CN113729729A

CN113729729A - Schizophrenia early detection system based on graph neural network and brain network

Info

Publication number: CN113729729A
Application number: CN202110932958.4A
Authority: CN
Inventors: 张冀聪; 常琪
Original assignee: Beihang University
Current assignee: Beihang University
Priority date: 2021-08-13
Filing date: 2021-08-13
Publication date: 2021-12-03
Anticipated expiration: 2041-08-13
Also published as: CN113729729B

Abstract

The system comprises an EEG signal acquisition unit, a signal processing unit and a signal processing unit, wherein the EEG signal acquisition unit is used for acquiring 128 channels of EEG signals of a tester during the auditory stimulation of an MMN paradigm; the source tracing analysis unit is used for dividing the EEG signals into four frequency bands, and performing source tracing analysis on the EEG signals of all the frequency bands respectively to reconstruct the EEG signals into endogenous electrical activity of the cortex in the brain; the brain network construction unit is used for constructing a brain network of each frequency band according to the endogenous electrical activity of the cortex layer in the brain; the neural network unit of the picture, is used for extracting the brain network characteristic of each corresponding frequency channel from the brain network of each said frequency channel, the neural network unit of the said picture includes the neural network module of the full connection, the said neural network module of the said full connection is used for receiving the demographic characteristic, cognitive characteristic, relevant electric potential characteristic of event and said brain network characteristic and classifying, output the classification result as the testing result. The detection system can improve the accuracy of early identification of schizophrenia.

Description

Schizophrenia early detection system based on graph neural network and brain network

Technical Field

The invention relates to the field of early detection of schizophrenia, in particular to an early detection system of schizophrenia based on a neural network and a brain network.

Background

Schizophrenia is the most common mental disease characterized by various disorders such as thinking, emotion and behavior, and by the incoordination of mental activities. At present, the cause of schizophrenia is not clear, biological indexes are lacked, early identification cannot be carried out, the attack process is slow and hidden, the schizophrenia is frequently repeatedly attacked, if the schizophrenia is not found as early as possible, the optimal diagnosis time is delayed, and even the schizophrenia possibly causes damage to oneself and other people.

The main symptoms of schizophrenia include negative symptoms, positive symptoms, and cognitive impairment, and the appearance of cognitive impairment occurs earlier before the appearance of the disease. The study of cognitive disorders in schizophrenia has focused on higher cognitive functions, but basal sensory impairments are present in schizophrenia as well and affect higher cognitive functions. The most studied of the fundamental sensory disturbances are auditory processing defects, i.e. impairment of the sound coding function. Therefore, finding a good clinical objective marker to help the detection of clinically good abnormalities from the impaired hearing function is the focus of the current study on the impaired sensation of schizophrenia.

The current diagnosis of schizophrenia mainly depends on detection and identification of symptoms by medical staff after clinical symptoms are shown, and therefore the diagnosis has strong subjectivity. Currently, there is a great need for objective indicators that effectively assist in clinical detection and early screening of schizophrenia.

Therefore, there is a need to develop an early schizophrenia detection system based on a graph neural network and a brain network to solve one or more of the above technical problems.

Disclosure of Invention

To solve at least one of the above technical problems, the applicant has studied to find that a dysfunction of functional interactions (i.e. brain networks) between different brain regions in schizophrenia is considered to be the cause of cognitive impairment in schizophrenia. Brain function is not performed independently by a single neuron or a single brain region, but rather by interactions between different regions of the brain. Thus, neurological dysfunction is associated with impaired brain connections (brain networks). In combination with the research results, according to an aspect of the present invention, there is provided an early schizophrenia detection system based on a neural network and a brain network, comprising:

the EEG signal acquisition unit is used for acquiring EEG signals of 128 channels of a tester during the auditory stimulation of the MMN paradigm;

the source tracing analysis unit is used for dividing the EEG signals into four frequency bands of delta (1.5-4Hz), theta (4-8Hz), alpha (8-13Hz) and beta (13-30Hz), and performing source tracing analysis on the EEG signals of all the frequency bands respectively to reconstruct the EEG signals into endogenous electrical activity of the brain cortex;

the brain network construction unit is used for constructing a brain network of each frequency band according to the endogenous electrical activity of the cortex layer in the brain;

the neural network unit of the picture, is used for extracting the brain network characteristic of each corresponding frequency channel from the brain network of each said frequency channel, the neural network unit of the said picture includes the neural network module of the full connection, the said neural network module of the said full connection is used for receiving the demographic characteristic, cognitive characteristic, relevant electric potential characteristic of event and said brain network characteristic and classifying, output the classification result as the testing result.

According to yet another aspect of the invention, the demographic characteristics include age, educational age, and IQ of the tester.

