CN115414054A - Epilepsia electroencephalogram detection system based on feedforward pulse neural network - Google Patents
Epilepsia electroencephalogram detection system based on feedforward pulse neural network Download PDFInfo
- Publication number
- CN115414054A CN115414054A CN202211178565.XA CN202211178565A CN115414054A CN 115414054 A CN115414054 A CN 115414054A CN 202211178565 A CN202211178565 A CN 202211178565A CN 115414054 A CN115414054 A CN 115414054A
- Authority
- CN
- China
- Prior art keywords
- pulse
- neural network
- snn
- neuron
- electroencephalogram
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
- A61B5/369—Electroencephalography [EEG]
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/40—Detecting, measuring or recording for evaluating the nervous system
- A61B5/4076—Diagnosing or monitoring particular conditions of the nervous system
- A61B5/4094—Diagnosing or monitoring seizure diseases, e.g. epilepsy
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/72—Signal processing specially adapted for physiological signals or for diagnostic purposes
- A61B5/7235—Details of waveform analysis
- A61B5/7264—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
- A61B5/7267—Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Health & Medical Sciences (AREA)
- Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Theoretical Computer Science (AREA)
- Public Health (AREA)
- Veterinary Medicine (AREA)
- Mathematical Physics (AREA)
- Pathology (AREA)
- Neurology (AREA)
- Heart & Thoracic Surgery (AREA)
- Medical Informatics (AREA)
- Evolutionary Computation (AREA)
- Surgery (AREA)
- Animal Behavior & Ethology (AREA)
- Physiology (AREA)
- General Physics & Mathematics (AREA)
- Psychiatry (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Neurosurgery (AREA)
- Software Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Fuzzy Systems (AREA)
- Signal Processing (AREA)
- Psychology (AREA)
- Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)
Abstract
The invention discloses an epilepsia electroencephalogram detection system based on a feedforward pulse neural network, which defines an error function of a pulse sequence and a learning rule of a synapse weight and explores a network model suitable for epilepsia electroencephalogram identification. In order to verify the performance of an algorithm reconstructed based on an expected pulse sequence, a three-layer feedforward pulse neural network structure is constructed, then an expected output pulse sequence of the pulse neural network is set, and a learning algorithm reconstructed based on the expected pulse sequence is used for adjusting the synaptic weights of the neural network. And judging that the learning algorithm framework reconstructed based on the expected pulse sequence has better learning capability according to the pulse sequence learning result and the change situation of the synaptic weight values before and after learning.
Description
Technical Field
The invention belongs to the technical field of electroencephalogram signal detection, and particularly relates to an epilepsia electroencephalogram detection system based on a feedforward pulse neural network.
Background
1. Current status of electroencephalogram identification of epilepsy
The pulse time coding and information processing mechanism in the impulse neural network is gradually becoming a popular research direction in recent years, and the research enthusiasm of researchers on the impulse neural network is aroused. The pulse neuron model has higher biological authenticity, and compared with the traditional artificial neural network based on pulse frequency coding information, the pulse neural network formed by the pulse neuron model serving as a basic unit has stronger computing power and is very suitable for processing and analyzing brain neural signals. Researchers construct different impulse neural network models and learning mechanisms, and the calculation models have remarkable effects in the aspects of pattern recognition, decision making, prediction and the like.
The accurate epilepsia electroencephalogram recognition result has very important practical significance for diagnosis and treatment of epileptics, when electroencephalogram data are processed and classified by the traditional machine learning technology, the accuracy of pattern recognition depends on the obtained characteristic values, and the selection of the characteristic values is usually determined by an extraction method of a user and personal experience of a researcher, so that the characteristic values sometimes cannot accurately represent the inherent attributes of the electroencephalogram data. How to overcome the difficulty faced by electroencephalogram signal identification to further promote the development of brain science and neurological disease diagnosis technology is a challenging research direction.
Early electroencephalogram signals are distinguished manually, time and labor are consumed, results are subjective, and the results are generally related to personal experience and professional level of an analyst. The subsequent study on the electroencephalogram signals mostly assumes that the electroencephalogram signals are relatively stable in a short time, and then the electroencephalogram signals are analyzed by using a traditional signal processing method. With the proposal of some non-linear electroencephalogram data classification methods, more and more new technologies are rapidly applied to various fields such as biology, medicine and the like. As a self-adaptive mode recognition technology, the artificial neural network becomes an important method for carrying out nonlinear analysis and recognition on electroencephalogram signals.
At present, researchers research on the technical aspects of depression identification, sleep detection, epilepsy diagnosis and the like, and the auxiliary diagnosis result is obtained by processing EEG signals. Besides, the practical application in the electroencephalogram signal analysis also comprises emotion analysis, risk assessment of stroke occurrence, schizophrenia diagnosis, alzheimer's disease detection and the like. The traditional artificial neural network is used for analyzing and identifying the electroencephalogram data, and generally the following processes are adopted: data acquisition, signal preprocessing, feature extraction, and finally, dividing EEG signals into different categories by using different types of neural network classifiers. The traditional artificial neural network has strong dependence on feature extraction, and the classification precision of the traditional artificial neural network is greatly influenced by the quality of the feature extraction, so that researchers in the past spend much time and energy on developing a proper feature extraction technology. After the impulse neural network model is provided, the impulse neural network model has the characteristic of processing complex spatiotemporal pattern data, so that the impulse neural network model is considered to have greater potential, is more suitable for discovering important hierarchical relations existing in the complex data through self algorithms, and does not need to laboriously and manually extract features.
Unlike traditional ANNs that process real-valued inputs and outputs, SNNs employ a pulse neuron model with inputs that are pulse sequences into which problem-specific data is encoded, and the processing and transmission of information in the network is performed in the form of pulse sequences. In addition, the SNN network structure adopting multi-synapse connection is also researched, and the information processing capability of the SNN is further expanded.
