阅读说明：本技术 发作间期痫样放电的检测方法、存储介质和设备 (Method for detecting seizure-like discharge between attacks, storage medium and device ) 是由 张冀聪 赵晓慧 韦博轩 于 2020-12-22 设计创作，主要内容包括：提供了一种癫痫患者发作间期痫样放电的检测方法、设备和存储介质,该检测方法包括：对预处理后的所述脑电信号进行分割,生成m个w*256的输入样本矩阵S～((m))；根据输入的m个w*256的输入样本矩阵S～((m))输出痫样电活动检测结果。利用痫样电活动在同时刻各传感器空间的第一关联性以及伪差在同时刻各传感器空间的偶发性和不同于第一关联性的第二关联性识别痫样电活动相关的特征,能够过滤无法通过滤波手段滤除的伪差,无需人工设计伪差滤除单元及特征工程即可完成自动检测。(A method, a device and a storage medium for detecting an inter-seizure epileptiform discharge of an epileptic patient are provided, the detection method comprising: segmenting the preprocessed electroencephalogram signals to generate m w-256 input sample matrixes S (m) (ii) a Input sample matrix S of m w 256 according to input (m) And outputting the epilepsy-like electrical activity detection result. The first relevance of the epilepsy-like electrical activity in the sensor spaces at the same time, the sporadic nature of the artifact in the sensor spaces at the same time and the second relevance different from the first relevance are utilized to identify the relevant characteristics of the epilepsy-like electrical activity, the artifact which cannot be filtered by a filtering means can be filtered, and automatic detection can be completed without manually designing an artifact filtering unit and a characteristic project.)

1. A detection method of epileptic discharge in seizure intervals of epileptic patients based on feature analysis of space electroencephalogram signals of a simultaneous sensor is characterized by comprising the following steps:

m w 256 input sample matrixes S through U-shaped semantic segmentation structural units^(m)Mapping and outputting as feature vector F^(m)＝[F1,F2,…,Fw]^T(ii) a The method comprises the following steps that Fw is a feature obtained after electroencephalogram signals of a w-th sensor channel are mapped through a U-shaped semantic segmentation structure, the U-shaped semantic segmentation structure unit comprises 4 superposed maximum pooling layers and 4 superposed deconvolution structures, two layers of 1 × 3 convolution layers are superposed in front of each maximum pooling layer, two layers of 1 × 3 convolution layers are superposed behind each deconvolution structure, and 1 × 1 convolution layers are superposed behind two layers of 1 × 3 convolution layers corresponding to the deconvolution structure on the uppermost layer;

according to the feature vector F^(m)＝[F1,F2,…,Fw]^TAnalyzing the characteristics of electroencephalogram signals of all sensor spaces at the same time, identifying characteristics related to epileptic discharge by utilizing the first relevance of epileptic discharge in all sensor spaces at the same time, the sporadic nature of artifact in all sensor spaces at the same time and the second relevance different from the first relevance, and further judging the S of a 1S time window^(m)Whether the segments contain epileptic-like discharge or not is judged, and the results of the m segments are summarized to realize the detection of the epileptic-like discharge in the whole segment of electroencephalogram signals;

wherein the m w 256 input sample matrices S^(m)＝[S1,S2,…,Sw]^TW is the number of sensor channels of the electroencephalogram signal, and Sw is a vector formed by single-channel electroencephalogram signal segments with the duration of 1 s.

2. The detection method according to claim 1, wherein the characteristics of the electroencephalogram signals of w sensor spaces at the same time are input through a gate control cycle unit, the input is decided in a chain form of a repeating network module, and a decision result corresponding to each time is output;

judging S of the 1S time window by utilizing a sigmod full-connection network according to the corresponding decision result of each moment^(m)And (4) whether the segments contain epileptic-like discharge or not, summarizing the results of the m segments, and realizing the detection of the epileptic-like discharge in the whole segment of electroencephalogram signals.

3. The method of claim 1 or 2, wherein a time window with a window width of 1S and an overlap of 50% is selected, the preprocessed electroencephalogram signal is segmented, and m w × 256 input sample matrices S are generated^(m)。

4. A detection method of epileptic discharge in seizure intervals of epileptic patients based on feature analysis of space electroencephalogram signals of a simultaneous sensor is characterized by comprising the following steps:

collecting an electroencephalogram signal of a patient, wherein the electroencephalogram signal is a multi-lead scalp electroencephalogram signal;

filtering and normalizing the electroencephalogram signals;

segmenting the preprocessed electroencephalogram signals to generate m w-256 input sample matrixes S^(m)＝[S1,S2,…,Sw]^TW is the number of sensor channels of the electroencephalogram signal, and Sw is a vector formed by single-channel electroencephalogram signal segments with the duration of 1 s;

inputting m w 256 input sample matrixes S^(m)Classifying and outputting epilepsy discharge detection results;

wherein the input sample matrix S of m w 256 to be input^(m)The classification and output of epileptiform discharge detection results comprises:

inputting m w-256 input sample matrixes S through U-shaped semantic segmentation structure units^(m)Mapping and outputting as feature vector F^(m)＝[F1,F2,…,Fw]^T(ii) a The method comprises the following steps that Fw is a feature obtained after electroencephalogram signals of a w-th sensor channel are mapped through a U-shaped semantic segmentation structure, the U-shaped semantic segmentation structure unit comprises 4 superposed maximum pooling layers and 4 superposed deconvolution structures, two layers of 1 × 3 convolution layers are superposed in front of each maximum pooling layer, two layers of 1 × 3 convolution layers are superposed behind each deconvolution structure, and 1 × 1 convolution layers are superposed behind two layers of 1 × 3 convolution layers corresponding to the deconvolution structure on the uppermost layer;

according to the feature vector F^(m)＝[F1,F2,…,Fw]^TAnalyzing the characteristics of electroencephalogram signals of all sensor spaces at the same time, identifying characteristics related to epileptic discharge by utilizing the first relevance of epileptic discharge in all sensor spaces at the same time, the sporadic nature of artifact in all sensor spaces at the same time and the second relevance different from the first relevance, and further judging the S of a 1S time window^(m)And (4) whether the segments contain epileptic-like discharge or not, summarizing the results of the m segments, and realizing the detection of the epileptic-like discharge in the whole segment of electroencephalogram signals.

5. A computer-readable storage medium, in which a computer program is stored which is adapted to be loaded and executed by a processor to cause a computer device having said processor to carry out the method of any one of claims 1 to 4.

6. A computer device, comprising: a processor and a memory;

the processor is coupled to a memory, wherein the memory is configured to store a computer program, and the processor is configured to invoke the computer program to cause the computer device to perform the method of any of claims 1-4.

Technical Field

The invention relates to the field of biomedical detection, in particular to a device and a method for detecting seizure-like electrical activity of an epileptic in an inter-seizure period, and especially relates to a device and a method for detecting seizure-like electrical activity of an epileptic in an inter-seizure period based on feature analysis of a space electroencephalogram signal of a simultaneous sensor.

Background

Scalp electroencephalography (sEEG) is a non-invasive signal acquisition method for clinical recording of brain-related activities. It has higher time resolution and plays an important role in the detection, diagnosis, treatment, curative effect evaluation and pathological research of epilepsy. The current diagnosis of epilepsy based on scalp electroencephalogram is mainly the detection of epileptic seizures, however, epileptic seizures are not common. In contrast, nearly 80% of epileptic patients have neuronal abnormal Discharges during the inter-seizure period, and these abnormal Discharges observed in sEEG are called inter-seizure like Discharges (IEDs). IED is powerful support for epilepsy diagnosis or risk assessment of epileptic seizures, and is also a key factor for determining potential focus of epileptic origin. Thus, IEDs detection based on the sog et is of great importance for the diagnosis of epilepsy.

Unlike epileptic seizures, IEDs appear more frequently and in various forms in scalp electroencephalogram signals, such as spikes, spike-slow complexes, multi-spike complexes, etc., which are present in the sog waveform for varying lengths of time (only 20-70ms short, long lasting for more than 1 s), and are easily confused with artifacts (caused by chewing, sensor slippage, heartbeat, etc.) in the sog waveform that cannot be filtered out by filtering means, making detection of IEDs in the sog waveform extremely challenging.

Currently, the detection of IEDs in long-range eeg still needs manual labeling by experienced technicians and doctors. Finding anomalous segments in milliseconds from a large signal, up to hours and even more than a day, is time consuming for the clinician and its accuracy is inevitably subject to large deviations due to the individual's subjective experience and fatigue.

Over the past decades, methods have been proposed to automatically describe and detect IEDs by measuring the similarity of features of interest in sEEG, such as dictionary learning, template matching, subband decomposition, etc. In addition, researchers have also built feature engineering from the time domain, frequency domain or nonlinear features of the sEEG, and have implemented automatic detection of IEDs using classifiers with one or more features, such as decision trees, artificial neural networks, gradient elevators, etc. Although the above-mentioned automatic IEDs detection method has achieved certain effects, there are some obstacles in clinical application. The main challenges are as follows:

(1) complex pre-processing. The sEEG contains a large amount of artifacts which cannot be filtered by a filtering algorithm, and the artifacts and IEDs are difficult to be accurately distinguished by using the traditional method, so that the traditional method is mostly provided with an individually designed artifact removing algorithm or an artificial artifact removing criterion, the engineering quantity is large, and pathological information contained in signals is easily damaged.

(2) The performance is poor among different patients. The external manifestations of different types, ages and sexes of epileptic patients, such as the forms and durations of IEDs, may differ greatly, and the clinical records may differ greatly between different hospitals due to the differences of the factors such as equipment, reference electrodes and environment. In addition, one or more poor channels may be present in the sEEG of certain patients. These reasons cause that the feature engineering set by the signal processing experience in the traditional method cannot fully and deeply mine the sEEG features, so that the differentiation between IEDs and background signals lacks robustness and is difficult to cope with complicated and variable patient situations in practical clinical application.

In view of the above, the present invention aims to provide an eclamptic electrical activity detection device and method during an eclamptic episode to solve one or more of the above technical problems.

