CN-116304575-B - Multi-task epileptic electroencephalogram automatic detection and prediction system

Abstract

The design is a multitasking epileptic electroencephalogram automatic detection and prediction system. The system comprises a preprocessing module, a detection and prediction module, a prediction module and a prediction module, wherein the preprocessing module is used for receiving an electroencephalogram signal to be recognized of an epileptic patient or an electroencephalogram signal of the acquired patient for the training module, denoising, filtering, segmenting, normalizing and the signals and outputting the processed electroencephalogram signal of the patient for the training module to the training module, the training module is used for training and testing the acquired electroencephalogram signal of the patient for the training module, the characteristic extraction is carried out through a multi-scale convolution network, the detection and prediction branch added with a dual-attention mechanism is trained, and the trained model parameters are stored for the detection and prediction module, and the detection and prediction result is given to the electroencephalogram signal to be recognized of the epileptic patient in the detection and prediction module. The invention adopts a modularized design, realizes the real-time detection and prediction of the unbalanced brain electrical signal with noise, and provides a reliable system for practical application.

Inventors

• LI BING
• Yu Shangning
• LIU XIA

Assignees

• 哈尔滨理工大学

Dates

Publication Date

20260505

Application Date

20230405

Claims (1)

1. 1. The system is characterized by comprising a signal preprocessing module, a training module and a detection and prediction module; The model preprocessing module is used for receiving relevant electroencephalogram signals input by a user, processing the signals and outputting the processed information to the training module, and comprises the following steps: s1.1, classifying epileptic brain electrical signals into five types, namely epileptic seizure stage, epileptic seizure pre-stage 1, epileptic seizure pre-stage 2, epileptic seizure pre-stage 3 and epileptic seizure interval; S1.2, using a wavelet function to process EEG, removing the interference of electro-oculogram, electrocardio and myoelectricity noise, using a band-stop filter to filter the power frequency interference of 57-63Hz and 117-123Hz in the electroencephalogram data, and performing standardization processing on the data by using dispersion standardization; S1.3, dividing an epileptic electroencephalogram signal into segments of 2S, wherein the segments in the epileptic seizure period have 75% overlapping parts, and the segments divided in other periods have no overlapping parts; S1.4, extracting time-frequency characteristics of each piece of data by adopting a continuous wavelet transformation and short-time Fourier transformation fusion method, carrying out splicing fusion on a time-frequency matrix generated by the two methods after resampling, and transmitting a characteristic diagram to a prediction network in a training module; The training module is used for training the electroencephalogram signals after the preprocessing module through a neural network and storing trained model parameters, and is used for detecting and predicting the module, and the training module comprises the following steps: s2.1, performing feature extraction by utilizing a multi-scale convolution network formed by multi-scale convolution blocks, performing pooling operation by using a K maximum pooling layer to replace a maximum pooling layer, and transmitting the extracted information to a detection network by adopting a double-attention module after each step of multi-scale convolution block; S2.2, a double-attention module respectively aggregates the input features along the horizontal and vertical directions into two independent feature graphs by utilizing the maximum pooling and average pooling operation, the two feature graphs respectively capture the relationship between the input feature graphs and time and the relationship between the channels along the time and channel directions, and then fusion of the feature graphs is carried out; S2.3, detecting epileptic brain electrical signals by using stacked residual networks, wherein each stacked residual network consists of three residual blocks, each residual block sequentially comprises point-by-point convolution (Pointwise Conv d), batch Normalization (BN) processing, depth separable convolution (DEPTHWISE CONV d), BN processing, pointwise Conv d, BN processing and a correction linear unit; S2.4, predicting epileptic brain electrical signals by utilizing a Bi-directional long-short-term memory (Bi-LSTM) network, and adding a double-attention module at the last layer of the network, wherein the double-attention module consists of time attention and channel attention, and the time attention module and the channel attention module adopt a parallel mode; S2.5, training the detection branch network and the prediction branch network respectively, obtaining a loss function (GHMLoss) of the detection branch network and a loss function (CrossEntropyLoss) of the prediction branch network, calculating through the two branch loss functions to obtain a total loss function, and training for preset iteration times to complete model establishment; the detection and prediction module is used for giving out a detection and prediction result to the acquired brain electrical signals of the patient.