For the single-layer pulse neural network, because the structure is simple and the structure of a learning algorithm is relatively easy, at present, many SNN (single layer network) supervised learning algorithms based on a single-layer structure exist and can be successfully applied to the fields of space-time pattern recognition and the like. However, for some complex problems, single-layer SNN gradually shows its limitations, and research on multi-layer network learning algorithms becomes an urgent practical requirement. Therefore, a new algorithm is provided on the basis of the correlation theory of pulse sequence inner product and SNN supervised learning, and can be used for the supervised learning training of the multilayer feedforward pulse neural network.
2. Research of existing patent on epilepsia electroencephalogram
Chinese patent CN107153825A is an epilepsia electroencephalogram classification method based on a support vector machine, and provides a signal classification method based on particle swarm optimization to optimize support vector machine parameters according to particle swarm optimization and support vector machine theory aiming at the problems of low epilepsia electroencephalogram classification accuracy and few classification categories in the prior art. The invention is suitable for classification of electroencephalogram signals of normal, epileptic seizure and epileptic seizure states.
The Chinese patent CN107616780A discloses an electroencephalogram detection method and device using a wavelet neural network, and the electroencephalogram detection method comprises the steps of acquiring electroencephalogram signals, filtering and denoising the electroencephalogram signals, extracting relative amplitude and relative fluctuation indexes, training a classifier and calculating a predicted value. The method utilizes the relative amplitude and the relative fluctuation index with better characteristic effect to carry out characteristic extraction on the acquired and preprocessed electroencephalogram data, and sends the extracted characteristic vector into the classifier obtained by the wavelet neural network, thereby obtaining the mark of the abnormal electroencephalogram signal, not only reducing the workload of a clinician for distinguishing large-scale electroencephalogram data, but also improving the timeliness of detecting the abnormal electroencephalogram.
Chinese patent CN105249962B scalp electroencephalogram retrospective epileptic seizure point detection method and system provides a scalp electroencephalogram retrospective epileptic seizure point detection method and system. The method retrospectively analyzes the electroencephalogram signals without various artifacts by a nonlinear dynamics sample entropy threshold detection method to determine epileptic seizure points. The retrospective epileptic seizure point detection system for the scalp electroencephalogram signals comprises an electroencephalogram signal receiving module, an epileptic seizure point determining module and an information output module. The electroencephalogram signal receiving module is used for receiving clinically collected original electroencephalogram signals. The epileptic seizure point determining module is used for analyzing and determining retrospective epileptic seizure points through the electroencephalogram signals received by the electroencephalogram signal receiving module. The information output module is used for outputting the retrospective epileptic seizure point determined by the epileptic seizure point determination module.
3. Disadvantages of the background Art
There are two conventional diagnostic methods, scalp brain waves and deep brain waves, respectively, and multiple sampling points are required.
The traditional diagnostic method has the following five defects:
1. time consumed for diagnosis: the doctor monitors the brain waves for a long time and then makes a diagnosis, and the method is long in time consumption, needs certain human resources, and wastes time and labor.
2. The diagnosis accuracy is as follows: most data noise and interference exist in the diagnosis process, and the diagnosis accuracy is reduced to a certain extent.
3. Data processing: when the traditional machine learning technology processes and classifies electroencephalogram data, the accuracy of pattern recognition depends on the obtained characteristic values, and the selection of the characteristic values is usually determined by the extraction method used and the personal experience of researchers. Sometimes, the selected characteristic value cannot accurately represent the inherent attribute of the electroencephalogram data, and the accuracy and timeliness of data processing are low.
4. Difficulty of diagnosis: the traditional method for diagnosing epilepsy mainly comprises four types of medical history data, physical examination, auxiliary examination and symptom examination. The four methods for diagnosing epilepsy require detailed medical history, detailed examination of the medical system, examination of the human and nervous system, and detailed and correct description of symptoms of family members.
5. And (3) updating the technology: the traditional epilepsy diagnosis methods such as medical history data, physical examination, auxiliary examination and symptom examination are old, the technical updating space is small, and the diagnosis of diseases is not facilitated.
Disclosure of Invention
The invention provides an epilepsia electroencephalogram detection system based on a feedforward pulse Neural network, which provides a supervised learning algorithm based on expected pulse sequence reconstruction through the supervised learning theory knowledge of the pulse Neural network (SNN), constructs an electroencephalogram (EEG) data classification model, and aims to provide an epilepsia diagnosis auxiliary function. At present, the abnormal electroencephalogram signals are manually monitored for a long time and then diagnosed, time and labor are consumed, medical resources are extruded, the system can well solve the defect by providing diagnosis assistance, and the development of artificial intelligence in the medical field is further promoted.