Disclosure of Invention

In order to solve one or more technical problems in the prior art, according to an aspect of the present invention, a method for detecting seizure discharges in an epileptic patient during an episode period based on feature analysis of a simultaneous sensor spatial electroencephalogram signal is provided, which includes the following steps:

m w 256 input sample matrixes S through U-shaped semantic segmentation structural units^(m)Mapping and outputting as feature vector F^(m)＝[F1,F2,…,Fw]^T(ii) a Wherein Fw is the w-th sensor channelThe electroencephalogram signal of the channel is mapped by a U-shaped semantic segmentation structure to obtain characteristics, the U-shaped semantic segmentation structure unit comprises 4 superposed maximum pooling layers and 4 superposed deconvolution structures, two layers of 1 × 3 convolution layers are superposed in front of each maximum pooling layer, two layers of 1 × 3 convolution layers are superposed behind each deconvolution structure, and 1 × 1 convolution layer is superposed behind two layers of 1 × 3 convolution layers corresponding to the deconvolution structure at the uppermost layer;

according to the feature vector F^(m)＝[F1,F2,…,Fw]^TAnalyzing the characteristics of electroencephalogram signals of all sensor spaces at the same time, identifying characteristics related to epileptic discharge by utilizing the first relevance of epileptic discharge in all sensor spaces at the same time, the sporadic nature of artifact in all sensor spaces at the same time and the second relevance different from the first relevance, and further judging the S of a 1S time window^(m)Whether the segments contain epileptic-like discharge or not is judged, and the results of the m segments are summarized to realize the detection of the epileptic-like discharge in the whole segment of electroencephalogram signals;

wherein the m w 256 input sample matrices S^(m)＝[S1,S2,…,Sw]^TW is the number of sensor channels of the electroencephalogram signal, and Sw is a vector formed by single-channel electroencephalogram signal segments with the duration of 1 s.

According to another aspect of the invention, the characteristics of the electroencephalogram signals of the w sensor spaces at the same time are input through a gate control circulation unit, the input is decided in a chain form of a repeating network module, and a decision result corresponding to each time is output;

judging S of the 1S time window by utilizing a sigmod full-connection network according to the corresponding decision result of each moment^(m)And (4) whether the segments contain epileptic-like discharge or not, summarizing the results of the m segments, and realizing the detection of the epileptic-like discharge in the whole segment of electroencephalogram signals.

According to another aspect of the invention, a time window with a window width of 1S and an overlap of 50% is selected, the preprocessed electroencephalogram signal is segmented, and m w 256 input sample matrixes S are generated^(m)。

According to yet another aspect of the present invention, there is also provided a computer-readable storage medium, characterized in that a computer program is stored in the computer-readable storage medium, the computer program being adapted to be loaded and executed by a processor, so as to cause a computer device having the processor to perform the aforementioned method.

According to still another aspect of the present invention, a computer apparatus, comprising: a processor and a memory;

the processor is connected to a memory, wherein the memory is used for storing a computer program, and the processor is used for calling the computer program to make the computer device execute the method.

According to another aspect of the present invention, there is provided an epileptic seizure electrical activity detection device based on feature analysis of simultaneous sensor spatial electroencephalogram signals for an epileptic patient, comprising:

the electroencephalogram signal sensor is used for collecting an electroencephalogram signal of a patient, and the electroencephalogram signal is a multi-lead scalp electroencephalogram signal;

the electroencephalogram signal preprocessing unit is used for filtering and normalizing the electroencephalogram signals;

an EEG signal segmentation unit for segmenting the preprocessed EEG signal to generate m w 256 input sample matrixes S^(m)＝[S1,S2,…,Sw]^TW is the number of sensor channels of the electroencephalogram signal, and Sw is a vector formed by single-channel electroencephalogram signal segments with the duration of 1 s; and

an end-to-end binary unit for classifying the eclamptic electric activity signal and the background signal according to the input m w x 256 input sample matrixes S^(m)Outputting a seizure-like electrical activity detection result;

wherein, the end-to-end epileptic-like electrical activity signal and background signal two-classification unit comprises:

the U-shaped semantic division structure unit comprises 4 superposed maximum pooling layers and 4 superposed deconvolution structures, wherein two layers of 1-3 convolution layers are superposed in front of each maximum pooling layer, two layers of 1-3 convolution layers are superposed behind each deconvolution structure, 1-1 convolution layers are superposed behind two layers of 1-3 convolution layers corresponding to the deconvolution structure at the uppermost layer, and the U-shaped semantic division structure unit is used for inputting m w 256 input samplesThe matrix S^(m)Output feature vector F^(m)＝[F1,F2,…,Fw]^T(ii) a Fw is a feature obtained after the electroencephalogram signal of the w-th sensor channel is mapped by a U-shaped semantic segmentation structure; and

a feature analysis unit based on the space electroencephalogram signals of the simultaneous sensors and used for analyzing the feature vectors F^(m)＝[F1,F2,…,Fw]^TAnalyzing the characteristics of electroencephalogram signals of sensor spaces at the same time, identifying characteristics related to seizure-like electrical activity by utilizing the first relevance of seizure-like electrical activity in the sensor spaces at the same time, the sporadic nature of artifact in the sensor spaces at the same time and the second relevance different from the first relevance, and further judging the S of a 1S time window^(m)And (4) whether the segments contain epilepsy-like electrical activity or not, summarizing the results of the m segments, and realizing the detection of the epilepsy-like electrical activity in the whole segment of electroencephalogram signals.

According to another aspect of the invention, the feature analysis unit based on the simultaneous sensor space electroencephalogram signal comprises:

the gate control cycle unit is used for making a decision on input in a chain type of a repeating network module according to the characteristics of the electroencephalogram signals of the w sensor spaces at the same time and outputting a decision result corresponding to each time;

a sigmod fully-connected network for determining S of a 1S time window based on the input decision results corresponding to the respective time instants^(m)And (4) whether the segments contain epilepsy-like electrical activity or not, summarizing the results of the m segments, and realizing the detection of the epilepsy-like electrical activity in the whole segment of electroencephalogram signals.

According to still another aspect of the present invention, the electroencephalogram signal preprocessing unit includes: the band-pass filter with the cut-off frequency of 0.5-45Hz is used for filtering the input electroencephalogram signals; the reference electrodes are arranged at the ear lobes on the two sides of the patient; and the normalization unit is used for performing normalization processing on the re-referenced electroencephalogram signals.

According to another aspect of the invention, the sampling frequency of the electroencephalogram signal sensor is 256 Hz.

According to still another aspect of the present invention, the EEG signal segmentation unit selects a window width of 1s and an overlap of 50%An inter-window for dividing the preprocessed EEG signal to generate m w 256 input sample matrixes S^(m)。

According to another aspect of the invention, the invention also provides an epileptic-like electrical activity detection method based on feature analysis of a simultaneous sensor space electroencephalogram signal for an epileptic during an inter-seizure period, which is characterized by comprising the following steps:

collecting an electroencephalogram signal of a patient, wherein the electroencephalogram signal is a multi-lead scalp electroencephalogram signal;

filtering and normalizing the electroencephalogram signals through an electroencephalogram signal preprocessing unit;

segmenting the preprocessed electroencephalogram signals to generate m w-256 input sample matrixes S^(m)＝[S1,S2,…,Sw]^TW is the number of sensor channels of the electroencephalogram signal, and Sw is a vector formed by single-channel electroencephalogram signal segments with the duration of 1 s;

inputting m w 256 input sample matrixes S through end-to-end epilepsy-like electrical activity signal and background signal two-classification units^(m)Classifying and outputting seizure-like electrical activity detection results;

wherein, the input m w x 256 input sample matrixes S are input through an end-to-end epilepsy-like electrical activity signal and background signal two-classification unit^(m)The classification and output of epilepsy-like electrical activity detection results comprises:

inputting m w-256 input sample matrixes S through U-shaped semantic segmentation structure units^(m)Mapping and outputting as feature vector F^(m)＝[F1,F2,…,Fw]^T(ii) a The method comprises the following steps that Fw is a feature obtained after electroencephalogram signals of a w-th sensor channel are mapped through a U-shaped semantic segmentation structure, the U-shaped semantic segmentation structure unit comprises 4 superposed maximum pooling layers and 4 superposed deconvolution structures, two layers of 1 × 3 convolution layers are superposed in front of each maximum pooling layer, two layers of 1 × 3 convolution layers are superposed behind each deconvolution structure, and 1 × 1 convolution layers are superposed behind two layers of 1 × 3 convolution layers corresponding to the deconvolution structure on the uppermost layer;

according to the feature vector F^(m)＝[F1,F2,…,Fw]^TAnalyzing sensors at the same timeThe characteristics of spatial electroencephalogram signals are used for identifying characteristics related to seizure-like electrical activity by utilizing the first relevance of seizure-like electrical activity in each sensor space at the same time, the sporadic nature of artifact in each sensor space at the same time and the second relevance different from the first relevance, and then the S of a 1S time window is judged^(m)And (4) whether the segments contain epilepsy-like electrical activity or not, summarizing the results of the m segments, and realizing the detection of the epilepsy-like electrical activity in the whole segment of electroencephalogram signals.

According to another aspect of the invention, the characteristics of the electroencephalogram signals of the w sensor spaces at the same time are input through a gate control circulation unit, the input is decided in a chain form of a repeating network module, and a decision result corresponding to each time is output;

judging S of the 1S time window by utilizing a sigmod full-connection network according to the corresponding decision result of each moment^(m)And (4) whether the segments contain epilepsy-like electrical activity or not, summarizing the results of the m segments, and realizing the detection of the epilepsy-like electrical activity in the whole segment of electroencephalogram signals.

According to still another aspect of the present invention, the electroencephalogram signal preprocessing unit includes: the band-pass filter with the cut-off frequency of 0.5-45Hz is used for filtering the input electroencephalogram signals; the reference electrodes are arranged at the ear lobes on the two sides of the patient; and the normalization unit is used for performing normalization processing on the re-referenced electroencephalogram signals.

According to yet another aspect of the invention, the electroencephalogram signal of the patient is acquired by an electroencephalogram signal sensor having a sampling frequency of 256 Hz.

According to another aspect of the invention, a time window with a window width of 1S and an overlap of 50% is selected, the preprocessed electroencephalogram signal is segmented, and m w 256 input sample matrixes S are generated^(m)。

Compared with the prior art, the invention has one or more of the following technical effects:

(1) the first relevance of the epilepsy-like electrical activity in the sensor spaces at the same time, the sporadic nature of the artifact in the sensor spaces at the same time and the second relevance different from the first relevance are utilized to identify the relevant characteristics of the epilepsy-like electrical activity, the artifact which cannot be filtered by a filtering means can be filtered, and automatic detection can be completed without manually designing an artifact filtering unit and a characteristic project.

(2) For the U-type semantic segmentation structure used by the present invention: firstly, the superposition of multiple layers of pooling layers provides a multi-scale observation visual angle for the feature mapping of the model, so that the model can acquire corresponding morphological features in the sEEG for a single spike wave which is as small as 20ms in a 1s time window or a spike-slow complex wave which is as large as the whole time window; secondly, the model can guarantee that the model can not be influenced by factors such as gradient explosion and the like while excavating feature information of IED in the sEEG in enough depth due to the jump type cross-layer connection, a splicing mode is used for connecting the shallow layer features and the deep layer features, and the network can self-adaptively select the specific gravity of the shallow layer features and the deep layer features according to the difference features of the IED and the background sEEG, so that the model has better identification capability on abnormal discharge waveforms in different forms and modes in a single channel; and finally, the model uses a self-adaptive convolution filling mode to finish an automatic feature extraction mode from sEEG coding to decoding, and the intact time sequence and channel correlation in the signal are kept while the morphological features of the original signal are obtained, so that sufficient information is provided for further decision making.

(3) For the gated loop units used in the present invention: the model creatively considers the characteristics acquired by different sensors at the same time in the U-shaped semantic segmentation structure as sequence information representing the incidence relation among channels, and a gating cycle unit is used for acquiring the electroencephalogram signals of all sensor spaces at each time (corresponding to each sampling point in 256 sampling points) in a 1s time window, so that the analysis and decision of abnormal discharge of the detected morphological abnormality from a nervous system or pseudo-errors caused by random disturbance are made. The dynamic decision of each element information in the sequence by the update gate and the reset gate is greatly reduced, so that the influence of the overall decision of bad channels in the sEEG is greatly reduced; meanwhile, the output result is the comprehensive decision of the gating cycle unit on the whole sequence, so the arrangement sequence of elements in the sequence does not have great influence on the final decision, and the model has higher applicability to sEEG collected in different environments.

Drawings

So that the manner in which the above recited features of the present invention can be understood in detail, a more particular description of the invention, briefly summarized above, may be had by reference to embodiments. The drawings relate to preferred embodiments of the invention and are described below:

fig. 1 is a schematic structural diagram of an end-to-end epileptic-like electrical activity signal and background signal two-classification unit adopted by an epileptic patient inter-seizure epileptic electrical activity detection device according to a preferred embodiment of the invention;

FIG. 2 is a schematic representation of seizure-like electrical activities (IEDs) captured within a 1s time window in accordance with the present invention.

Detailed Description

Reference will now be made in detail to the various embodiments, one or more examples of which are illustrated in each figure. The examples are provided by way of explanation and are not meant as limitations. For example, features illustrated or described as part of one embodiment can be used on or in conjunction with any other embodiment to yield yet a further embodiment. It is intended that the present invention encompass such modifications and variations.

In the following description of the drawings, like reference numerals designate identical or similar structures. Generally, only the differences between the individual embodiments will be described. Descriptions of parts or aspects in one embodiment can also be applied to corresponding parts or aspects in another embodiment, unless explicitly stated otherwise.

Example 1

Referring to fig. 1-2, wherein fig. 1 is a schematic structural diagram of an end-to-end epileptic-like electrical activity signal and background signal classifying unit adopted by an inter-seizure epileptic patient epileptic-like electrical activity detection device according to a preferred embodiment of the invention; FIG. 2 is a schematic representation of seizure-like electrical activities (IEDs) captured within a 1s time window in accordance with the present invention.

According to a preferred embodiment of the present invention, there is provided an epileptic seizure electrical activity detection device based on feature analysis of simultaneous sensor spatial electroencephalogram signals for an epileptic patient, comprising:

the electroencephalogram signal sensor is used for collecting an electroencephalogram signal of a patient, and the electroencephalogram signal is a multi-lead scalp electroencephalogram signal;

the electroencephalogram signal preprocessing unit is used for filtering and normalizing the electroencephalogram signals;

an electroencephalogram signal segmentation unit (input signal segmentation unit) for segmenting the preprocessed electroencephalogram signal to generate m w-256 input sample matrices S^(m)＝[S1,S2,…,Sw]^TW is the number of sensor channels of the electroencephalogram signal, and Sw is a vector formed by single-channel electroencephalogram signal segments with the duration of 1 s; and

an end-to-end binary unit for classifying the eclamptic electric activity signal and the background signal according to the input m w x 256 input sample matrixes S^(m)And outputting the epilepsy-like electrical activity detection result.

Preferably, referring to fig. 1, the end-to-end seizure-like electrical activity signal and background signal classification unit includes:

the U-shaped semantic division structure unit comprises 4 superposed maximum pooling layers and 4 superposed deconvolution structures, wherein two layers of 1-3 convolution layers are superposed in front of each maximum pooling layer, two layers of 1-3 convolution layers are superposed behind each deconvolution structure, 1-1 convolution layers are superposed behind two layers of 1-3 convolution layers corresponding to the deconvolution structure at the uppermost layer, and the U-shaped semantic division structure unit is used for dividing the structure according to m input sample matrixes S of w 256 input^(m)Output feature vector F^(m)＝[F1,F2,…,Fw]^T(ii) a Fw is a feature obtained after the electroencephalogram signal of the w-th sensor channel is mapped by a U-shaped semantic segmentation structure; and

a feature analysis unit (feature analysis unit based on the spatial information of the simultaneous sensors) based on the spatial EEG signals of the simultaneous sensors, and is used for analyzing the feature vector F according to the feature vector^(m)＝[F1,F2,…,Fw]^TAnalyzing the characteristics of electroencephalogram signals of sensor spaces at the same time, identifying characteristics related to seizure-like electrical activity by utilizing the first relevance of seizure-like electrical activity in the sensor spaces at the same time, the sporadic nature of artifact in the sensor spaces at the same time and the second relevance different from the first relevance, and further judging the S of a 1S time window^(m)Whether the segment contains epilepsy-like electrical activity or not, and summarizing m segmentsAnd the result of the segment realizes the detection of epileptic-like electrical activity in the whole segment of electroencephalogram signals.

Advantageously, the invention can complete automatic detection of IEDs without manually designing a pseudo-error filtering unit and a characteristic project, and realizes IED detection with high accuracy and low false positive rate under the cross-crowd and multi-center background according with the clinical application environment, thereby reducing the burden of long-range EEG interpretation of clinical workers.

It will be appreciated that the types of artifacts that are difficult to filter by these filtering units, which are easily confused with epileptic electrical activity, are very large, and some, in addition to incidental artifacts that occur only in some or some of the sensor spaces, occur in most, or even all, of the sensor spaces. For example electrocardio-induced artefacts that produce a transient variation of the same size in all sensor spaces similar to the shape of a spike in seizure-like electrical activity. The research shows that the correlation of the real epileptic-like electrical activity and the real epileptic-like electrical activity in the sensor space at the same time is different. That is, the correlation of artifacts in sensor space is not consistent with the disease characteristics of epilepsy, e.g., true seizure-like electrical activity exhibits a first correlation in contemporaneous sensor space, while electrocardiographically induced artifacts exhibit a second correlation in contemporaneous sensor space that is different from the first correlation.

It should be noted that the method can distinguish artifacts which are easily confused with epileptic-like electrical activity in electroencephalogram signals, and the chain structure is used for obtaining relevance information among different sensors and reducing the influence of bad channels and element arrangement in electroencephalogram, so that the system has the potential of being applied to different scenes and crowds, namely the system has higher robustness in different medical centers.

According to another preferred embodiment of the present invention, the feature analysis unit based on the simultaneous sensor spatial electroencephalogram signals comprises:

the gate control cycle unit is used for making a decision on input in a chain type of a repeating network module according to the characteristics of the electroencephalogram signals of the w sensor spaces at the same time and outputting a decision result corresponding to each time;

a sigmod fully-connected network for determining S of a 1S time window based on the input decision results corresponding to the respective time instants^(m)And (4) whether the segments contain epilepsy-like electrical activity or not, summarizing the results of the m segments, and realizing the detection of the epilepsy-like electrical activity in the whole segment of electroencephalogram signals.

According to still another preferred embodiment of the present invention, the electroencephalogram signal preprocessing unit includes: the band-pass filter with the cut-off frequency of 0.5-45Hz is used for filtering the input electroencephalogram signals; the reference electrodes are arranged at the ear lobes on the two sides of the patient; and the normalization unit is used for performing normalization processing on the re-referenced electroencephalogram signals.

According to another preferred embodiment of the present invention, the sampling frequency of the electroencephalogram signal sensor is 256 Hz.

According to another preferred embodiment of the present invention, the electroencephalogram signal segmentation unit selects a time window having a window width of 1S and an overlap degree of 50%, segments the preprocessed electroencephalogram signal, and generates m w × 256 input sample matrices S^(m)。

According to another preferred embodiment of the present invention, there is provided a method for detecting epileptic-like electrical activity during an epileptic seizure based on feature analysis of simultaneous sensor spatial electroencephalograms, comprising the steps of:

collecting an electroencephalogram signal of a patient, wherein the electroencephalogram signal is a multi-lead scalp electroencephalogram signal;

filtering and normalizing the electroencephalogram signals through an electroencephalogram signal preprocessing unit;

segmenting the preprocessed electroencephalogram signals to generate m w-256 input sample matrixes S^(m)＝[S1,S2,…,Sw]^TW is the number of sensor channels of the electroencephalogram signal, and Sw is a vector formed by single-channel electroencephalogram signal segments with the duration of 1 s;

inputting m w 256 input sample matrixes S through end-to-end epilepsy-like electrical activity signal and background signal two-classification units^(m)Classifying and outputting seizure-like electrical activity detection results;

wherein, theInputting m w 256 input sample matrixes S through end-to-end epilepsy-like electrical activity signal and background signal two-classification units^(m)The classification and output of epilepsy-like electrical activity detection results comprises:

inputting m w-256 input sample matrixes S through U-shaped semantic segmentation structure units^(m)Mapping and outputting as feature vector F^(m)＝[F1,F2,…,Fw]^T(ii) a The method comprises the following steps that Fw is a feature obtained after electroencephalogram signals of a w-th sensor channel are mapped through a U-shaped semantic segmentation structure, the U-shaped semantic segmentation structure unit comprises 4 superposed maximum pooling layers and 4 superposed deconvolution structures, two layers of 1 × 3 convolution layers are superposed in front of each maximum pooling layer, two layers of 1 × 3 convolution layers are superposed behind each deconvolution structure, and 1 × 1 convolution layers are superposed behind two layers of 1 × 3 convolution layers corresponding to the deconvolution structure on the uppermost layer;

according to the feature vector F^(m)＝[F1,F2,…,Fw]^TAnalyzing the characteristics of electroencephalogram signals of sensor spaces at the same time, identifying characteristics related to seizure-like electrical activity by utilizing the first relevance of seizure-like electrical activity in the sensor spaces at the same time, the sporadic nature of artifact in the sensor spaces at the same time and the second relevance different from the first relevance, and further judging the S of a 1S time window^(m)And (4) whether the segments contain epilepsy-like electrical activity or not, summarizing the results of the m segments, and realizing the detection of the epilepsy-like electrical activity in the whole segment of electroencephalogram signals.

According to another preferred embodiment of the invention, the characteristics of the electroencephalogram signals of the w sensor spaces at the same time are input through a gate control cycle unit, the input is decided in a chain form of a repeating network module, and a decision result corresponding to each time is output;

judging S of the 1S time window by utilizing a sigmod full-connection network according to the corresponding decision result of each moment^(m)And (4) whether the segments contain epilepsy-like electrical activity or not, summarizing the results of the m segments, and realizing the detection of the epilepsy-like electrical activity in the whole segment of electroencephalogram signals.

According to still another preferred embodiment of the present invention, the electroencephalogram signal preprocessing unit includes: the band-pass filter with the cut-off frequency of 0.5-45Hz is used for filtering the input electroencephalogram signals; the reference electrodes are arranged at the ear lobes on the two sides of the patient; and the normalization unit is used for performing normalization processing on the re-referenced electroencephalogram signals.

According to another preferred embodiment of the present invention, the electroencephalogram signal of the patient is acquired by an electroencephalogram signal sensor with a sampling frequency of 256 Hz.

According to another preferred embodiment of the present invention, a time window with a window width of 1S and an overlap of 50% is selected, and the preprocessed electroencephalogram signal is divided to generate m w × 256 input sample matrices S^(m)。

According to another preferred embodiment of the present invention, there is provided a method for detecting epileptic-like electrical activity during an epileptic seizure by fusing multi-scale time domain information and sensor information in scalp electroencephalogram, the method mainly comprising the steps of:

step 1, selecting a band-pass filter with the cutoff frequency of 0.5-45Hz, filtering the input original multi-lead scalp electroencephalogram signals, setting electrodes (A1 and A2) at the ear lobes of two sides of a patient as reference electrodes, calculating the sEEG after re-reference, and carrying out normalization processing on the sEEG by using a maximum and minimum value method;

step 2, adjusting the sampling frequency of the sEEG to 256Hz by calculation, then selecting a time window with the window width of 1S and the overlapping degree of 50 percent, segmenting the preprocessed sEEG signal, and generating m w 256 input sample matrixes S^(m)＝[S1,S2,…,Sw]^TWhere w is the number of channels of the re-referenced sEEG signal. Sw is a vector formed by single-channel electroencephalogram signal segments with the duration of 1 s;

step 3, synchronously extracting S by utilizing U-shaped semantic segmentation network^(m)The morphological characteristics of IED in the single-channel signal output a characteristic vector F with the same size^(m)＝[F1,F2,…,Fw]^TAnd then analyzed by gated cycle unit F^(m)IED characteristics of the single-channel sEEG on a time scale are output, decision results of each channel are output, then, the gating cycle unit is used for the connected EEG channels to mine spatial information related to the IED on the sensor level, and S with the duration of 1S is finally determined^(m)Whether the segment contains IED events or not is judged, and finally m IED events are summarizedAnd as a result of the segments, the labeling of IED events in the whole long-range sEEG can be realized, so that the workload of clinical workers is reduced.

Preferably, the inter-seizure epileptic-like electrical activity detection model for epileptic patients is a model for realizing two classifications of IEDs and background signals in sEEG end to end, and the structural diagram is shown in FIG. 1. Let m sEEG signals of input model be S^(m)＝[S1,S2,…,Sw]^T，w∈[1，19]On the left side of the U-shaped semantic segmentation structure, the model obtains 4 progressively increased views under different scales by means of 4 superposed maximum pooling layers, extracts the local features of the single sEEG channel Sw under the current view by means of a convolution layer with 1 × 3 of two layers of convolution kernels superposed before pooling each time, and finally completes the coding process of the sEEG under different scales on the time domain at the bottom end of the U-shaped semantic structure; on the right side of the U-shaped semantic segmentation structure, the model performs up-sampling by means of 4 deconvolution structures with the same number of pooled layers, so that morphological features on the single channel of the sEEG coded by the model are decoded, the corresponding position relation of each feature and the original signal on the time domain is recovered, and the F-shaped differential expression of the abnormal waveform and the background in each channel in the 1s time window is realized^(m)＝[F1，F2，…，Fw]^T，w∈[1，19](ii) a The feature expression of the fine sampling point (i.e. each time) obtained by the U-type semantic segmentation structure is sent to the subsequent gating cycle unit, the gating cycle unit makes a decision on input in a chain form of repeated network modules, signals from 19 sensors are sequentially input into the gating cycle unit, wherein the output y of each module in the chain structure_tInput x from the current module_tAnd the state h of the previous module_t-1Via a reset gate r_tAnd update gate z_tAnd (3) making a common decision after calculation, specifically calculating as follows:

r_t＝σ(W_r·[h_t-1，x_t]) (2)

z_t＝σ(W_z·[h_t-1，x_t]) (3)

y_t＝σ(W_y·h_t) (6)

wherein]Is vector connected, is the product of matrix, σ () is the activation function, W_r，W_z，W_yRespectively, the weight matrix learned in the training for the corresponding gate. Namely, the final output of the gating cycle unit is a comprehensive decision result obtained by dynamically weighting semantic morphological characteristics acquired at 19 sensors. Finally, a feature vector formed by comprehensive decision results of 19 sensors in a time domain is input into a full-connection network with sigmod as an activation function, and a judgment result of whether the samples in the current 1s time window contain seizure-like abnormal electrical activity is obtained.