Description

Multi-task epileptic electroencephalogram automatic detection and prediction system Technical Field The invention relates to the technical field of electroencephalogram signal processing, in particular to an epileptic electroencephalogram detection and prediction system based on deep learning, and specifically relates to a multi-task epileptic electroencephalogram automatic detection and prediction system. Background Epilepsy is a neurological disorder that transiently affects brain function, resulting from sudden abnormal discharges of brain neurons, which continue to repeat attacks without inducement. The number of epileptics in China is about 900 ten thousand, and about 40 ten thousand people are newly added each year, wherein teenagers under 18 years old account for 2/3. After the correct and economical treatment with the drug, about 70% of patients can avoid attacks. Therefore, in preventing and treating epilepsy, it is important to diagnose the disease accurately. Under the current situation, the diagnosis of the illness state mainly depends on the doctor with abundant experience to observe the visual signals such as the waveform, the amplitude and the like of the electroencephalogram of the patient. However, diagnosis of the disease by only a doctor is inefficient, and erroneous judgment occurs in long-term diagnosis. In order to rapidly diagnose epilepsy, timely treatment reduces seizure risk. The neural network is used for detecting and predicting the epileptic brain electrical signals, automatically detecting the period of the epileptic, and predicting the seizure of the epileptic in the future period. Therefore, the designed network is particularly important to detect and predict more quickly and accurately. Compared with the traditional neural network, the convolutional neural network can automatically extract the characteristics, has higher running speed, and has better effect in the detection and prediction of the brain electrical signals. The attention mechanism can enable the model to keep more space, time and main characteristic information in the channel, so that the accuracy of the model is improved. Disclosure of Invention The invention aims to provide a multi-task epileptic electroencephalogram automatic detection and prediction system, which solves the problems that the existing epileptic electroencephalogram system is affected by noise, power frequency interference, unbalanced electroencephalogram data, slower in epileptic period detection speed and lower in epileptic seizure prediction accuracy rate, and can provide reliable references for practical engineering application. In order to achieve the purpose, the invention provides the technical scheme that the system for detecting and predicting the multi-task epileptic brain electrical signals is high in detection and prediction accuracy and short in time, and comprises a signal preprocessing module, a training module and a detection and prediction module. The model preprocessing module is used for receiving relevant electroencephalogram signals input by a user and processing the signals, and comprises the steps of receiving data, dividing electroencephalogram signal periods, denoising the signals, filtering the signals, carrying out standardization processing on the signals, segmenting the signals, extracting spectrograms of the signals, and outputting processed information to the training module, wherein the preprocessing module comprises the following steps. S1.1 acquiring electroencephalogram (EEG) data of epileptic patients under sEEG monitoring, wherein the EEG data is derived from an EEG of 24 refractory epileptic seizures of 23 children patients in Boston childhood Hospital. S1.2, the epileptic brain electrical signals are divided into five types of epileptic seizure stage, epileptic seizure pre-stage 1, epileptic seizure pre-stage 2, epileptic seizure pre-stage 3 and epileptic seizure interval. And S1.3, using a wavelet function to process EEG, removing the interference of electro-oculogram, electrocardio and myoelectricity noise, using a band-stop filter to filter the power frequency interference of 57-63Hz and 117-123Hz in the electroencephalogram data, and performing normalized processing on the data by using dispersion normalization. S1.4, dividing the epileptic brain electrical signals into segments of 2S, wherein the segments in the epileptic seizure period have 75% overlapping parts, and the segments divided in other periods have no overlapping parts. And S1.5, extracting the time-frequency characteristics of each segment of data by adopting a continuous wavelet transformation and short-time Fourier transformation fusion method, carrying out resampling on a time-frequency matrix generated by the two methods, then carrying out splicing fusion to obtain richer image characteristics, and transmitting the characteristic images to a prediction network in a training module. The training module is used for training the electr