Therefore, the invention adopts the following technical scheme:
an epilepsia electroencephalogram detection system based on a feedforward pulse neural network comprises the following steps:
1) Collecting and establishing a data set, wherein experimental data of the data set is derived from a CHB-MIT scalp electroencephalogram database;
2) Constructing an electroencephalogram signal preprocessing module, namely an encoding module, and converting the electroencephalogram signals acquired in the step 1) into pulse sequences by adopting BSA (bovine serum albumin) encoding;
3) Constructing an impulse response model: constructing an impulse response model with an explicit analytic expression;
assuming that the spiking neuron has N input synapses, G is co-introduced through the ith synapse in the simulation interval i The number of pulses is such that,representing the firing time of the g-th pulse of a pre-synaptic neuron i, the internal state V (t) of this neuron at time t can be represented as follows:
ε (t) and ρ (t) are the impulse response function and refractory period function, w i Representing the weight of the ith synapse; wherein the impulse response function epsilon (t) represents the influence of the pre-synaptic neuron to fire the impulse on the post-synaptic neuron, and the formula is as follows:
where τ is a time constant that affects the morphology of ε (t);
the expression of the refractory period function ρ (t) is:
where θ is the firing threshold of the spiking neuron, τ R Is a delay constant that determines the nature of the refractory period function;
4) Constructing a multilayer feedforward pulse neural network (SNN): the multilayer feedforward pulse neural network SNN adopts a three-layer feedforward pulse neural network structure and comprises an input layer, a hidden layer and an output layer; the network is a fully-connected structure, namely each neuron is connected with all neurons of the next layer to form positive connection, and the synaptic weight of the positive connection between the hidden layer and the output layer is w oh ;
The SNN also comprises reverse connection, and the output layer neurons and all the neurons of the upper layer, namely the hidden layer, are connected to form the reverse connection; in the reverse connection, the input-to-output layer neuron pulse sequence isThe existence of synaptic weight between output layer neuron and hidden layer neuron is w oh ';
5) Training a multi-layer feedforward impulse neural network (SNN): inputting the pulse sequence subjected to BSA encoding in the step 2) into the multilayer feedforward pulse neural network constructed in the step 4) for pulse network learning;
6) Analyzing the classification result: when the SNN output results of two or more consecutive times are considered to have the epileptic seizure condition, judging that the patient enters the epileptic seizure state; in addition, when the front and rear samples which are continuous with the recording time of a certain sample are classified into the same class, the class of the sample is adjusted to be consistent with the front and rear samples; it can be considered that the recording time corresponding to the sample that is continuously determined as the epileptic seizure is more likely to be the actual epileptic seizure period.
Further, the detailed steps of BSA encoding in step 2) are as follows:
1) Estimated stimulation signal S est Expressed as:
wherein, t i Representing the firing time of the ith pulse of the neuron; h (t) represents the impulse response of the linear filter; x (t) represents a pulse sequence that a neuron fires, where x (t) can be written as:
2) Calculating two errors, wherein the BSA coding is based on finite impulse response, and the pulse sequence is converted into an analog signal by using an FIR reconstruction filter, and the conversion formula is as follows:
o (t) is analog information obtained by converting the pulse sequence, and the pulse distribution is calculated as follows: at each time τ the following two errors are calculated:
3) Judging whether to send out pulses or not; when error 1 ≤error 2 -threshold, where threshold is the threshold of the filter, set at 0.86, sending a pulse, while s (k + τ) ← s (k + τ) -h (k).
Further, in the step 5), the detailed steps of the pulse network learning are as follows:
1) Construction of a learning algorithm based on reconstruction of an expected pulse sequence
For a given input pulse sequence s i The actual output pulse sequence of the SNN is realized by adjusting the synaptic weight matrix WAnd a desired output pulse sequenceThe error of SNN at time t can be defined as:
in the formula N O The number of neurons in the SNN output layer; therefore, based on the pulse sequence inner product theory, the total error of SNN in the time interval Γ can be expressed as:
2) Synaptic weight adjustment using delta update rule
In SNN, a supervised learning mechanism based on gradient descent is used, namely, a gradient value E of an error function E to a synapse weight w is calculated, and synapse weight adjustment is carried out by using a delta update rule, wherein a neuron weight update formula is as follows:
e can be represented by the integral of the derivative of E (t) to the weight w over the time interval Γ, η being the learning rate;
3) Based on an SNN supervised learning algorithm of expected pulse sequence reconstruction, CHB-MIT epileptic brain data is normalized and BSA coded into 250 samples, the pulse sequences in a [0,1024ms ] interval obtained after coding are converted into a range of analog duration [0,256ms ], and 23 input pulse sequences are obtained after coding and conversion, so that 23 SNN input neurons are input.
Further, evaluating an epilepsia electroencephalogram data classification model based on the multilayer feedforward pulse neural network by adopting three indexes of Accuracy, sensitivity and Specificity; defining FP as the total number of non-epileptic seizure samples that were erroneously classified as seizure; FN is the total number of seizure samples that were mistakenly classified into non-epileptic seizure phases; TP and TN represent the total number of correctly classified epileptic seizure and non-seizure samples, respectively; the calculation formula of the three indexes is as follows:
in diagnostic tests, the higher the sensitivity, the lower the rate of missed diagnosis; the higher the specificity is, the lower the misdiagnosis rate is; the accuracy represents the percentage of correctly classified samples to the total number of samples in the data set, the sensitivity represents the capability of the multi-layer feedforward pulse neural network classification model of leak-proof epileptic seizure samples, and the specificity represents the capability of the model of not falsely judging normal electroencephalogram as epileptic seizure.
The main functions of the epilepsia electroencephalogram detection system comprise two parts, namely system maintenance and system platform operation, wherein a system maintenance module is mainly used for training an algorithm model based on a pulse neural network by using the existing CHB-MIT electroencephalogram data set to generate corresponding model parameters; the platform operation module mainly inputs the electroencephalogram data of the patient into the trained algorithm model, generates an auxiliary diagnosis result and provides reference for medical staff or the patient.
The system has the characteristics that:
1. and (4) predicting the epileptic disease. The doctor who is responsible for epileptic diagnosis reads patient's brain electrical data from the system, and the system can predict whether this patient has epileptic according to the brain electrical data of selecting to show the prediction result on user interface. In addition, the electroencephalogram data may need to be acquired from the patient for many times, so that the system is required to be capable of predicting multiple segments of electroencephalograms at the same time. The disease prediction operation is simple, and a user does not need to pay excessive attention to a specific prediction process of the model.