According to yet another aspect of the invention, the event-related potential signature comprises the amplitude and latency of a mismatched negative wave in the EEG signal.

According to another aspect of the invention, the system for detecting the early schizophrenia based on the graph neural network and the brain network further comprises a complete set of test units for cognitive functions of the schizophrenia, and the complete set of test units is used for acquiring the cognitive features.

According to another aspect of the present invention, the early schizophrenia detection system based on the neural network and the brain network further comprises a detection result output unit, the fully-connected neural network module outputs four classification results for the brain network features of four frequency bands, the detection result output unit outputs a detection result according to the four classification results, and the detection result is determined by three or more same classification results.

According to another aspect of the present invention, there is also provided a method for early detection of schizophrenia based on a neural network and a brain network, comprising the steps of:

acquiring 128-channel EEG signals of a test subject during auditory stimulation in the MMN paradigm;

dividing the EEG signals into four frequency bands of delta (1.5-4Hz), theta (4-8Hz), alpha (8-13Hz) and beta (13-30Hz), and respectively carrying out source tracing analysis on the EEG signals of each frequency band to reconstruct the EEG signals into endogenous electrical activity of the cortical layer in the brain;

constructing a brain network of each frequency band according to the endogenous electrical activity of the cortex layer in the brain;

extracting corresponding brain network characteristics of each frequency band from the brain networks of each frequency band by using a graph neural network;

and receiving and classifying the demographic characteristics, the cognitive characteristics, the event-related potential characteristics and the brain network characteristics through a fully-connected neural network module, and outputting a classification result as a detection result.

According to yet another aspect of the invention, the event-related potential signature comprises the amplitude and latency of a mismatch negative wave in the EEG signal.

According to yet another aspect of the invention, the cognitive characteristics are acquired using a cognitive function test suite of schizophrenia.

According to another aspect of the present invention, four classification results are output for the brain network features of four frequency bands by the fully-connected neural network module, and a detection result is output according to the four classification results, wherein the detection result is determined by three or more same classification results.

The invention can obtain one or more of the following technical effects:

the Graph Neural Network (GNN) model in the invention combines the characteristics, the demographic characteristics and the cognitive characteristics of MMN (Mismatch negative wave) event-related potential, improves the early identification accuracy rate of schizophrenia, and can provide reference for doctors;

the obstacle of functional interaction between different brain areas in schizophrenia can be identified by constructing a brain network and extracting brain network characteristics, so that the classification accuracy of first schizophrenia, chronic schizophrenia and healthy subjects is improved;

the brain network characteristics of four frequency bands and four corresponding classification results are obtained through one-time detection, and three or more same classification results are determined as detection results, so that the detection accuracy is further improved.

Drawings

The present invention will be described in further detail with reference to the accompanying drawings and specific embodiments.

Fig. 1 is a schematic diagram of an early schizophrenia detection system based on a graph neural network and a brain network according to a preferred embodiment of the present invention.

Fig. 2 is a schematic diagram of brain networks constructed in different frequency bands for different populations (FESZ, CSZ and HC).

Fig. 3 is a graph comparing the effect (fMMN, frequency deviation stimulation) of the neural network and a classical Support Vector Machine (SVM) for extracting brain network features.

Fig. 4 is a graph comparing the effect (dMMN, duration-biased stimulation) of neural networks and a classical Support Vector Machine (SVM) for extracting brain network features.

Detailed Description

The best mode for carrying out the present invention will be described in detail with reference to the accompanying drawings, wherein the detailed description is for the purpose of illustrating the invention in detail, and is not to be construed as limiting the invention, as various changes and modifications can be made therein without departing from the spirit and scope thereof, which are intended to be encompassed within the appended claims.

Example 1

According to a preferred embodiment of the present invention, referring to fig. 1 to 3, there is provided an early schizophrenia detection system based on a graph neural network and a brain network, comprising:

the brain network construction unit is used for constructing a brain network (namely a brain network topological graph) of each frequency band according to the endogenous electrical activity of the cortical layer in the brain;

According to a further preferred embodiment of the invention, the demographic characteristics include age, educational age and IQ of the tester.

According to a further preferred embodiment of the invention, the event-related potential signature comprises the amplitude and latency of a mismatch negative wave in the EEG signal.

According to another preferred embodiment of the present invention, the system for detecting schizophrenia based on the graph neural network and the brain network further comprises a complete set of test units for cognitive function of schizophrenia, which is used for acquiring the cognitive characteristics.

According to another preferred embodiment of the present invention, the early schizophrenia detection system based on a neural network and a brain network of a graph further comprises a detection result output unit, the fully-connected neural network module outputs four classification results for brain network features of four frequency bands, the detection result output unit outputs a detection result according to the four classification results, and the detection result is determined by three or more same classification results.