Preferably, the detection method of the invention and the existing comparative model are quantitatively evaluated in effect, wherein, on one hand, a balance training set is used for cross validation of 5 folds across population, and the performance of the invention and the existing end-to-end detection model is contrastively analyzed, and on the other hand, an independent test set is used for evaluating the usability and robustness of the invention in a clinical background. The quantitative evaluation results are as follows.

TABLE 1 average test results of 5-fold cross validation of different models in a balance training set

Shown in table 1 are model evaluation results of the method of the present invention in cross validation of 5-fold across population, and a typical convolutional neural network model VGGNet under the same conditions is listed, compared with the IEDs detection effect of the method of the present invention. It can be seen that the method of the present invention has the best effect on various indexes. Compared with a classical convolutional neural network, the U-type semantic network obtained by comparing VGGNet with UNet can better acquire multi-scale morphological information of IEDs in a time domain; similarly, comparing UNet with the method of the present invention, the sensor information at the same time is introduced through the gate control cycle unit for analysis and decision, and the capability of distinguishing IEDs from background signals in the sEEG is greatly improved.

TABLE 2 detection of IEDs in independent test sets by the method of the present invention

Table 2 shows the effect of the method of the present invention on the detection of IEDs in a separate test set. Long-range sEEG full-scale input models up to 1 hour per patient were evaluated for performance with a 1s time window and 50% overlap. Unlike the balance training set, the proportion of IEDs to background signal in each sample of the independent test set is severely unbalanced, and the proportion of the sample with the smallest difference is 1: 23, therefore, the precision obtained by the test and the F1-Score effect closely related to the precision are lower than the result of 5-fold cross validation. However, the advantage of high sensitivity (i.e. recall) of the model can ensure that IEDs in the patient sEEG are acquired as completely as possible, thereby greatly avoiding the huge workload brought by a large number of backgrounds in the long-range sEEG, particularly the artifact easily confused with the IEDs, for clinical diagnosis, saving a large amount of time and energy for doctors to interpret the sEEG, and providing a high-quality means for interpreting the sEEG of epileptic patients in remote areas without high-quality medical resources.

Compared with the prior art, the invention has one or more of the following technical effects:

(1) the first relevance of the epilepsy-like electrical activity in the sensor spaces at the same time, the sporadic nature of the artifact in the sensor spaces at the same time and the second relevance different from the first relevance are utilized to identify the relevant characteristics of the epilepsy-like electrical activity, the artifact which cannot be filtered by a filtering means can be filtered, and automatic detection can be completed without manually designing an artifact filtering unit and a characteristic project.

(2) For the U-type semantic segmentation structure used by the present invention: firstly, the superposition of multiple layers of pooling layers provides a multi-scale observation visual angle for the feature mapping of the model, so that the model can acquire corresponding morphological features in the sEEG for a single spike wave which is as small as 20ms in a 1s time window or a spike-slow complex wave which is as large as the whole time window; secondly, the model can guarantee that the model can not be influenced by factors such as gradient explosion and the like while excavating feature information of IED in the sEEG in enough depth due to the jump type cross-layer connection, a splicing mode is used for connecting the shallow layer features and the deep layer features, and the network can self-adaptively select the specific gravity of the shallow layer features and the deep layer features according to the difference features of the IED and the background sEEG, so that the model has better identification capability on abnormal discharge waveforms in different forms and modes in a single channel; and finally, the model uses a self-adaptive convolution filling mode to finish an automatic feature extraction mode from sEEG coding to decoding, and the intact time sequence and channel correlation in the signal are kept while the morphological features of the original signal are obtained, so that sufficient information is provided for further decision making.

(3) For the gated loop units used in the present invention: the model creatively considers the characteristics acquired by different sensors at the same time in the U-shaped semantic segmentation structure as sequence information representing the incidence relation among channels, and a gating cycle unit is used for acquiring the electroencephalogram signals of all sensor spaces at each time (corresponding to each sampling point in 256 sampling points) in a 1s time window, so that the analysis and decision of abnormal discharge of the detected morphological abnormality from a nervous system or pseudo-errors caused by random disturbance are made. The dynamic decision of each element information in the sequence by the update gate and the reset gate is greatly reduced, so that the influence of the overall decision of bad channels in the sEEG is greatly reduced; meanwhile, the output result is the comprehensive decision of the gating cycle unit on the whole sequence, so the arrangement sequence of elements in the sequence does not have great influence on the final decision, and the model has higher applicability to sEEG collected in different environments.

While the foregoing is directed to embodiments of the present invention, other and further embodiments of the invention may be devised without departing from the basic scope thereof, and the scope thereof is determined by the claims that follow.

The above embodiments are merely preferred embodiments of the present invention, which are not intended to limit the present invention, and the features of the embodiments that do not violate each other may be combined with each other. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.