2. And (5) data persistence. The data required to be stored permanently in the system mainly comprises basic data of a user and a patient, electroencephalogram data of the patient and a diagnosis result predicted according to the electroencephalogram data. From the perspective of system safety and operation specifications, user information is stored in background data and is called by an administrator from a hospital personnel management system database. The user system contains a username and an initial login password. The basic data of the patient is managed and maintained by the hospital, and the user directly calls the patient information when using the system. The patient's electroencephalogram data is stored directly on the system by the technician responsible for taking the photograph. The predictive diagnosis result is stored by the doctor in charge of diagnosis. If the doctor approves the prediction result of the system, the prediction result can be directly stored, otherwise, the doctor can correct the prediction result and then store the result.
The invention has the following beneficial effects:
1. the invention constructs an efficient pulse neural network supervised learning algorithm, trains the pulse neural network to have the capability of detecting epileptic abnormal electroencephalogram signals, and has simple operation in the diagnosis process. The diagnosis is quick, the hands of medical personnel are greatly liberated to a certain extent, certain medical pressure is relieved, and certain medical resources are released; the supervised learning training of the impulse neural network in which the algorithm is actually applied to the model is provided, the epilepsy recognition accuracy rate of 95.38% is obtained through experiments, the sensitivity and specificity of the model are respectively 97.44% and 93.89%, and the diagnosis accuracy is greatly improved; the classification accuracy of 73.60% is obtained, and after the classification result is processed, the epilepsy recognition accuracy of the whole model can reach 95.38%, so that the accuracy and timeliness of data processing are greatly improved; the diagnosis of the system only needs to use the provided equipment to record the electroencephalogram and then upload the electroencephalogram for diagnosis, the whole process is short in time consumption, simple to operate and accurate in diagnosis, and the difficulty in diagnosing epileptic diseases is increased to a certain extent by the series of requirements.
2. The invention defines the error function of the pulse sequence and the learning rule of the synapse weight by establishing the inner product of the pulse sequence and the conversion relation thereof, and explores a network model suitable for electroencephalogram identification of epilepsy. In order to verify the performance of an algorithm reconstructed based on an expected pulse sequence, a three-layer feedforward pulse neural network structure is constructed, then an expected output pulse sequence of the pulse neural network is set, and a learning algorithm reconstructed based on the expected pulse sequence is used for adjusting the synaptic weights of the neural network. And judging that the learning algorithm framework reconstructed based on the expected pulse sequence has better learning capability according to the pulse sequence learning result and the change situation of the synaptic weight values before and after learning.
3. The invention designs an epilepsia EEG data classification model based on a pulse neural network, and uses a CHB-MIT data set collected in Boston children hospital to carry out experiments. The experiment utilizes the characteristic that the presentation form of the electroencephalogram data is very fit with the input mode of the pulse neural network, BSA (bovine serum albumin) coding is carried out on the electroencephalogram signals of each channel during sample preprocessing, the coded pulse sequence is used as the input of the pulse neural network, and the learning algorithm based on expected pulse sequence reconstruction is utilized to carry out supervised learning training on the pulse neural network, so that the classification result is obtained. For medical staff, the achievement of each stage of the project has great reference value. The whole research idea can not only provide reference for automatic analysis and detection of epilepsia electroencephalogram, but also be popularized to identification and diagnosis and treatment of similar neurological diseases.
Drawings
FIG. 1 is a schematic diagram of a neural information encoding and decoding process;
FIG. 2 is a schematic representation of the results of BSA encoding;
fig. 3 is a schematic diagram of a forward connection and an implicit reverse connection in SNN;
FIG. 4 is a schematic diagram of a SNN-based brain spatiotemporal data classification framework;
FIG. 5 is a graph of experimental results on a training set;
FIG. 6 is a schematic diagram of the recognition result of the SNN to the actual epileptic electroencephalogram data.
Detailed Description
The invention is further illustrated by the following specific examples and figures:
an epilepsia electroencephalogram detection system based on a feedforward pulse neural network comprises the following detailed steps:
1. step 1) collecting and establishing a data set, wherein experimental data of the data set is derived from a CHB-MIT scalp electroencephalogram database.
The experimental data of this example are derived from a CHB-MIT Scalp electroencephalogram Database (CHB-MIT scale EEG Database) and can be downloaded from a PhysioNet website. This database was collected from boston children's hospital, including 24 electroencephalographic recordings of 23 epileptic patients, where chb21 was EEG data generated 1.5 years after chb01 by the same female patient. Of these patients, 5 were male, between the ages of 3 and 22 years; 17 women between the ages of 1.5 and 19 years; the information for the chb24 patient is unknown. Before monitoring, they had discontinued antiepileptic drugs.
In the data set, all signals were sampled at 256Hz with a resolution of 16-bit. The Data is stored in the Format of an EDF (European Data Format) file and can be viewed using software such as an EDF browser. Most files contain 23 channels of electroencephalogram signals, which may be increased or decreased in some cases. These recordings were obtained using bipolar leads, with electrode positions following the 10-20 international standard lead system.
2. And 2) constructing an electroencephalogram signal preprocessing module-coding module, and converting the electroencephalogram signals collected in the step 1) into pulse sequences by adopting BSA (bovine serum albumin) coding.
Neural information encoding and decoding process as shown in fig. 1, when a neuron senses a stimulus, a stimulus signal is encoded into a specific pulse sequence through an encoding process as a response, and conversely, the pulse sequence data can be converted into an estimate of the stimulus signal through decoding.
In practice, almost all natural world signals are analog signals that are continuous in time. Therefore, these signals need to be converted into pulse sequences by an encoding module before the SNN is used to process the data. In this process, it is of utmost importance to ensure that errors and information loss caused by the conversion are minimized. Bens Spiker Algorithm (BSA) is an efficient Algorithm for converting analog signals into pulse sequences, the basic idea being to deconvolute the stimulus signal. Generally, the BSA encoding method is used to convert sound data into a pulse sequence, and since electroencephalogram data is also distributed in a frequency domain, BSA is also suitable for encoding EEG data.
