, and the third frequency difference sequence are obtained. Based on the energy deviations of the first, second, and third frequency band signals in the two states, a first energy difference sequence, a second energy difference sequence, and a third energy difference sequence are obtained. For each frequency difference sequence, extract frequency difference features and construct three frequency difference feature matrices. For each energy difference sequence, extract energy difference features and construct three energy difference feature matrices. A multi-channel, multi-scale neural network is used to process three frequency difference feature matrices and three energy difference feature matrices to obtain an EEG abnormality assessment value. When the EEG abnormality assessment value exceeds the threshold, a risk warning is issued. The process of constructing three frequency difference feature matrices and three energy difference feature matrices includes: Both the frequency difference sequence and the energy difference sequence are divided into N parts to obtain multiple frequency difference subsequences and multiple energy difference subsequences, where N is a positive integer; Extract the mean frequency difference, frequency difference fluctuation value, and frequency difference discrete span from the frequency difference subsequence; Extract the mean energy difference, energy difference fluctuation value, and energy difference discrete span from the energy difference subsequences respectively; The mean frequency difference, frequency difference fluctuation, and frequency difference discrete span of each frequency difference subsequence of the same frequency difference sequence are used as elements to construct an N×3 frequency difference feature matrix; The mean energy difference, energy difference fluctuation value, and energy difference discrete span of each energy difference subsequence of the same energy difference sequence are used as elements to construct an N×3 energy difference feature matrix; The mean frequency difference is the mean of each frequency difference coefficient in the frequency difference subsequence, the frequency difference fluctuation is the variance of each frequency difference coefficient in the frequency difference subsequence, and the frequency difference dispersion span is the difference between the maximum and minimum frequency difference coefficients in the frequency difference subsequence. The mean energy difference is the mean of the energy difference coefficients in the energy difference subsequence, the fluctuation of the energy difference is the variance of the energy difference coefficients in the energy difference subsequence, and the discrete span of the energy difference is the difference between the maximum and minimum energy difference coefficients in the energy difference subsequence.

2. The data analysis and early warning method for human health monitoring according to claim 1, characterized in that, The frequency range of the first frequency band signal is 0~10Hz, the frequency range of the second frequency band signal is 10~20Hz, and the frequency range of the third frequency band signal is 20~40Hz.

3. The data analysis and early warning method for human health monitoring according to claim 1, characterized in that, The process of obtaining the first frequency difference sequence, the second frequency difference sequence, and the third frequency difference sequence includes: The first frequency difference coefficient is obtained by subtracting the spectral frequency centroid of the first frequency band signal in the awake state from the spectral frequency centroid of the first frequency band signal in the sleep state, and then normalizing the difference in spectral frequency centroids. The second frequency difference coefficient is obtained by subtracting the spectral frequency centroid of the second frequency band signal in the sleep state from the spectral frequency centroid of the second frequency band signal in the awake state, and then normalizing the difference in spectral frequency centroids. The third frequency difference coefficient is obtained by subtracting the spectral frequency centroid of the third frequency band signal in the sleep state from the spectral frequency centroid of the third frequency band signal in the awake state, and then normalizing the difference in spectral frequency centroids. Arrange the first frequency difference coefficients in chronological order of their occurrence to obtain the first frequency difference sequence; Arrange the second frequency difference coefficients in chronological order of their occurrence to obtain the second frequency difference sequence; Arrange the third frequency difference coefficients in chronological order to obtain the third frequency difference sequence.

4. The data analysis and early warning method for human health monitoring according to claim 1, characterized in that, The process of obtaining the first energy difference sequence, the second energy difference sequence, and the third energy difference sequence includes: The energy of the first frequency band signal in the waking state is subtracted from the energy of the first frequency band signal in the sleeping state, and the energy difference is normalized to obtain the first energy difference coefficient. The energy of the second frequency band signal in the waking state is subtracted from the energy of the second frequency band signal in the sleeping state, and the energy difference is normalized to obtain the second energy difference coefficient. The energy of the third frequency band signal in the waking state is subtracted from the energy of the third frequency band signal in the sleeping state, and the energy difference is normalized to obtain the third energy difference coefficient. Arrange the first energy difference coefficients in chronological order to obtain the first energy difference sequence; Arrange the second energy difference coefficients in chronological order to obtain the second energy difference sequence; Arrange the third energy difference coefficients in chronological order to obtain the third energy difference sequence.

5. The data analysis and early warning method for human health monitoring according to claim 1, characterized in that, The multi-channel multi-scale neural network includes: a first feature fusion enhancement channel, a second feature fusion enhancement channel, a third feature fusion enhancement channel, a multi-scale feature extraction unit, a splicing layer, and a fully connected layer; The first input terminal of the first feature fusion enhancement channel is used to input the first frequency difference feature matrix, and its second input terminal is used to input the first energy difference feature matrix; the first input terminal of the second feature fusion enhancement channel is used to input the second frequency difference feature matrix, and its second input terminal is used to input the second energy difference feature matrix; the first input terminal of the third feature fusion enhancement channel is used to input the third frequency difference feature matrix, and its second input terminal is used to input the third energy difference feature matrix. The first input of the multi-scale feature extraction unit is connected to the output of the first feature fusion enhancement channel, its second input is connected to the output of the second feature fusion enhancement channel, and its third input is connected to the output of the third feature fusion enhancement channel. The input of the splicing layer is connected to the first and second outputs of the multi-scale feature extraction unit, respectively, and its output is connected to the input of the fully connected layer. The output of the fully connected layer serves as the output of the multi-channel multi-scale neural network.

6. The data analysis and early warning method for human health monitoring according to claim 5, characterized in that, The first feature fusion enhancement channel, the second feature fusion enhancement channel, and the third feature fusion enhancement channel all include: a feature fusion module and a feature enhancement module; The feature fusion module is used to extract features from the frequency difference feature matrix and the energy difference feature matrix, and perform feature fusion to obtain fused features; the feature enhancement module is used to enhance the fused features to obtain enhanced features.

7. The data analysis and early warning method for human health monitoring according to claim 6, characterized in that, The feature fusion module includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, a fourth convolutional layer, and an adder A1; The input of the first convolutional layer serves as the first input of the feature fusion module, and its output is connected to the input of the second convolutional layer. The input of the third convolutional layer serves as the second input of the feature fusion module, and its output is connected to the input of the fourth convolutional layer. The input of adder A1 is connected to the outputs of the second and fourth convolutional layers, respectively, and its output serves as the output of the feature fusion module.

8. The data analysis and early warning method for human health monitoring according to claim 6, characterized in that, The multi-scale feature extraction unit includes: a fifth convolutional layer, a sixth convolutional layer, a seventh convolutional layer, an eighth convolutional layer, a ninth convolutional layer, a tenth convolutional layer, adder A2, and adder A3; The input of the fifth convolutional layer is connected to the input of the sixth convolutional layer and serves as the first input of the multi-scale feature extraction unit; the input of the seventh convolutional layer is connected to the input of the eighth convolutional layer and serves as the second input of the multi-scale feature extraction unit; the input of the ninth convolutional layer is connected to the input of the tenth convolutional layer and serves as the third input of the multi-scale feature extraction unit. The input of adder A2 is connected to the output of the fifth convolutional layer, the output of the seventh convolutional layer, and the output of the ninth convolutional layer, respectively, and its output serves as the first output of the multi-scale feature extraction unit. The input of adder A3 is connected to the outputs of the sixth, eighth and tenth convolutional layers, respectively, and its output serves as the second output of the multi-scale feature extraction unit.

Images