CN112741638A

CN112741638A - Medical diagnosis auxiliary system based on EEG signal

Info

Publication number: CN112741638A
Application number: CN202110010464.0A
Authority: CN
Inventors: 王磊; 梁锦威; 刘洋; 石岩; 张书源
Original assignee: Beihang University
Current assignee: Beihang University
Priority date: 2021-01-05
Filing date: 2021-01-05
Publication date: 2021-05-04
Anticipated expiration: 2041-01-05
Also published as: CN112741638B

Abstract

The invention provides a medical diagnosis auxiliary system based on an EEG signal, and belongs to the field of medical instruments. The system comprises an EEG signal acquisition device for acquiring and transmitting EEG signals of a patient and an upper computer medical diagnosis auxiliary system for EEG signal processing, analysis and mental disease prediction. The invention uses the self-made EEG signal acquisition device to acquire, process and transmit EEG signals of patients, extracts time domain characteristics, frequency domain characteristics and nonlinear characteristics of the EEG signals on an auxiliary system for medical diagnosis of an upper computer, predicts the probability of the patients suffering from mental diseases such as epilepsy, schizophrenia, depression and autism by adopting a trained random forest model based on the characteristics, and provides a definite auxiliary basis for doctors to diagnose the mental diseases.

Description

Medical diagnosis auxiliary system based on EEG signal

Technical Field

The invention relates to the field of EEG signal processing analysis and medical auxiliary instruments, in particular to a medical diagnosis auxiliary system based on an EEG signal.

Background

The brain electrical wave (EEG) is the comprehensive result of postsynaptic potentials of human brain neurons, and is the general reflection of electrophysiological activity of brain neurons on the surface of the cerebral cortex or scalp. In a natural state without external stimulation, spontaneous electroencephalogram signals generated by the human brain are generally regarded as random signals with non-stationarity and outstanding stability. The spontaneous electroencephalogram signals contain a large amount of physiological and disease information, and in the aspect of clinical medicine, the spontaneous electroencephalogram signals can provide diagnosis basis for certain brain diseases. For example, clinical studies show that even in the absence of epileptic seizures, the brain can emit abnormal waves such as spike waves, sharp waves, spike-slow waves and slow-scattered waves; the brain electricity of a schizophrenic patient has abnormal discharge conditions, such as greater or less wave amplitude and longer or shorter wavelength than normal persons; the brain waves of depression patients generally show that basic electrical activity becomes slow, slow activity becomes more, sharp waves are occasionally seen, and the like; the brain electricity of the autistic patient shows non-drug-induced rapid wave increase, focal discharge, rhythm slowing, rhythm imbalance and the like. Therefore, the abnormal brain discharge result is processed and analyzed by monitoring the EEG signal of the patient, auxiliary basis except clinical diagnosis can be provided for doctors, the doctors are helped to quickly and accurately diagnose mental diseases, and the diagnosis accuracy is greatly improved.

Therefore, the invention designs an EEG signal acquisition device from the extraction of EEG signals, constructs a complete set of auxiliary system for medical diagnosis of mental diseases, aims to process and analyze EEG signals of patients, predicts the probability of mental diseases such as epilepsy, schizophrenia, depression, autism and the like of the patients through a random forest model, and provides auxiliary basis for doctors to diagnose the mental diseases.

Disclosure of Invention

The invention aims to provide a medical diagnosis auxiliary system based on an EEG signal, which is used for carrying out diagnosis and analysis on mental health of a patient by acquiring and processing the EEG signal of the patient and helping medical personnel to accurately diagnose whether the patient suffers from mental diseases and what kind of mental diseases the patient may suffer from. The system is designed based on EEG dynamic characteristic signals, can monitor the brain waves of patients, predicts the probability of the patients suffering from mental diseases such as epilepsy, schizophrenia, depression and autism by extracting and classifying the characteristics of the brain waves, and can provide reliable medical diagnosis assistance for doctors.

The invention provides the following technical scheme: a medical diagnosis assistance system based on EEG signals. The invention comprises the processes of collecting, processing, transmitting, analyzing and predicting the mental disease classification of the patient EEG signal. The EEG signal acquisition device realizes the acquisition, preliminary processing and signal transmission of EEG signals and transmits the EEG signals to the upper computer; the upper computer medical diagnosis auxiliary system realizes further processing and analysis of the EEG signals and classification prediction of mental diseases and carries out visual display.

The invention firstly realizes the collection, processing and transmission of the EEG signals of patients based on a self-made EEG signal collecting device, and the specific steps are as follows:

the method comprises the following steps: collecting EEG signals of a patient within a period of time through a self-made EEG signal collecting device;

step two: and the acquired EEG signals are amplified, filtered and the like, converted into digital signals and transmitted to an upper computer medical diagnosis auxiliary system.

And on the upper computer medical diagnosis auxiliary system, the subsequent processing of the EEG signals is realized, including data noise reduction, feature extraction, classifier training, mental disease classification prediction and the like. The method comprises the following specific steps:

the method comprises the following steps: the upper computer medical diagnosis auxiliary system carries out preprocessing such as filtering and denoising on the EEG signal;

step two: extracting time domain features, frequency domain features and nonlinear features of the EEG signal;

step three: training a classifier by using a random forest algorithm to construct a multi-classification system;

step four: based on the extracted features, the trained random forest model is used for carrying out classification and prediction on mental diseases of the patient, and a clear auxiliary basis is provided for diagnosis of doctors.