BSA utilizes a stimulation reconstruction technique that is widely used in neural information coding, i.e., the stimulation of a biological neuron can be estimated by a pulse sequence filtered by a linear filter. Estimated stimulation signal s est Can be expressed as:
wherein, t i Representing the firing time of the ith pulse of the neuron; h (t) represents the impulse response of the linear filter; x (t) represents a pulse sequence that a neuron fires.
After the EEG signal is converted to a pulse sequence by BSA encoding, the original waveform can also be reconstructed from the encoded pulse sequence by a decoding process. This allows the coding scheme to be checked for validity and shows how well the coded pulse sequence mimics the original EEG waveform.
FIG. 2 (a) shows the original 4s FP1-F7 channel EEG fragment, for 1024 data points; FIG. 2 (b) shows a pulse sequence consisting of 0 and 1 obtained after BSA encoding; FIG. 2 (c) shows the pulse sequence with time interval [0,1024] in (b) being converted into time interval [0,256], the conversion formula is as follows:
wherein spike represents the Pulse emission time after the conversion to the time interval [0, simultime ], simultime is the set analog duration, and Total Pulse represents the Total number of 0 and 1 in the Pulse sequence before the conversion; in fig. 2 (d), the solid line is the normalized original EEG signal, and the dotted line represents the reconstructed analog signal after decoding, and it can be seen that the analog signal obtained after decoding has a high degree of matching with the original EEG signal, thus indicating that this coding method is effective.
3. Step 3), constructing an impulse response model: and constructing an impulse response model with an explicit analytic expression.
Assuming that the spiking neuron has N input synapses, G is co-introduced through the ith synapse in the simulation interval i The number of pulses is such that,representing the firing time of the g-th pulse of a pre-synaptic neuron i, the internal state V (t) of this neuron at time t can be represented as follows:
ε (t) and ρ (t) are the impulse response function and refractory period function, w i Representing the weight of the ith synapse; wherein the impulse response function epsilon (t) represents the influence of the pre-synaptic neuron to fire the impulse on the post-synaptic neuron, and the formula is as follows:
where τ is a time constant that affects the morphology of ε (t);
the expression of the refractory period function ρ (t) is:
where θ is the firing threshold of the spiking neuron, τ R Is a delay constant that determines the nature of the refractory period function.
4. And 4) constructing a multilayer feedforward pulse neural network (SNN).
In the biological nervous system, the signal generated and transmitted by neurons is a short electrical pulse, and this pattern is derived from the special electrochemical properties of the nerve cell membrane. The ion concentration inside and outside the nerve cell membrane is different, so that there is a potential difference between the inside and outside of the neuron, when the cell is stimulated, the ion permeability of the cell membrane and the potential difference between the inside and outside of the membrane are changed, and the cell generates excitation or inhibition phenomenon.
The topological structure of the SNN can be classified into three types, namely, a feedforward type impulse neural network, a recursive type impulse neural network, and a hybrid type impulse neural network. The feedforward type network adopts a one-way structure, neurons of which are arranged in layers, and each layer of neurons receives an input of a previous layer and generates an output to be transmitted to a next layer. There is no connection between multiple neurons of each layer. In general, the first layer is called the input layer of the neural network, and this layer does not contain the calculation of neurons and only represents the network input; the last layer is called output layer and is responsible for providing data analysis results. As shown in fig. 1, in a multi-layer network, there may be one or more hidden layers in between the input and output layers.
To simplify the description of the algorithm derivation process, a three-layer feedforward impulse neural network structure is used herein, comprising an input layer, a hidden layer and an output layer; the network is fully connected, namely, each neuron is connected with all the neurons in the next layer; no multi-synaptic connection approach is used. For a feedforward impulse neural network with multiple hidden layers, the derivation process of the learning rule is similar.
Pulse trains can be used as abstract representations of neural activity, and thus SNNs employ pulse trains to encode, process, and convey information. In a specific simulation time interval Γ = [0,T ]]In, pulse train s = { t = f F =1,., N } represents an ordered sequence of N discrete pulse firing instants, described by the formula:wherein t is f Represents the release time of the f-th pulse of the neuron, and delta (x) isDirac delta function, δ (x) =1 when x =0, otherwise δ (x) =0.
In SNN, supervised learning mainly involves: input pulse sequence s i Actual output pulse trainAnd a desired output pulse sequenceFor any pulse neuron, the presynaptic neuron input is assumed to be s i Its own actual output isThere is a linear combination relationship between multiple inputs and their own actual outputs:
wherein, w oi Denotes the synaptic weight, i.e. the strength of the connection between the pre-synaptic neuron i and the post-synaptic neuron o. N represents the total number of presynaptic neurons connected to the postsynaptic neuron o.
The SNN also comprises reverse connection, and the output layer neurons and all the neurons of the upper layer, namely the hidden layer, are connected to form the reverse connection; in the reverse connection, the input-to-output layer neuron pulse sequence isThe existence of synaptic weight between output layer neuron and hidden layer neuron is w oh '。
5. And 5) training the multilayer feedforward pulse neural network SNN.
The method comprises the steps of constructing an SNN supervised learning algorithm based on expected pulse sequence reconstruction, and defining an error function of a pulse sequence and a learning rule of a synapse weight. The aim being to give a given input pulse sequence s i The actual output pulse sequence of the SNN is realized by adjusting the synaptic weight matrix WAnd a desired output pulse sequenceThe error of SNN at time t can be defined as:
in the formula N O The number of SNN output layer neurons is shown. Therefore, based on the pulse sequence inner product theory, the total error of SNN in the time interval Γ can be expressed as:
similar to the BP algorithm of the conventional ANN, in SNN, a supervised learning mechanism based on gradient descent can also be used, that is, a gradient value ∑ E of the error function E for the synapse weight w is calculated, and then a delta update rule is used to perform synapse weight adjustment, where the expression is as follows:
e can be represented by the integral of the derivative of E (t) to the weight w over the time interval Γ. Eta is the learning rate.