Preferably, the classification results include first-onset schizophrenia (FESZ), Chronic Schizophrenia (CSZ) and health.

According to another preferred embodiment of the present invention, there is provided a method for early detection of schizophrenia based on a neural network and a brain network, comprising the steps of:

According to a further preferred embodiment of the invention, the cognitive characteristics are acquired using a test suite of cognitive functions for schizophrenia.

According to another preferred embodiment of the present invention, the fully-connected neural network module outputs four classification results for the brain network features of four frequency bands, and outputs a detection result according to the four classification results, wherein the detection result is determined by three or more same classification results.

Preferably, the demographic records the age, IQ and age of education of the subject. Cognitive assessment 7 psychological dimensions were assessed using the schizophrenia Cognitive function test set (MCCB), which comprises 10 point tests: (1) processing Speed (SOP), including link testing (TMT), symbol coding testing (SC), and semantic fluency testing (CF); (2) attention/alertness (Attention/Vigilance), i.e., continuous manipulation test (CPT-IP); (3) working Memory (WM) including digital sequence test (DS) and spatial breadth test (SS); (4) speech learning and memory, i.e., speech memory test (HVLT-R); (5) visual Learning and Memory (Visual Learning and Memory), i.e., Visual Memory test (BVMT-R); (6) reasoning and Problem Solving capabilities (learning and Problem Solving), namely maze test (MAZES); (7) social Cognition (Social Cognition), i.e., a test for emotion management.

Preferably, the raw brain electrical signals are acquired and a total of 825 auditory stimuli are acquired in the MMN paradigm, including 75 stimuli each (9% each, 150 stimuli total, 18%) and 675 stimuli (82%) of the two types of deviation stimuli. The sound stimuli were communicated to the subject through the headphones with a time interval of 500-. The frequency of the standard stimulation design was 1000Hz, the intensity was 75dB, and the duration was 50 ms. Two bias stimuli were included in the experiment, frequency bias (frequency 1500Hz, intensity 75dB, duration 50ms) and duration bias (frequency 1000Hz, intensity 75dB, duration 100 ms). The first 15 stimuli were set as standard stimulus types. Participants did not need to actively identify task stimuli during the experiment.

Preferably, the acquired EEG raw signals are first subjected to an offline series of pre-processing: re-referencing all the lead data, then performing band-pass filtering at 0.1Hz to 40Hz, removing eye movement artifacts by independent principal component analysis, and finally manually removing the artifacts by manually browsing all the data. Taking 100ms before each stimulation and 500ms after each stimulation, performing baseline calibration on the first 100ms, and performing superposition averaging on waveforms intercepted by all standard stimulations of each tested, wherein similarly, waveforms intercepted by two types of deviation stimulations are also respectively subjected to superposition averaging. The waveform obtained by superposing and averaging the standard stimulus is subtracted from the waveform obtained by stimulating with the frequency deviation, and the waveform of each tested frequency MMN is obtained. Likewise, the duration MMN may also be obtained. And performing superposition averaging on the waveforms of each group to obtain the MMN waveform of each group. The MMN component occurs in the middle of 100-.

Preferably, the scalp electroencephalographic signals are subjected to a traceability analysis using the eLORETA method to reconstruct the cortical time series in the brain. The activity of the central voxel source of 80 brain regions is extracted to represent each brain region based on AAL brain region.

Preferably, a brain network at the source level is constructed. Referring to fig. 2, a network is composed of nodes and edges connecting the nodes. The correlation between brain regions is evaluated by Phase Lagged Index (PLI) at the edges of the brain network, so that a large-scale brain function network with the brain regions as nodes is constructed. Phase lag exponent is a method of detecting asymmetry in the phase difference distribution between two signals, which reflects the consistency of the phase advance or retard of one signal relative to the other, and is an effective estimate of phase synchronization. PLI is not sensitive to the bulk conductor effect of the signals and may only be concerned with the coupling relationship between the signals.

PLI makes an estimate of the phase synchronism insensitive to signals from the same source (e.g., bulk conductor effects or active references) by calculating the asymmetry of the phase difference distribution law, which appears as a phase difference from 0. This distribution is horizontal when no phase coupling relationship exists between the two time series. Any deviation from this horizontal distribution is an indication of phase synchronism.

The asymmetry index of the phase difference distribution may be derived from a time t-sequence of phase differences

Obtaining:

sign is here a sign function, assuming:

then the variation range of PLIThe circumference is 0 to 1, i.e. PLI is more than or equal to 0 and less than or equal to 1. A PLI of 0 indicates that there is no coupling or a coupling with a phase difference close to 0mod π. PLI of 1 is indicated in

A perfect phase lock at a value different from 0mod pi. The stronger this non-0 phase lock, the closer the PLI is to 1.

Preferably, the GNN model is trained in advance. First, the convolutional layer extracts and learns network feature information useful for classification by extracting graph features in brain network data through hierarchical graph convolution and graph pooling, approximating the convolution kernel using chebyshev spectrogram convolution operator (ChebConv), while using node features and edge weights. Next, graph pooling is an important operation to avoid overfitting and reduce information redundancy and noise, and is achieved by toppkpooling. The brain network characteristics obtained by the graph convolution model can be regarded as compressed and refined representations of an original graph, the characteristics are reshaped into vectors and fed to a full-connection layer, the output of the vectors is connected with other quantitative indexes (including demographic characteristics, event-related potential characteristics and cognitive characteristics), and then the vectors are classified by a full-connection neural network module.

The classification accuracy of the five-fold cross validation of the invention is shown in table 1, and the whole is relatively high.

TABLE 1 figure neural network classification accuracy

Further, the invention compares the effect of the application of the GNN and the classical SVM to the extraction of brain network features, and the result is shown in Table 2 (the results of 10 times of five-fold cross validation of the SVM and GNN classifier on the first schizophrenia, the chronic schizophrenia and the healthy subjects), the GNN classification result is obviously superior to that of the SVM, and the invention provides an effective method for distinguishing FESZ, CSZ and HC groups.

TABLE 2 accuracy of GNN and SVM classification (%)

Further, see fig. 3, which shows the results of four performance indicators (accuracy, recall, accuracy and F1 score) for classification of FESZ, CSZ and HC patterns by GNN and SVM classifiers, respectively, in the case of frequency deviation stimulation (fMMN). See fig. 4, which shows the results of four performance indicators (accuracy, recall, accuracy and F1 score) for classification of FESZ, CSZ and HC patterns by GNN and SVM classifiers, respectively, in the case of duration deviation stimulation (dMMN). Graphs of the GNN and SVM classification results are shown to the left and right of each dashed line, respectively. FESZ: first-onset schizophrenia; CSZ: chronic schizophrenia; HC: healthy control.

By comparison, under the condition of duration deviation stimulation (dMMN), four performance indexes (accuracy, recall rate, precision and F1 score) for classifying the FESZ, CSZ and HC modes by the GNN classifier are best. Advantageously, during the acquisition of the electroencephalogram signals, the auditory stimulation of the duration deviation is given to the testee, so that the accuracy of the detection result is further improved.

The invention can obtain one or more of the following technical effects:

It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are described in the specification and illustrated only to illustrate the principle of the present invention, but that various changes and modifications may be made therein without departing from the spirit and scope of the present invention, which fall within the scope of the invention as claimed. The scope of the invention is defined by the appended claims and equivalents thereof.

Claims

1. An early schizophrenia detection system based on a graph neural network and a brain network, which is characterized by comprising:

2. The early detection system of schizophrenia based on graph neural network and brain network as claimed in claim 1, wherein the demographic characteristics comprise age, education age and IQ of testers.

3. The early schizophrenia detection system based on graph neural networks and brain networks as claimed in claim 2, wherein the event-related potential characteristics comprise amplitude and latency of mismatch negative waves in the EEG signals.

4. The early schizophrenia detection system based on graph neural network and brain network as claimed in any one of claims 1-3, further comprising a test set of cognitive function test unit for schizophrenia for obtaining the cognitive characteristics.

5. The early schizophrenia detection system based on neural network and brain network of claim 4, further comprising a detection result output unit, wherein the fully connected neural network module outputs four classification results for the brain network features of four frequency bands, the detection result output unit outputs a detection result according to the four classification results, the detection result is determined by three or more same classification results, and the MMN-paradigm auditory stimulation is preferably duration deviation stimulation.

6. A schizophrenia early detection method based on a graph neural network and a brain network is characterized by comprising the following steps:

7. The method of claim 6, wherein the demographic characteristics include age, educational age, and IQ of the tester.

8. The method of claim 7 wherein the event-related potential signature comprises the amplitude and latency of mismatched negative waves in the EEG signal.

9. The method for the early detection of schizophrenia based on neural networks and brain networks of graphs as claimed in any one of claims 6-9, wherein the cognitive characteristics are obtained by using a test set of cognitive functions of schizophrenia.

10. The method according to claim 9, wherein four classification results are outputted for the brain network features of four frequency bands by the fully-connected neural network module, and a detection result is outputted according to the four classification results, wherein the detection result is determined by three or more same classification results, and the MMN-paradigm auditory stimulation is preferably duration deviation stimulation.