Preferably, the homemade EEG signal acquisition device has the following characteristics:

(1) the EEG signal acquisition device is designed in a head-wearing manner, is integrally designed on the basis of an integrated chip, and is provided with two acquisition electrodes and two reference electrodes, wherein the two acquisition electrodes are respectively positioned on the left forehead and the right forehead, and the two reference electrodes are respectively positioned near the left ear and the right ear;

(2) the acquisition electrode and the reference electrode can be configured into a plurality of different lead acquisition modes: 1) the reference electrode at one ear corresponds to the forehead collecting electrode at the same side; 2) the reference electrodes at the left ear and the right ear are connected together to be used as reference electrodes and then matched with one/two collecting electrodes; 3) the collecting electrode at one side of the ear corresponds to the collecting electrode at the other side of the forehead;

(3) through different lead modes, a doctor can autonomously select an EEG signal acquisition area and flexibly record absolute values of EEG signal changes of different frontal areas;

(4) the EEG signal acquisition device is integrated with an amplification module and a hardware filtering module, and can perform signal amplification, low-pass filtering, high-pass filtering and other preliminary processing on the EEG signals acquired by the acquisition electrode;

(5) after the EEG signal is subjected to primary processing, the EEG signal is converted into a digital signal through an analog-to-digital converter, and the digital signal is transmitted to an upper computer medical diagnosis auxiliary system through a wireless transmission technology.

Preferably, the specific flow of EEG signal acquisition and transmission is:

the method comprises the following steps: a self-made EEG signal acquisition device is worn by a patient, an electrode lead mode is configured, the sampling rate is set, and EEG signal acquisition is started;

step two: performing primary processing on the EEG signal, including signal amplification, high-pass filtering, low-pass filtering, and the like;

step three: and converting the preprocessed EEG signals into digital signals, and transmitting the digital signals to an upper computer for further processing and analysis through wireless transmission technologies such as Bluetooth and WiFi.

Preferably, the specific process of the upper computer medical diagnosis auxiliary system for EEG signal processing and classifier training and prediction is as follows:

the method comprises the following steps: carrying out software filtering on an EEG signal transmitted by an EEG signal acquisition device to remove noise and 50Hz power frequency interference caused by mismatching of a capacitor and a resistor of a hardware filter and the like, so as to obtain a relatively pure EEG signal;

step two: cutting an EEG signal into fixed time lengths by adopting a sliding window, removing myoelectricity, ocular artifacts and the like by adopting an independent component analysis method, simultaneously processing EEG data by using methods such as statistical analysis, power spectrum estimation, sample entropy estimation and the like, and extracting time domain characteristics, frequency domain characteristics and nonlinear characteristics of the EEG signal;

step three: training a random forest classifier, and constructing an optimal random forest classification model based on an out of bag estimation error (out of bag error);

step four: predicting the EEG signals of the patient in the period of time by using an optimal random forest classification model based on the time domain characteristics, the frequency domain characteristics and the nonlinear characteristics of the EEG signals of the patient extracted in the step two, predicting the probability of the EEG signals of the patient suffering from mental diseases such as epilepsy, schizophrenia, depression and autism, and carrying out visual display;

step five: and the doctor makes accurate judgment according to the prediction result and by combining medical experience.

Preferably, the specific process of training and predicting the random forest classifier in the third step and the fourth step of the upper computer medical diagnosis auxiliary system is as follows:

the method comprises the following steps: collecting a large number of EEG signals of normal people and patients with mental diseases such as epilepsy, schizophrenia, depression, autism and the like;

step two: a professional doctor selects an EEG wave band with obvious characteristics in a period of time, and labels of categories such as '0', '1', '2', '3', '4', … ', J' and the like are respectively marked on EEG signals of normal people, epileptics, schizophrenic patients, depressive illness patients, autistic patients and the like;

step three: performing denoising pretreatment on each EEG signal in the second step, cutting the EEG signal into EEG signal segments with certain time length, extracting time domain characteristics, frequency domain characteristics and nonlinear characteristics of the EEG signal segments, constructing training samples by using the EEG characteristics, and setting the number of samples of a training set to be N and the number of characteristics of each training sample to be M;

preferably, the time domain features extracted for the EEG signal segments in this step include: the amplitude (maximum, minimum, mean, variance, etc.) of the EEG signal, the waveform (kurtosis, skewness, wave width, etc.). When the brain is damaged or has neurological diseases, abnormal brain electrical signals can appear in the brain electrical signals of a human, and the time domain characteristics of the abnormal brain electrical signals are obtained through a statistical analysis method and serve as first characteristic vectors.

Preferably, the frequency domain features extracted for the EEG signal segments in this step are EEG power spectra:

the power spectrum of the EEG signal transforms the time domain EEG signal with the amplitude changing along with the time into an EEG power spectrogram with the amplitude changing along with the frequency, the power spectrum reflects the frequency components of the EEG signal and the relative strength of each component, and the rhythm distribution and the change condition of the EEG signal are revealed from the frequency domain. The object of power spectrum estimation is a finite discrete time signal, and the periodogram method can be used in the invention to estimate the power spectrum of an EEG signal, and the basic idea is as follows: regarding S-point observation data X (n) { X (1), X (2),. ·, X (S) } of an input signal as an energy-limited signal, a frequency-domain signal sequence X is obtained after Fast Fourier Transform (FFT) of X (n) (1, 2.·, S) is directly taken_S(w) then taking X_SThe modulus of (w) is squared and divided by S as an estimate of the true signal power spectrum. With P_PER(w) represents the power spectrum estimated by the periodogram method, and the formula is as follows:

the time-frequency relationship between the EEG rhythm and the EEG power spectrum is established by:

1) the EEG signal is first segmented into 4 frequency bands: delta (1-3Hz), theta (4-7Hz), alpha (8-13Hz), beta (14-30 Hz);

2) respectively extracting power spectrum characteristics of 4 frequency bands;

3) combining the power spectral features of the 4 frequency bands into a power spectral vector that characterizes the relative strengths and weaknesses between the EEG rhythms from an energy perspective;

the power spectral vector of the EEG signal segment is taken as the second feature vector.

Preferably, the non-linear features extracted for the EEG signal segments in this step are sample entropies:

the sample entropy is a method for measuring the complexity of a time sequence, the physical meaning of the sample entropy is to reflect the complexity of the signal by measuring the probability of generating a new pattern in the signal, the smaller the sample entropy is, the higher the self-similarity of the corresponding sample sequence is, and the larger the sample entropy is, the more complex the corresponding sample sequence is. When the brain is abnormal, the electroencephalogram signals become complex, irregular and more random. Assuming that there is a sequence of original EEG signals x (n) { x (1), x (2),.., x(s) }, the sample entropy is calculated as follows:

1) reconstructing the l-dimensional vector sequence: x_l(i) X (i), x (i +1) ·, x (i + l-1) }, where 1 ≦ i ≦ S-l +1, the vector sequence representing l consecutive x values starting from the ith point;

2) definition of X_l(i) And X_l(j) Has a vector distance of d [ X ]_l(i)，X_l(j)]Namely:

wherein i is more than or equal to 1 and less than or equal to S-l +1, j is more than or equal to 1 and less than or equal to S-l +1, and j is not equal to i;

3) for a given X_l(i) (1 ≦ i ≦ S-l +1), and d [ X ] is counted under the condition of tolerance deviation distance r (r > 0)_l(i)，X_l(j)]The number of j (1. ltoreq. j. ltoreq. S-l +1, j. noteq. i) of r or less is marked as B_iThis number B_iAnd d [ X ]_l(i)，X_l(j)]The ratio of the number of (A) to (B) is recorded as:

4) all i's are determined to correspond to

Average value of (1), denoted as B^l(r)：

5) Increasing the dimension l by 1, repeating 1) to 4, with c ═ l +1), yields:

6) the sample entropy of the original EEG signal sequence is then:

in the actual calculation process, S takes a finite value, and the sample entropy is estimated as:

the EEG signal is segmented and decomposed into delta (1-3Hz), theta (4-7Hz), alpha (8-13Hz) and beta (14-30Hz)4 frequency bands, sample entropies of the frequency bands are respectively calculated, the sample entropies of the 4 frequency bands are combined into a sample entropy vector, the sample entropy vector is used for describing the complexity of each rhythm of the EEG signal segment, and the sample entropy vector is used as a third feature vector.

The first feature vector, the second feature vector and the third feature vector are combined into one feature vector as a training sample derived from the EEG signal segmentation. The training sample further comprises class labels, and the class labels are from the third step and the fourth step of the upper computer medical diagnosis auxiliary system, and the second step is a specific process for training and predicting the random forest classifier.

Step four: generally speaking, the more training samples, the better the generalization performance of the trained classifier, in order to fully utilize sample data and improve the generalization capability of the random forest classifier, a bootstrap sampling method (self-help sampling method) is used to randomly extract N samples from the training set, and the N samples are used as a new training set, and K times of repetition are performed to obtain K new different training sets;

step five: taking the K training sets obtained in the fourth step as training data, training a decision tree in each training set, randomly selecting M (M < M) feature subsets from M features (M is a sample feature number) by each node of the decision tree, selecting the optimal feature from the M features when the tree is split each time, recursively repeating the step until the decision tree grows to the maximum extent, splitting the nodes by adopting a CART method in the step, and taking a Gini index GINI value as the basis of splitting the nodes; k decision tree classifiers D are obtained through the step_j(j＝1，...，K)；

Step six: since the training set obtained by bootstrap sampling is used in training the decision tree, for each decision tree (assuming that for the ith tree), about 1/3 training examples do not participate in the generation of the ith tree, which are called out of bag (oob) samples of the ith tree, and the out of bag sample set can be used to evaluate the performance of the random forest classifier. The method comprises the following specific steps:

1) for each sample, calculating it as oob sample tree to classify it;

2) taking simple majority vote as the classification result of the sample;

3) the ratio of the number of the false scores to the total number of the samples is used as the estimation error outside the bag of the random forest;

step seven: adjusting the number K of decision trees and the splitting characteristic number m by adopting a grid search method, randomly scrambling data, and repeating the fourth step to the sixth step;

step eight: selecting the number K of decision trees and the classification feature number m which enable the random forest classifier to have the minimum out-of-bag estimation error, and determining an optimal random forest classification model;

step nine: and using the optimal random forest classification model for the classification prediction of the mental diseases of the patient. Generally speaking, the random forest classifier performs probability prediction on the class of an input sample through K decision trees, each decision tree takes the class with the maximum probability as the prediction class of the sample, and then determines the final classification result of the sample through simple majority voting_i＝[p_i1，p_i2，...，p_iJ](i-1.. K.) random forest combines (by weighted sum or multiplication and normalization) K predicted probability distributions to obtain a final probability distribution P '═ P'₁，p′₂，...，p′_J]I.e. the probability of illness for each disease.

Step ten: the EEG data of each patient is regularly applied, the random forest algorithm is updated and perfected, and the generalization capability of the classifier is improved.

The invention has the following innovation points:

(1) the invention predicts the probability that the patient has the mental diseases such as epilepsy, schizophrenia, depression, autism and the like by a random forest classification algorithm based on the dynamic EEG signal, and assists doctors in medical diagnosis. Generally speaking, a random forest classifier carries out probability prediction on the class of an input sample through K decision trees, each decision tree takes the class with the maximum probability as the prediction class of the sample, and then the final classification result of the sample is determined through simple majority voting;

(2) in order to acquire an EEG signal of a patient, the invention prepares an EEG signal acquisition device, the device is designed based on an integrated chip, a plurality of electrodes are preset to facilitate different lead configurations, so that enough EEG signals can be acquired for processing and analysis, and the effectiveness of an upper computer medical diagnosis auxiliary system is ensured;

(3) the EEG signal acquisition device is small and exquisite and convenient to carry, and the upper computer medical diagnosis auxiliary system can be used in a place with a computer, so that the EEG signal acquisition device is convenient for a doctor to carry and is beneficial for the patient to autonomously operate at any time to autonomously diagnose the mental state.

According to a first aspect of the present application, there is provided a first EEG signal based information processing system comprising: an EEG signal acquisition device and a computer; the EEG signal acquisition device realizes acquisition, amplification, filtering and signal transmission of EEG signals and transmits the EEG signals to the computer; the computer realizes the denoising, the feature extraction and the classification prediction of the EEG signal and carries out visual display.

According to the second EEG signal-based information processing system of the first aspect of the present application, the EEG signal acquisition device is a head-mounted design, and the whole EEG signal acquisition device is designed based on an integrated chip, and includes a detection electrode, a preamplifier, a low-pass filter, a high-pass filter, an analog-to-digital converter, a microcontroller, and a wireless transmission module, so as to realize the acquisition, amplification, low-pass filtering, high-pass filtering, and analog-to-digital conversion of the EEG signal, and transmit the digitized EEG signal to a computer through wireless transmission.

A third EEG signal based information processing system according to the first aspect of the present application, wherein the detecting electrodes comprise two reference electrodes and two collecting electrodes; the left ear electrode is a first reference electrode, the left forehead electrode is a first acquisition electrode, correspondingly, the right ear electrode is a second reference electrode, and the right forehead electrode is a second acquisition electrode; the multiple lead acquisition modes with configurable acquisition electrodes and reference electrodes comprise: in a first configuration mode, a reference electrode at one ear corresponds to a forehead collecting electrode at the same side; in the second configuration mode, the reference electrodes at the left ear and the right ear are connected together to be used as the reference electrode and then matched with one/two acquisition electrodes; or a third configuration, where the collecting electrode at one ear corresponds to the collecting electrode at the other forehead.

A fourth EEG signal based information processing system according to the first aspect of the present application, wherein the preamplifier is a differential amplifier for amplifying the difference between EEG signals of the collecting electrode and the reference electrode; the low-pass filter is a classical second-order active low-pass filter and is used for filtering high-frequency interference, and the high-frequency interference comprises inherent noise of a myoelectricity and/or an electronic device; the high-pass filter is a classical second-order active high-pass filter and is used for filtering low-frequency interference, and the low-frequency interference comprises polarization voltage; the analog-to-digital converter converts the EEG signal subjected to amplification and filtering pretreatment into a digital signal; the microcontroller transmits the digitized EEG signal to a computer via a wireless transmission module.

A fifth EEG signal based information processing system according to the first aspect of the present application, the computer comprising a processor and a memory; the processor implements the following method by executing software: carrying out software filtering on an EEG signal transmitted by an EEG signal acquisition device to remove noise and 50Hz power frequency interference caused by mismatching of a capacitor and a resistor of a hardware filter; the method comprises the steps of segmenting an EEG signal into segments with specified time length by adopting a sliding window, and extracting time domain characteristics, frequency domain characteristics and nonlinear characteristics of each EEG signal segment; and constructing an inference sample based on the features, predicting the classification of one or more inference samples by using a trained random forest classifier, obtaining the probability of the one or more inference samples belonging to the specified class, and performing visual display.

A sixth EEG signal based information processing system according to the first aspect of the present application, wherein the time domain features of the EEG signal segments are obtained by a statistical analysis method, the time domain features comprising maxima, minima, means, variance, kurtosis, skewness and/or wave width of the EEG signal, the obtained time domain features being used as the first feature vector.

A seventh EEG signal based information processing system according to the first aspect of the present application, wherein the frequency domain features of the EEG signal segments are obtained by a periodogram method, the frequency domain features comprise power spectra of the EEG signal segments, and the process of calculating the power spectra is as follows: the EEG signal segment includes S-point observation data X (n) { X (1), X (2),.. times, X (S) } as an energy-limited signal, where S is a positive integer, and n ═ 1, 2.. times, S, and X (n) (n ═ 1, 2.. times, S) is subjected to Fast Fourier Transform (FFT) to obtain a frequency-domain signal sequence X_S(w), where w is the frequency domain variable, then take X_S(w) and dividing by S as an estimate of the EEG signal segment power spectrum, with P_PER(w) represents the power spectrum estimated by the periodogram method, and the formula is as follows:

a time-frequency relationship is established for the power spectrum of the EEG rhythm and EEG signal segments by:

1) the EEG signal is segmented into 4 frequency bands: delta (1-3Hz), theta (4-7Hz), alpha (8-13Hz), beta (14-30 Hz);

2) respectively extracting power spectrum characteristics of 4 frequency bands by using the process of calculating the power spectrum;

the obtained power spectral vector is taken as a second feature vector of the EEG signal segment.

An eighth EEG signal based information processing system according to the first aspect of the present application, wherein the non-linear features of the EEG signal segments are obtained by sample entropy estimation, the process of calculating sample entropy is as follows: for a sequence of S points x (n) { x (1), x (2),. ·, x (S) } included in an EEG signal segment, where S is a positive integer and n ═ 1, 2.., S, the process of calculating the sample entropy is as follows:

1) reconstruction of dimension IVector sequence: x_l(i) (ii) x (i), x (i +1) ·, x (i + l-1) }, wherein 1 ≦ i ≦ S-l +1, the vector sequence representing l consecutive x values starting from the i-th point in the sequence of S points of the EEG signal segment;

3) for a given X_l(i) (1 ≦ i ≦ S-l +1), and d [ X ] is counted under the condition of tolerance deviation distance r (r > 0)_l(i)，X_l(j)]The number of j (1. ltoreq. j. ltoreq.S-l +1, j. noteq.i) ≦ r, this number being denoted B_iThis number B_iAnd d [ X ]_l(i)，X_l(j)]The ratio of the number of (A) to (B) is recorded as:

4) all i's are determined to correspond to

Average value of (1), denoted as B^l(r)：

6) the sample entropy of the sequence of EEG signal segments is:

decomposing the EEG signal segment into delta (1-3Hz), theta (4-7Hz), alpha (8-13Hz) and beta (14-30Hz)4 frequency bands, respectively calculating the sample entropies of the frequency bands by using the process of calculating the sample entropies, combining the sample entropies of the 4 frequency bands into a sample entropy vector, wherein the sample entropy vector is used for describing the complexity of each rhythm of the EEG signal segment, and taking the obtained sample entropy vector as a third feature vector of the EEG signal segment.

A ninth EEG signal based information processing system according to the first aspect of the present application, wherein the processor by executing software constructs training samples for training the random forest classifier also from a plurality of EEG signals; the construction process of the training sample of the random forest classifier is as follows: acquiring a plurality of EEG signals, deriving a first plurality of EEG signal segments having a specified length of time from the acquired EEG signals;

extracting the first feature vector, the second feature vector and the third feature vector for each EEG signal segment of a first plurality of EEG signal segments, wherein the first feature vector, the second feature vector and the third feature vector comprise M features, constructing a sample segment using the first feature vector, the second feature vector and the third feature vector, and adding a label indicating a class to each sample segment, wherein the label indicates the class of the EEG signal to which the sample segment belongs; and taking a sample segment as a training sample of the random forest classifier, wherein a set of training samples is called a training set, the size of the training set is N, and M, N is a positive integer.

According to the applicationA tenth EEG signal based information processing system of the first aspect, wherein the processor, by executing software, further trains the random forest classifier with the training samples; the process of training the random forest classifier with the training samples is as follows: randomly selecting training samples from the training set by using a bootstrap sampling method to obtain K different training subsets, wherein each training subset comprises N training samples; each node of the decision tree randomly selects the optimal feature from the m features of the training samples to split by using one of K training decision trees of K training subsets, so that each decision tree grows to the maximum extent to obtain K decision tree classifiers D_i(j ═ 1.., K); adjusting the number K of decision trees and the splitting characteristic number m by using a grid search method, and constructing an optimal random forest classifier based on the number K of decision trees and the splitting characteristic number m corresponding to the minimum out-of-bag estimation error; wherein K, M, N, M is a positive integer, and M < M.

An eleventh EEG signal based information processing system according to the first aspect of the present application, wherein the processor predicts the classification of one or more inferred samples using a trained random forest classifier by executing software; the process of predicting the classification of one or more inferred samples using a trained random forest classifier is as follows: acquiring an EEG signal, deriving a second plurality of EEG signal segments having a specified length of time from the acquired EEG signal; constructing an inference sample with the first feature vector, the second feature vector, and the third feature vector of each EEG signal segment of a second plurality of EEG signal segments, wherein first feature vector, second feature vector, and third feature vector collectively comprise M features; the random forest classifier predicts the inferred samples, and K decision trees have K prediction probability distributions P_i＝[p_i1，p_i2，…，p_iJ](i ═ 1.,. K.) combining the K predicted probability distributions results in a final probability distribution P '═ P'₁，p′₂，…，p′_J]The probability distribution characterizes the probability of the class of the extrapolated sample, where J is the number of classes and K, J is a positive integer.

A twelfth EEG signal based information processing system according to the first aspect of the present application, wherein the processor updates the trained random forest classifier periodically by executing software; and constructing a new training sample, and updating the trained random forest classifier by using the new training sample so as to improve the generalization capability of the random forest classifier.

Drawings

FIG. 1 is a schematic diagram of an embodiment of the present invention.

Fig. 2 is a front view of an EEG signal acquisition device according to an embodiment of the invention.

Fig. 3 is a top view of an EEG signal acquisition device according to an embodiment of the invention.

Fig. 4 is a schematic diagram of the working principle of an EEG signal acquisition device according to an embodiment of the present invention.

Fig. 5 is a flowchart of a computer-implemented process of the medical diagnosis support system according to the embodiment of the present invention.

FIG. 6 is a flow chart of a random forest classifier construction according to an embodiment of the present invention.

Detailed Description

The present invention will be described in detail below with reference to the accompanying drawings and examples.

The method comprises the steps of firstly sampling EEG signals of normal people and patients with mental diseases such as epilepsy, schizophrenia, depression and autism, marking different category labels, preprocessing the EEG signals, segmenting the signals, extracting electroencephalogram characteristics and the like to construct a training data set, using the training data set to train a random forest classifier, and optimizing parameters of the random forest classifier by a grid search method to obtain an optimal random forest model. Meanwhile, EEG signals of a patient are preprocessed and feature extracted and then are led into an optimal random forest model to predict the probability of mental diseases, and a random forest classifier gives the probability that the patient suffers from the mental diseases such as epilepsy, schizophrenia, depression and autism to assist doctors in medical diagnosis.

In this embodiment, the following concrete steps are performed:

the method comprises the following steps: firstly, collecting a large number of EEG signals of normal people and patients with mental diseases such as epilepsy, schizophrenia, depression, autism and the like;

step two: a medical professional selects an EEG wave band with obvious characteristics within a period of time (30 s in the embodiment), and marks category labels such as '0', '1', '2', '3', '4', … ', J' and the like on EEG signals of normal people, epileptics, schizophrenia patients, depression patients, autism patients and the like respectively;

step three: denoising and preprocessing each EEG signal, cutting the EEG signal into a certain time length (3 s in the embodiment), and extracting time domain information such as amplitude (maximum value, minimum value, mean value, variance and the like), waveform (kurtosis, skewness, wave width and the like) and the like of the EEG signal as time domain features (first feature vectors) by a statistical analysis method; extracting power spectrums of four frequency bands (delta (1-3Hz), theta (4-7Hz), alpha (8-13Hz) and beta (14-30Hz)) in the EEG signal as frequency domain characteristics (second characteristic vector) by a periodogram method; extracting sample entropies of four frequency bands (delta (1-3Hz), theta (4-7Hz), alpha (8-13Hz) and beta (14-30Hz)) in the EEG signal as nonlinear features (third feature vectors) through sample entropy estimation, and combining the first feature vectors, the second feature vectors, the third feature vectors and the sample class labels into training samples;

step four: constructing a plurality of different training sets for training a random forest classifier by using a bootstrap sampling method for training samples obtained in the third step, splitting nodes by using a CART method in the training process, taking Gini index GINI values as the basis of the split nodes, and obtaining an optimal random forest classification model based on the estimation error outside the bag by using a grid search method;

step five: applying the optimal random forest classification model to an upper computer medical diagnosis auxiliary system;

step six: wearing a self-made EEG signal acquisition device for a patient, configuring an electrode lead mode (for example, configuring a reference electrode at the left ear corresponding to a forehead acquisition electrode on the same side), setting a sampling rate (for example, 500Hz), and acquiring an EEG signal of the patient within a period of time;

step seven: the acquired EEG signals are subjected to preliminary processing such as signal amplification, high-pass filtering, low-pass filtering and the like, because the frequency range of the EEG signals is generally 1-30Hz, the cut-off frequency of the high-pass filtering is set to be 0.1Hz, and the cut-off frequency of the low-pass filtering is set to be 100Hz in the embodiment;

step eight: the preprocessed EEG signals are converted into digital signals through an analog-to-digital converter and transmitted to an upper computer through wireless transmission technologies such as Bluetooth and WiFi;

step nine: the upper computer medical diagnosis auxiliary system performs software filtering on the received EEG signal to remove noise and 50Hz power frequency interference caused by mismatching of a capacitor and a resistor of a hardware filter, and obtains a relatively pure EEG signal;

step ten: cutting the EEG signal obtained in the ninth step into fixed lengths (3 s in the embodiment) by adopting a sliding time window, removing myoelectricity, ocular artifacts and the like by adopting an independent component analysis method, and simultaneously extracting time domain characteristics, frequency domain characteristics and nonlinear characteristics of the EEG signal, wherein the characteristic extraction process is as described in the third step;

step eleven: predicting the EEG signal of the patient in the period of time by using an optimal random forest model based on the time domain feature, the frequency domain feature and the nonlinear feature of the EEG signal of the patient, predicting the probability of the EEG signal of the patient suffering from mental diseases such as epilepsy, schizophrenia, depression and autism, and carrying out visual display; such as: assuming that the number of decision trees in the stochastic forest is K, the prediction probability of each decision tree for the patient's EEG signal to be normal, epilepsy, schizophrenia, depression, autism (the number of categories of mental disease in this embodiment is J-5, which may be a larger number in practical use) is P_i＝[p_i1，p_i2，p_i2，p_i4，p_i5](i-1.,. K), random forests combine the predicted probability distributions of K decision trees (by weighting or sum direct multiplication and normalization, in this embodiment by weighting and combining K probability distributions), resulting in a final predicted probability distribution P '═ P'₁，p′₂，p′₃，p′₄，p′₅]Namely:

wherein a is_i∈[0，1]As a function of the weighting coefficients of the decision tree,

step twelve: the doctor makes accurate judgment according to the prediction result and by combining medical experience; assuming that the predicted probability distribution P' obtained in the step eleven is [0.1, 0.61, 0.12, 0.1, 0.07], the upper computer medical diagnosis assistance system predicts that the patient is a normal person with a probability of 10%, an epileptic patient with a probability of 61%, a schizophrenia patient with a probability of 12%, a depression patient with a probability of 10%, and an autism patient with a probability of 7%, and the doctor judges whether the patient has a mental disease or not based on the prediction result and clinical manifestations of the patient and the clinical experience of the doctor;

step thirteen: and the EEG data of each patient is periodically applied to update and perfect the random forest model, so that the generalization capability of the random forest model is improved.

Referring to fig. 1, according to an embodiment of the present application, an EEG signal is acquired. EEG signals are from both normal and psychiatric patients. The EEG signal acquisition device acquires the EEG signal and provides the EEG signal to the computer to realize the acquisition of the EEG signal. An EEG signal acquisition apparatus according to an embodiment of the present application is depicted in fig. 2 and 3.

With continued reference to fig. 1, an EEG signal segment (1) is extracted from the acquired EEG signal. EEG signal segmentation is obtained by performing denoising, segmentation and other processing on the EEG signal. Each EEG signal segment has a specified length of time, for example 3 seconds.

From each EEG signal segment, its time domain features (denoted as first feature vector), frequency domain features (denoted as second feature vector) and non-linear features (denoted as third feature vector) are extracted (2.1). And constructing a training sample for training the random forest classifier by using the first feature vector, the second feature vector and the third feature vector, and adding a classification label (2.2) to the training sample. The class labels correspond to attributes of the person (normal person, or with a specified mental illness) for the EEG signal segment.

It will be appreciated that the extrapolated samples are also constructed by way of EEG signal segmentation. The inference sample includes the first feature vector, the second feature vector, and the third feature vector, but does not include the classification label.

A plurality of training sets (training sample sets) are constructed by a bootstrap sampling method (self-help sampling method). The random forest classifier comprises K decision trees (D)₁，D₂，...，D_K) The training of each decision tree requires 1 sample set (3), so that the number of sample sets is also K, K being a positive integer. The number of training samples of each training set is N, and N is a positive integer.

To train the random forest classifier, parameters K and m are set, K being the number of decision trees and m being the splitting characteristic number of the decision trees.

The corresponding decision tree (4) is trained with the sample set. And the performance of the trained random forest classifier was evaluated with out-of-bag estimation errors (5.1, 5.2). For each combination of parameters K and m < K, m >, the corresponding out-of-bag estimation error is obtained. And adjusting the values of the parameters K and m, retraining the random forest classifier, evaluating the performance of the random forest classifier (6), selecting the parameters K and m corresponding to the optimal estimation error outside the bag, and obtaining the optimal random forest classifier (7).

The inference samples are provided to a trained random forest classifier (8). And each decision tree of the random forest classifier outputs a predicted category label and probability distribution thereof, and the predicted probability distribution of the K decision trees is combined to obtain final predicted probability distribution (9).

Referring to fig. 2, the EEG signal acquisition device of the present application is a head-mounted design, based on an integrated chip design as a whole, having two acquisition electrodes located on the left and right forehead, respectively, and two reference electrodes located near the left and right ear, respectively. The acquisition electrode and the reference electrode can be configured into a plurality of different lead acquisition modes: 1) the reference electrode at one ear corresponds to the forehead collecting electrode at the same side; 2) the reference electrodes at the left ear and the right ear are connected together to be used as reference electrodes and then matched with one/two collecting electrodes; 3) the collecting electrode at one side of the ear corresponds to the collecting electrode at the other side of the forehead; through different lead modes, a doctor can autonomously select an EEG signal acquisition area and flexibly record the absolute values of EEG signal changes of different frontal areas. The EEG signal acquisition device is integrated with an amplification module and a hardware filtering module, and can perform signal amplification, low-pass filtering, high-pass filtering and other preliminary processing on EEG signals acquired by the acquisition electrode. After the EEG signal is subjected to primary processing, the EEG signal is converted into a digital signal through an analog-to-digital converter, and the digital signal is transmitted to an upper computer of the medical diagnosis auxiliary system through a wireless transmission technology.

Fig. 3 is a top view of the EEG signal acquisition apparatus of the present application.

Fig. 4 is a schematic diagram of the working principle of an EEG signal acquisition device according to an embodiment of the invention. Firstly, amplifying the difference of EEG signals between a collecting electrode and a reference electrode through a pre-differential amplifier to obtain EEG signals with larger amplitude; carrying out low-pass filtering on the EEG signal to filter out high-frequency interference such as myoelectricity, inherent noise of an electronic device and the like, and carrying out high-pass filtering to filter out low-frequency interference such as polarization voltage and the like; and then an analog-to-digital converter is used for converting the EEG signals subjected to preliminary processing such as amplification, filtering and the like into digital signals, and the digital signals are transmitted to a computer through a wireless transmission module.

The upper computer of the medical diagnosis support system of the present application is, for example, a computer or a server. The upper computer includes, for example, a processor and a memory, and the memory stores a program. The processor executes the program to implement one or more process flows according to embodiments of the application, such as the process flows provided in connection with fig. 5 and 6.

Referring to fig. 5, EEG signal processing according to an embodiment of the present application includes:

and carrying out software self-adaptive filtering on the EEG signal transmitted by the EEG signal acquisition device to remove noise and 50Hz power frequency interference caused by unmatched capacitors and resistors of a hardware filter and the like, so as to obtain a relatively pure EEG signal. And cutting the EEG signal into EEG signal segments with fixed time length by adopting a sliding window, and removing myoelectricity, ocular artifacts and the like by adopting an independent component analysis method. And processing the EEG data by using methods such as statistical analysis, power spectrum estimation, sample entropy estimation and the like, and extracting time domain characteristics, frequency domain characteristics and nonlinear characteristics of the EEG signal segments. And (3) constructing a training sample by using the EEG signal segmentation, training a random forest classifier, and constructing an optimal random forest classification model based on an out of bag estimation error (out of bag error).

And constructing an inference sample by using time domain characteristics, frequency domain characteristics and nonlinear characteristics of EEG signal segments from the patient, predicting the EEG signal of the patient in the time period by using a trained random forest classifier, predicting the probability of the patient suffering from mental diseases such as epilepsy, schizophrenia, depression, autism and the like, and carrying out visual display. And optionally, the doctor makes an accurate judgment according to the prediction result and by combining medical experience. It is to be appreciated that the process by which the doctor makes a decision based on the prediction of the random forest classifier is not part of the process flow performed by the computer according to embodiments of the present application.

FIG. 6 is a flow chart of a random forest classifier construction according to an embodiment of the present invention. The process of obtaining training samples and training the random forest classifier with the training samples has been described in detail above and will not be repeated here.

Claims

1. An information processing system based on EEG signals, comprising: an EEG signal acquisition device and a computer; the EEG signal acquisition device realizes acquisition, amplification, filtering and signal transmission of EEG signals and transmits the EEG signals to the computer; the computer realizes the denoising, the feature extraction and the classification prediction of the EEG signal and carries out visual display.

2. The information processing system of claim 1, wherein the EEG signal acquisition device is a head-mounted design, based on an integrated chip design as a whole, comprising detection electrodes, a preamplifier, a low-pass filter, a high-pass filter, an analog-to-digital converter, a microcontroller and a wireless transmission module, and is configured to acquire, amplify, low-pass filter, high-pass filter and analog-to-digital convert the EEG signals and transmit the digitized EEG signals to the computer via wireless transmission.

3. The information handling system of claim 2, wherein the detection electrodes comprise two reference electrodes and two acquisition electrodes; the left ear electrode is a first reference electrode, the left forehead electrode is a first acquisition electrode, correspondingly, the right ear electrode is a second reference electrode, and the right forehead electrode is a second acquisition electrode; the multiple lead acquisition modes with configurable acquisition electrodes and reference electrodes comprise:

in a first configuration mode, a reference electrode at one ear corresponds to a forehead collecting electrode at the same side;

in the second configuration mode, the reference electrodes at the left ear and the right ear are connected together to be used as the reference electrode and then matched with one/two acquisition electrodes; or

In a third configuration mode, the collecting electrode at one side of the ear corresponds to the collecting electrode at the other side of the forehead;

the preamplifier is a differential amplifier and is used for amplifying the difference between EEG signals of the collecting electrode and the reference electrode;

the low-pass filter is a classical second-order active low-pass filter and is used for filtering high-frequency interference, and the high-frequency interference comprises inherent noise of a myoelectricity and/or an electronic device;

the high-pass filter is a classical second-order active high-pass filter and is used for filtering low-frequency interference, and the low-frequency interference comprises polarization voltage;

the analog-to-digital converter converts the EEG signal subjected to amplification and filtering pretreatment into a digital signal;

the microcontroller transmits the digitized EEG signal to a computer via a wireless transmission module.

4. The information handling system of claim 1, the computer comprising a processor and a memory; the processor implements the following method by executing software:

carrying out software filtering on an EEG signal transmitted by an EEG signal acquisition device to remove noise and 50Hz power frequency interference caused by mismatching of a capacitor and a resistor of a hardware filter; the method comprises the steps of segmenting an EEG signal into segments with specified time length by adopting a sliding window, and extracting time domain characteristics, frequency domain characteristics and nonlinear characteristics of each EEG signal segment; constructing an inference sample based on the features, predicting the classification of one or more inference samples by using a trained random forest classifier, obtaining the probability that one or more inference samples belong to a specified class, and performing visual display;

the time domain features of the EEG signal segments are obtained through a statistical analysis method, the time domain features comprise the maximum value, the minimum value, the mean value, the variance, the kurtosis, the skewness and/or the wave width of the EEG signal, and the obtained time domain features serve as first feature vectors.

5. The information processing system according to claim 4, wherein the frequency domain features of the EEG signal segment are acquired by periodogram method, the frequency domain features comprising power spectra of the EEG signal segment, the process of calculating the power spectra is as follows:

the EEG signal segment includes S-point observation data X (n) { X (1), X (2),.. times, X (S) } as an energy-limited signal, where S is a positive integer, and n ═ 1, 2.. times, S, and X (n) (n ═ 1, 2.. times, S) is subjected to Fast Fourier Transform (FFT) to obtain a frequency-domain signal sequence X_S(w), where w is the frequency domain variable, then take X_S(w) and dividing by S as an estimate of the EEG signal segment power spectrum, with P_PER(w) represents the power spectrum estimated by the periodogram method, and the formula is as follows:

6. The information processing system according to claim 4 or 5, wherein the non-linear features of the EEG signal segments are obtained by sample entropy estimation, the process of calculating sample entropy is as follows:

for a sequence of S points x (n) { x (1), x (2),. ·, x (S) } included in an EEG signal segment, where S is a positive integer and n ═ 1, 2.., S, the process of calculating the sample entropy is as follows:

1) reconstructing the l-dimensional vector sequence: x_l(i) (ii) x (i), x (i +1) ·, x (i + l-1) }, wherein 1 ≦ i ≦ S-l +1, the vector sequence representing l consecutive x values starting from the i-th point in the sequence of S points of the EEG signal segment;

4) all are solvedi corresponds to

Average value of (1), denoted as B^l(r)：

6) the sample entropy of the sequence of EEG signal segments is:

7. The information handling system of claim 6, wherein the processor, by executing software, constructs training samples for training the random forest classifier further from a plurality of EEG signals;

the construction process of the training sample of the random forest classifier is as follows:

acquiring a plurality of EEG signals, deriving a first plurality of EEG signal segments having a specified length of time from the acquired EEG signals;

8. The information handling system of claim 7, wherein the processor, by executing software, further trains the random forest classifier with the training samples;

the process of training the random forest classifier with the training samples is as follows:

randomly selecting training samples from the training set by using a bootstrap sampling method to obtain K different training subsets, wherein each training subset comprises N training samples; each node of the decision tree randomly selects the optimal feature from the m features of the training samples to split by using one of K training decision trees of K training subsets, so that each decision tree grows to the maximum extent to obtain K decision tree classifiers D_j(j ═ 1.., K); adjusting the number K of decision trees and the splitting characteristic number m by using a grid search method, and constructing an optimal random forest classifier based on the number K of decision trees and the splitting characteristic number m corresponding to the minimum out-of-bag estimation error; wherein K, M, N, M is a positive integer, and M < M.

9. The information handling system of claim 8, wherein the processor predicts a classification of one or more inferred samples using a trained random forest classifier by executing software;

the process of predicting the classification of one or more inferred samples using a trained random forest classifier is as follows:

acquiring an EEG signal, deriving a second plurality of EEG signal segments having a specified length of time from the acquired EEG signal;

constructing an inference sample with the first feature vector, the second feature vector, and the third feature vector of each EEG signal segment of a second plurality of EEG signal segments, wherein first feature vector, second feature vector, and third feature vector collectively comprise M features;

the random forest classifier predicts the inferred samples, and K decision trees have K prediction probability distributions P_i＝[p_i1，p_i2，...，p_iJ](i ═ 1.,. K.) combining the K predicted probability distributions results in a final probability distribution P '═ P'₁，p′₂，…，p′_J]The probability distribution characterizes the probability of the class of the extrapolated sample, where J is the number of classes and K, J is a positive integer.

10. The information handling system of claim 9, wherein the processor periodically updates the trained random forest classifier by executing software;

and constructing a new training sample, and updating the trained random forest classifier by using the new training sample so as to improve the generalization capability of the random forest classifier.