It can be assumed that in the SNN network structure, besides the existence of the forward connection, there is an implicit reverse connection, in which the synapse weight is also used, but is not related to the explicit forward network synapse weight. Thereby providing a desired pulse sequence for output layer neurons from a spiking neural networkReverse reconstruction of hidden layer neuron expected outputThe method of (1). In FIG. 3, the dashed lines indicate the existence of a synaptic weight of w between SNN output layer neurons and hidden layer neurons oh ' reverse connection; for convenience of illustration, only one forward connection (solid line) is drawn between them, and synaptic weights are used for w oh However, in practice, both the positive and negative connections in the network use the full connection scheme.
6. And 6) analyzing the classification result.
EEG Data belongs to space-time mode Data (Spatio-and Spectro-Temporal Data) and has the characteristics of multi-scale, multi-dimension and dynamic correlation. In general, conventional ANN focuses on handling real-valued inputs and outputs, and often loses time information contained in large amounts of data. Current research shows that SNN has significant advantages in learning target pulse delivery times and capturing multi-level features from spatio-temporal pattern data.
Fig. 4 shows a classification framework based on SNN adopted in the CHB-MIT epilepsy electroencephalogram classification experiment, which includes four modules, namely, a data acquisition module, an encoding module, a neural network module and a result processing module. The first part is that the working result of acquiring electroencephalograms of epileptics is provided by a CHB-MIT database, and screened EEG is segmented into small samples by taking 4s duration as a unit for convenient processing; the second part of the coding module uses a Bens spike algorithm to convert continuous value input data into a pulse sequence, the conversion is effective, and the process is fast and accurate; the third part, the pulse sequence obtained after coding is used as the input of the SNN, and at the moment, the SNN supervised learning algorithm based on the expected pulse sequence reconstruction and proposed in the text is used for training a pulse neural network to realize the complex space-time pattern recognition, so that the learning effect of the algorithm is verified; and in the last part, processing SNN output to obtain a final classification result.
It can be seen that the classification model, whether it is a data coding module or supervised learning of the impulse neural network, adopts a threshold-based method and is spread around the impulse sequence, which is very similar to the biological neural activity. When BSA coding is carried out on input space-time data, when two calculated error values exceed a threshold value, a pulse is generated at the moment; activation of neurons in the SNN also requires that the membrane potential change caused by the received pulse reach a fixed threshold. In addition, in the supervised learning process of SNN, the goal is to have neurons produce output pulses at specified times. Therefore, it is necessary to set a desired output of the network. In the experiments in this chapter, it is desirable to set in the form of a specified pulse sequence.
And verifying the effectiveness of the SNN supervised learning algorithm training pulse neural network based on the expected pulse sequence reconstruction and the solving capability of the pulse neural network on the brain space-time mode data identification problem through a CHB-MIT epilepsia electroencephalogram classification experiment.
First, 250 samples obtained by processing were processed in a manner of 2:3 into training and test sets and then using multichannel EEG for automatic classification recognition. Thus, 23 EEG data samples were normalized and BSA encoded, and the pulse sequences in the [0,1024ms ] interval obtained after encoding were transformed into a range of analog durations [0,256ms ]. Since 23 input pulse sequences are obtained after encoding and conversion, 23 SNN input neurons are provided in total. Three layers of feedforward SNN are selected in the experiment, and the training iteration times are 20. The rest network structures and parameters are set as follows: 60 hidden layer neurons and 1 output layer neuron; learning rate 0.000001; weight range [0,0.5]; when the time constant is adjusted to control the first iteration, the number of actual output pulses of the network is not too large; the expected pulse sequence is set as a specified pulse sequence, the class A sends pulses in 135ms, the class B sends pulses in 140ms, and the class C sends pulses in 130ms; and respectively calculating error values of the SNN actual output and the three types of expected errors, wherein when the error is minimum, the type corresponding to the expected pulse is the identification type.
Fig. 5 shows the experimental results of the algorithm proposed in this example on a training set of 100 samples. It can be seen that as the number of iterations increases, the classification accuracy of the learning algorithm based on the expected pulse sequence reconstruction improves. The algorithm reaches a maximum accuracy of 81% at iteration 18. For different types of samples, the SNN can accurately identify non-epileptic seizure EEG data (class A samples) and epileptic seizure EEG data (class C samples), while the classification accuracy of class B samples is low, and one very probable reason is that electroencephalograms at the beginning of epilepsia include both normal electroencephalogram characteristics and epileptic electroencephalogram characteristics. The test set contained a total of 150 EEG samples, 49 of which were of class a, 47 of which were of class B and 54 of which were of class C. As can be seen from the graph, the algorithm can accurately distinguish non-epileptic seizure periods from epileptic seizure periods, a small part of epileptic brain electricity can be mistaken for normal, but the algorithm can hardly recognize EEG samples at the beginning of epileptic seizure, and about half of B samples are mistaken for A samples.
FIG. 6 shows the recognition result of the classification model of epileptic brain electrical data in the text on the electroencephalogram record of a segment of 800s of patient CHB01 in the database of CHB-MIT. Fig. 6 (a) shows the actual seizure time of the patient in red. Fig. 6 (b) shows the classification result determined directly from the actual output of the SNN in the model. Since the classification accuracy of the neural network is difficult to reach one hundred percent, more false judgments exist. However, in reality, it is not possible for a patient to experience multiple seizures over a period of minutes or even seconds. Therefore, in order to improve the classification accuracy of the model, a determination method is adopted in which the patient is determined to enter the epileptic seizure state when two consecutive SNN outputs are considered to have the epileptic seizure state. In addition, when the front and rear samples that are time-continuous with a certain sample are classified into the same class, the class of the sample is adjusted to be consistent with the front and rear samples. The recognition result after the above strategy processing is shown in fig. 6 (c). It can be considered that the recording time corresponding to the sample that is continuously judged as the epileptic seizure is more likely to be the actual epileptic seizure period. Short-time seizure determination results indicate the possibility of seizures.
Similar to the existing research, the SNN-based epilepsia electroencephalogram data classification model is also evaluated by adopting three indexes of Accuracy (Accuracy), sensitivity (Sensitivity) and Specificity (Specificity). Defining FP as the total number of non-epileptic seizure samples that were erroneously classified as seizure phase; FN is the total number of seizure samples that were mistakenly classified into non-epileptic seizure phases; TP and TN represent the total number of correctly classified epileptic and non-seizure samples, respectively. The calculation formula of the three indexes is as follows:
in diagnostic tests, the higher the sensitivity, the lower the rate of missed diagnosis; the higher the specificity, the lower the misdiagnosis rate. Therefore, in this study, the accuracy represents the percentage of correctly classified samples to the total number of samples in the data set, the sensitivity represents the ability of the SNN classification model to fail to detect seizure samples, and the specificity represents the ability of the model to fail to falsely identify normal electroencephalogram as seizure.
When the SNN output results of two or more consecutive times are considered to have the epileptic seizure condition, judging that the patient enters the epileptic seizure state; in addition, when the front and rear samples which are continuous with the recording time of a certain sample are classified into the same class, the class of the sample is adjusted to be consistent with the front and rear samples; it can be considered that the recording time corresponding to the sample that is continuously determined as the epileptic seizure is more likely to be the actual epileptic seizure period.
Claims (4)
1. An epilepsia electroencephalogram detection system based on a feedforward pulse neural network is characterized by comprising the following steps:
1) Collecting and establishing a data set, wherein experimental data of the data set is derived from a CHB-MIT scalp electroencephalogram database;
2) Constructing an electroencephalogram signal preprocessing module, namely an encoding module, and converting the electroencephalogram signals acquired in the step 1) into pulse sequences by adopting BSA (bovine serum albumin) encoding;
3) Constructing an impulse response model: constructing an impulse response model with an explicit analytic expression;
assuming a spiking neuron has N input synapses, G is co-afferent through the ith synapse in a simulation time interval i The number of pulses is such that,representing the firing time of the g-th pulse of a pre-synaptic neuron i, the internal state V (t) of this neuron at time t can be represented as follows:
ε (t) and ρ (t) are the impulse response function and refractory period function, w i Representing the weight of the ith synapse; wherein the impulse response function epsilon (t) represents the influence of the pre-synaptic neuron to fire the impulse on the post-synaptic neuron, and the formula is as follows:
where τ is a time constant that affects the morphology of ε (t);
the expression of the refractory period function ρ (t) is:
where θ is the firing threshold of the spiking neuron, τ R Is a delay constant that determines the nature of the refractory period function;
4) Constructing a multilayer feedforward pulse neural network (SNN): the multilayer feedforward pulse neural network SNN adopts a three-layer feedforward pulse neural network structure and comprises an input layer, a hidden layer and an output layer; the network is a fully-connected structure, namely each neuron is connected with all neurons of the next layer to form positive connection, and the synaptic weight of the positive connection between the hidden layer and the output layer is w oh ;
SNN further comprisesReverse connection is formed by constructing connections between the neurons of the output layer and all the neurons of the upper layer, namely the hidden layer; in the reverse connection, the input-to-output layer neuron pulse sequence isThe existence of synaptic weight between output layer neuron and hidden layer neuron is w oh ';
5) Training a multi-layer feedforward impulse neural network (SNN): inputting the pulse sequence subjected to BSA encoding in the step 2) into the multilayer feedforward pulse neural network constructed in the step 4) for pulse network learning;
6) Analyzing the classification result: when the SNN output results of two or more consecutive times are considered to have the epileptic seizure condition, judging that the patient enters the epileptic seizure state; in addition, when the front and rear samples which are continuous with the recording time of a certain sample are classified into the same class, the class of the sample is adjusted to be consistent with the front and rear samples; it can be considered that the recording time corresponding to the sample that is continuously determined as the epileptic seizure is more likely to be the actual epileptic seizure period.
2. The system for detecting the epilepsia electroencephalogram based on the feed-forward pulse neural network, which is characterized in that the BSA encoding in the step 2) comprises the following detailed steps:
1) Estimated stimulation signal S est Expressed as:
wherein, t i Representing the firing time of the ith pulse of the neuron; h (t) represents the impulse response of the linear filter; x (t) represents a pulse sequence that a neuron fires, where x (t) can be written as:
2) Calculating two errors, wherein the BSA code is based on finite impulse response, and the FIR reconstruction filter is used for converting the impulse sequence into an analog signal, and the conversion formula is as follows:
o (t) is analog information obtained by converting the pulse sequence, and the pulse distribution is calculated as follows: at each time τ the following two errors are calculated:
3) Judging whether to send out pulses or not; when error 1 ≤error 2 -threshold, where threshold is the threshold of the filter, set at 0.86, sending a pulse, while s (k + τ) ← s (k + τ) -h (k).
3. The epilepsia electroencephalogram detection system based on the feed-forward pulse neural network as claimed in claim 1, wherein in the step 5), the detailed steps of pulse network learning are as follows:
1) Construction of a learning algorithm based on reconstruction of an expected pulse sequence
For a given input pulse sequence s i The actual output pulse sequence of the SNN is realized by adjusting the synaptic weight matrix WAnd a desired output pulse sequenceThe error of SNN at time t can be defined as:
In the formula N O The number of neurons in the SNN output layer; therefore, based on the pulse sequence inner product theory, the total error of SNN in the time interval Γ can be expressed as:
2) Synaptic weight adjustment using delta update rule
In SNN, a supervised learning mechanism based on gradient descent is used, i.e. the gradient value of the error function E to the synaptic weight w is calculatedAnd then adjusting the synaptic weights by using a delta updating rule, wherein the neuron weight updating formula is as follows:
can be represented by the integral of the derivative of E (t) to the weight w in a time interval Γ, η is the learning rate;
3) Based on an SNN supervised learning algorithm of expected pulse sequence reconstruction, CHB-MIT epileptic brain data is normalized and BSA coded into 250 samples, the pulse sequences in a [0,1024ms ] interval obtained after coding are converted into a range of analog duration [0,256ms ], and 23 input pulse sequences are obtained after coding and conversion, so that 23 SNN input neurons are input.
4. The epilepsy electroencephalogram detection system based on the feedforward pulse neural network as claimed in claim 1, wherein three indexes of Accuracy, sensitivity and Specificity are adopted to evaluate the epilepsy electroencephalogram data classification model based on the multilayer feedforward pulse neural network; defining FP as the total number of non-epileptic seizure samples that were erroneously classified as seizure; FN is the total number of seizure samples that were mistakenly classified into non-epileptic seizure phases; TP and TN represent correctly classified total number of epileptic seizure and non-seizure samples, respectively; the calculation formula of the three indexes is as follows:
in diagnostic tests, the higher the sensitivity, the lower the missed diagnosis rate; the higher the specificity is, the lower the misdiagnosis rate is; the accuracy represents the percentage of correctly classified samples to the total number of samples in the data set, the sensitivity represents the capability of the multi-layer feedforward pulse neural network classification model of leak-proof epileptic seizure samples, and the specificity represents the capability of the model of not falsely judging normal electroencephalogram as epileptic seizure.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211178565.XA CN115414054A (en) | 2022-09-27 | 2022-09-27 | Epilepsia electroencephalogram detection system based on feedforward pulse neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202211178565.XA CN115414054A (en) | 2022-09-27 | 2022-09-27 | Epilepsia electroencephalogram detection system based on feedforward pulse neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN115414054A true CN115414054A (en) | 2022-12-02 |
Family
ID=84205867
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202211178565.XA Pending CN115414054A (en) | 2022-09-27 | 2022-09-27 | Epilepsia electroencephalogram detection system based on feedforward pulse neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN115414054A (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117493955A (en) * | 2023-11-01 | 2024-02-02 | 北京大学 | Training method of electroencephalogram signal classification model of epileptic patient |
-
2022
- 2022-09-27 CN CN202211178565.XA patent/CN115414054A/en active Pending
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117493955A (en) * | 2023-11-01 | 2024-02-02 | 北京大学 | Training method of electroencephalogram signal classification model of epileptic patient |
CN117493955B (en) * | 2023-11-01 | 2024-06-07 | 北京大学 | Training method of electroencephalogram signal classification model of epileptic patient |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US20200337580A1 (en) | Time series data learning and analysis method using artificial intelligence | |
CN110070105B (en) | Electroencephalogram emotion recognition method and system based on meta-learning example rapid screening | |
Buscema et al. | The IFAST model allows the prediction of conversion to Alzheimer disease in patients with mild cognitive impairment with high degree of accuracy | |
Tokuno et al. | Usage of emotion recognition in military health care | |
CN113786204A (en) | Epilepsia intracranial electroencephalogram early warning method based on deep convolution attention network | |
O’Shea et al. | Deep learning for EEG seizure detection in preterm infants | |
CN109106384B (en) | Psychological stress condition prediction method and system | |
Huang et al. | Automatic epileptic seizure detection via attention-based CNN-BiRNN | |
Li et al. | Patient-specific seizure prediction from electroencephalogram signal via multichannel feedback capsule network | |
Abdelhameed et al. | An efficient deep learning system for epileptic seizure prediction | |
Liang et al. | Obstructive sleep apnea detection using combination of CNN and LSTM techniques | |
CN112244868A (en) | Epilepsia electroencephalogram signal classification method based on ANFIS | |
Jana et al. | Epileptic seizure prediction from EEG signals using DenseNet | |
CN113974655A (en) | Epileptic seizure prediction method based on electroencephalogram signals | |
Mouleeshuwarapprabu et al. | Nonlinear vector decomposed neural network based EEG signal feature extraction and detection of seizure | |
Wan et al. | Automatic detection of HFOs based on singular value decomposition and improved fuzzy c-means clustering for localization of seizure onset zones | |
CN115414054A (en) | Epilepsia electroencephalogram detection system based on feedforward pulse neural network | |
Taşkıran et al. | A deep learning based decision support system for diagnosis of Temporomandibular joint disorder | |
Qin et al. | Patient-specific seizure prediction with scalp EEG using convolutional neural network and extreme learning machine | |
CN111613338A (en) | Method and system for constructing spike-slow complex wave detection model | |
CN111248907A (en) | Risk prediction method based on electroencephalogram signal characteristics of mental disease clinical high-risk group | |
TWI808579B (en) | Automatic evolution method for brain wave database and automatic evolving system for detecting brain wave | |
Al-hajjar et al. | Epileptic seizure detection using feature importance and ML classifiers | |
CN115530754A (en) | Epilepsy early warning method and device based on deep learning | |
Hermawan et al. | A multi representation deep learning approach for epileptic seizure detection |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |