WO2022135449A1 - 癫痫患者发作间期痫样电活动检测装置和方法 - Google Patents

癫痫患者发作间期痫样电活动检测装置和方法 Download PDF

Info

Publication number
WO2022135449A1
WO2022135449A1 PCT/CN2021/140341 CN2021140341W WO2022135449A1 WO 2022135449 A1 WO2022135449 A1 WO 2022135449A1 CN 2021140341 W CN2021140341 W CN 2021140341W WO 2022135449 A1 WO2022135449 A1 WO 2022135449A1
Authority
WO
WIPO (PCT)
Prior art keywords
eeg signal
electrical activity
signal
eeg
epileptiform
Prior art date
Application number
PCT/CN2021/140341
Other languages
English (en)
French (fr)
Inventor
张冀聪
韦博轩
赵晓慧
Original Assignee
北京航空航天大学
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 北京航空航天大学 filed Critical 北京航空航天大学
Publication of WO2022135449A1 publication Critical patent/WO2022135449A1/zh

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
    • A61B5/40—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
    • A61B5/7203—Signal processing specially adapted for physiological signals or for diagnostic purposes for noise prevention, reduction or removal
    • 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
    • A61B5/7225—Details of analog processing, e.g. isolation amplifier, gain or sensitivity adjustment, filtering, baseline or drift compensation
    • 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
    • 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

Definitions

  • the invention relates to the field of biomedical detection, in particular to a device and method for detecting interictal epileptiform electrical activity in epilepsy patients, in particular to an interictal epileptiform electric activity detection method for epilepsy patients based on feature analysis of spatial EEG signals from sensors at the same time Activity detection apparatus and method.
  • Scalp electroencephalography is a non-invasive signal acquisition method for clinical recording of brain-related activity. It has high temporal resolution and plays an important role in the detection, diagnosis, treatment, efficacy evaluation and pathological research of epilepsy.
  • the current epilepsy diagnosis based on scalp EEG is mainly the detection of seizures, however, seizures are uncommon.
  • nearly 80% of epilepsy patients have abnormal neuronal discharges between seizures, and these abnormal discharges observed in sEEG are called Interictal Epileptiform Discharges (IEDs).
  • IEDs Interictal Epileptiform Discharges
  • the IED is a strong support for epilepsy diagnosis or seizure risk assessment and is a key factor in identifying the underlying lesions of epilepsy origin. Therefore, the detection of IEDs based on sEEG is of great significance for the diagnosis of epilepsy.
  • IEDs are more frequent and appear in the scalp EEG in various forms, such as spikes, sharp waves, spike-slow complexes, multi-spike complexes, etc., and their duration in sEEG varies (The short one is only 20-70ms, the long one lasts for more than 1s), it is easy to be confused with the artifact (caused by chewing, sensor sliding, heartbeat, etc.) that cannot be filtered out by filtering means in sEEG, which makes the detection of IEDs in sEEG difficult. extremely challenging.
  • the present invention aims to provide an interictal epileptiform electrical activity detection device and method for epilepsy patients to solve one or more of the above technical problems.
  • a device for detecting interictal epileptiform electrical activity in epileptic patients based on feature analysis of spatial EEG signals from sensors at the same time which is characterized by: include:
  • an EEG signal sensor for collecting the EEG signal of a patient, the EEG signal being a multi-lead scalp EEG signal;
  • an EEG signal preprocessing unit for filtering and normalizing the EEG signal
  • the end-to-end epileptiform electrical activity signal and background signal binary classification unit is used to output the epileptiform electrical activity detection result according to the input m w*256 input sample matrix S (m) ;
  • the end-to-end epileptiform electrical activity signal and background signal binary classification unit includes:
  • U-shaped semantic segmentation structural unit including 4 superimposed max pooling layers and 4 superimposed deconvolution structures, each max pooling layer is superimposed with two layers of 1*3 convolution layers, each deconvolution structure There are two layers of 1*3 convolutional layers superimposed on the back, and the two layers of 1*3 convolutional layers corresponding to the topmost deconvolution structure are superimposed with a 1*1 convolutional layer.
  • the first correlation of the electrical activity in each sensor space at the same time and the sporadic nature of the artifact in each sensor space at the same time and the second correlation different from the first correlation identify the features related to the epileptiform electrical activity, and then determine the 1s time Whether the S (m) segment of the window contains epileptiform electrical activity, the results of m segments are summarized to realize the detection of epileptiform electrical activity in the entire EEG signal.
  • the feature analysis unit based on the spatial EEG signal of the simultaneous sensor includes:
  • the gated loop unit is used to make decisions on the input in the form of a chain of repeated network modules according to the characteristics of the EEG signals at the same time in the input w sensor spaces, and output the corresponding decision results at each moment;
  • the sigmod fully-connected network is used to determine whether the S (m) segment of the 1s time window contains epileptiform electrical activity according to the corresponding decision results of the inputted time, and summarize the results of m segments to realize the whole EEG signal. Detection of epileptiform electrical activity.
  • the EEG signal preprocessing unit includes: a band-pass filter with a cutoff frequency of 0.5-45 Hz, used to filter the input EEG signal; reference electrodes, arranged on both sides of the patient At the earlobe; the normalization unit is used to normalize the EEG signal after re-reference.
  • the sampling frequency of the brain electrical signal sensor is 256 Hz.
  • the EEG signal segmentation unit selects a time window with a window width of 1s and an overlap of 50%, and segments the preprocessed EEG signals to generate m inputs of w*256 Sample matrix S (m) .
  • a method for detecting interictal epileptiform electrical activity in epilepsy patients based on feature analysis of spatial EEG signals from sensors at the same time which is characterized by comprising the following steps:
  • the EEG signal is a multi-lead scalp EEG signal
  • classifying the input m w*256 input sample matrices S (m) through the end-to-end epileptiform electrical activity signal and background signal binary classification unit and outputting the epileptiform electrical activity detection results include:
  • the U-shaped semantic segmentation structural unit includes 4 superimposed max pooling layers and 4 superimposed deconvolution structures, each maximum There are two layers of 1*3 convolution layers superimposed in front of the pooling layer, and two layers of 1*3 convolution layers are superimposed after each deconvolution structure.
  • the top layer of the deconvolution structure corresponds to two layers of 1*3 convolution layers.
  • a 1*1 convolutional layer is superimposed after the layer;
  • F (m) [F1, F2, .
  • the difference between the sporadicness of each sensor space at the same time and the second correlation different from the first correlation identify the features related to the epileptiform electrical activity, and then determine whether the S (m) segment of the 1s time window contains the epileptiform electrical activity, Summarize the results of m segments to realize the detection of epileptiform electrical activity in the entire EEG signal.
  • the feature of the EEG signals of the input w sensor spaces at the same moment is used to make decisions on the input in a chained form of repeated network modules through the gated loop unit, and the corresponding decision results at each moment are output;
  • the sigmod fully connected network is used to determine whether the S (m) segment of the 1s time window contains epileptiform electrical activity according to the corresponding decision results at each moment, and the results of m segments are summarized to realize the epileptiform electrical activity in the entire EEG signal. Activity detection.
  • the EEG signal preprocessing unit includes: a band-pass filter with a cutoff frequency of 0.5-45 Hz, used to filter the input EEG signal; reference electrodes, arranged on both sides of the patient At the earlobe; the normalization unit is used to normalize the EEG signal after re-reference.
  • the EEG signal of the patient is collected through an EEG signal sensor with a sampling frequency of 256 Hz.
  • a time window with a window width of 1 s and an overlap of 50% is selected, the preprocessed EEG signal is segmented, and m input sample matrices S (m) of w*256 are generated.
  • the present invention has one or more of the following technical effects:
  • the superposition of multi-layer pooling layers provides a multi-scale observation perspective for the feature mapping of the model, so that for a single spine as small as 20ms in a 1s time window Whether the wave is large enough to fill the entire time window, the model can obtain the corresponding morphological features in the sEEG;
  • the jump cross-layer connection used by the model enables the model to ensure sufficient depth to mine the feature information of the IED in the sEEG At the same time, it will not be affected by factors such as gradient explosion, in which the connection method of shallow and deep features is spliced, and the network can adaptively select shallow and deep features according to the difference between IED and background sEEG features.
  • the proportion of sEEG in the model makes the model have a good ability to identify abnormal discharge waveforms in different shapes and modes in a single channel.
  • the model uses an adaptive convolution filling method to complete the automatic feature extraction method from sEEG encoding to decoding.
  • the original signal morphological features are preserved while preserving the intact timing and channel correlations in the signal, providing sufficient information for further decision-making.
  • the model creatively regards the features acquired by different sensors in the U-shaped semantic segmentation structure at the same time as sequence information representing the relationship between channels, and uses the gated recurrent unit to obtain all The EEG signals of the sensor space at each moment (corresponding to each sampling point in the 256 sampling points) in the 1s time window, so as to analyze and decide whether the detected morphological abnormalities come from abnormal discharges of the nervous system or are only caused by random disturbances Artifacts.
  • the dynamic decision-making of the update gate and the reset gate on the information of each element in the sequence of the present invention greatly reduces the influence of the bad channel in the sEEG on the overall decision-making; at the same time, since the output result is the synthesis of the overall sequence by the gated cyclic unit Therefore, the arrangement order of elements in the sequence will not have a significant impact on the final decision, making the model more applicable to sEEG collected in different environments.
  • FIG. 1 is a schematic structural diagram of an end-to-end epileptiform electrical activity signal and a background signal two-classification unit adopted by a device for detecting interictal epilepsy-like electrical activity in epilepsy patients according to a preferred embodiment of the present invention
  • FIG. 2 is a schematic diagram of intercepting epileptiform electrical activities (IEDs) in a 1s time window in the present invention.
  • IEDs epileptiform electrical activities
  • FIG. 1 is a schematic structural diagram of an end-to-end epileptiform electrical activity signal and a background signal binary classification unit adopted by a device for detecting interictal epilepsy-like electrical activity in epilepsy patients according to a preferred embodiment of the present invention
  • Figure 2 is a schematic diagram of intercepting epileptiform electrical activities (IEDs) in a 1s time window in the present invention.
  • IEDs epileptiform electrical activities
  • a device for detecting interictal epileptiform electrical activity in epilepsy patients based on feature analysis of spatial EEG signals from sensors at the same time is provided, which is characterized by comprising:
  • an EEG signal sensor for collecting the EEG signal of a patient, the EEG signal being a multi-lead scalp EEG signal;
  • an EEG signal preprocessing unit for filtering and normalizing the EEG signal
  • the end-to-end epileptiform electrical activity signal and background signal binary classification unit is used to output the epileptiform electrical activity detection result according to the input sample matrix S (m) of m w*256.
  • the end-to-end epileptiform electrical activity signal and background signal binary classification unit includes:
  • U-shaped semantic segmentation structural unit including 4 superimposed max pooling layers and 4 superimposed deconvolution structures, each max pooling layer is superimposed with two layers of 1*3 convolution layers, each deconvolution structure There are two layers of 1*3 convolutional layers superimposed on the back, and the two layers of 1*3 convolutional layers corresponding to the topmost deconvolution structure are superimposed with a 1*1 convolutional layer.
  • the characteristics of the EEG signals in each sensor space are identified by using the first correlation of the epileptiform electrical activity in each sensor space at the same time, the sporadic nature of the artifact in each sensor space at the same time, and the second correlation different from the first correlation
  • the features related to epileptiform electrical activity are then determined to determine whether the S (m) segment in the 1s time window contains epileptiform electrical activity, and the results of m segments are summarized to realize the detection of epileptiform electrical activity in the entire EEG signal.
  • the present invention can complete the automatic detection of IEDs without the need to manually design an artifact filtering unit and feature engineering, and realize IEDs with high accuracy and low false positive rate under the cross-population and multi-center background that conforms to the clinical application environment.
  • the detection can reduce the burden on the interpretation of long-range sEEG for clinicians.
  • the feature analysis unit based on the spatial EEG signal of the simultaneous sensor includes:
  • the gated loop unit is used to make decisions on the input in the form of a chain of repeated network modules according to the characteristics of the EEG signals of the input w sensor spaces at the same time, and output the corresponding decision results at each moment;
  • the sigmod fully-connected network is used to determine whether the S (m) segment of the 1s time window contains epileptiform electrical activity according to the corresponding decision results of the inputted time, and summarize the results of m segments to realize the whole EEG signal. Detection of epileptiform electrical activity.
  • the EEG signal preprocessing unit includes: a band-pass filter with a cut-off frequency of 0.5-45 Hz, for filtering the input EEG signal; a reference electrode, which is set on the patient At the bilateral earlobes; the normalization unit is used to normalize the re-referenced EEG signals.
  • the sampling frequency of the brain electrical signal sensor is 256 Hz.
  • the EEG signal segmentation unit selects a time window with a window width of 1s and an overlap of 50%, and segments the preprocessed EEG signals to generate m w*256 The input sample matrix S (m) of .
  • a method for detecting interictal epileptiform electrical activity in epilepsy patients based on feature analysis of spatial EEG signals from sensors at the same time which is characterized by comprising the following steps:
  • the EEG signal is a multi-lead scalp EEG signal
  • classifying the input m w*256 input sample matrices S (m) through the end-to-end epileptiform electrical activity signal and background signal binary classification unit and outputting the epileptiform electrical activity detection results include:
  • the U-shaped semantic segmentation structural unit includes 4 superimposed max pooling layers and 4 superimposed deconvolution structures, each maximum There are two layers of 1*3 convolution layers superimposed in front of the pooling layer, and two layers of 1*3 convolution layers are superimposed after each deconvolution structure.
  • the top layer of the deconvolution structure corresponds to two layers of 1*3 convolution layers.
  • a 1*1 convolutional layer is superimposed after the layer;
  • F (m) [F1, F2, .
  • the difference between the sporadicness of each sensor space at the same time and the second correlation different from the first correlation identify the features related to the epileptiform electrical activity, and then determine whether the S (m) segment of the 1s time window contains the epileptiform electrical activity, Summarize the results of m segments to realize the detection of epileptiform electrical activity in the entire EEG signal.
  • the characteristics of the EEG signals of the input w sensor spaces at the same time are used to make decisions on the input in the form of a chain of repeated network modules through the gated loop unit, and the corresponding decisions at each moment are output. result;
  • the sigmod fully connected network is used to determine whether the S (m) segment of the 1s time window contains epileptiform electrical activity according to the corresponding decision results at each moment, and the results of m segments are summarized to realize the epileptiform electrical activity in the entire EEG signal. Activity detection.
  • the EEG signal preprocessing unit includes: a band-pass filter with a cut-off frequency of 0.5-45 Hz, for filtering the input EEG signal; a reference electrode, which is set on the patient At the bilateral earlobes; the normalization unit is used to normalize the re-referenced EEG signals.
  • the EEG signal of the patient is collected by an EEG signal sensor with a sampling frequency of 256 Hz.
  • a time window with a window width of 1 s and an overlap of 50% is selected, the preprocessed EEG signal is segmented, and m input sample matrices S (m ) .
  • a method for detecting interictal epileptiform electrical activity in epilepsy patients by fusing multi-scale time domain information and sensor information in scalp EEG which mainly includes the following steps:
  • Step 1 Select a band-pass filter with a cutoff frequency of 0.5-45Hz, filter the input original multi-lead scalp EEG signal, and then set the electrodes ("A1", "A2") at the patient's bilateral earlobes as reference Electrodes, calculate the re-referenced sEEG and normalize it using the maximum-minimum method;
  • Spatial information is used to finally determine whether the 1s-long S (m) segment contains IED events, and finally the results of m segments are aggregated to realize the labeling of IED events in the entire long-range sEEG, thereby reducing the workload of clinicians. .
  • the interictal epileptiform electrical activity detection model of epilepsy patients of the present invention is an end-to-end model that realizes the two-classification of IEDs and background signals in sEEG, and its structural diagram is shown in FIG. 1 .
  • the model relies on 4 superimposed max pooling
  • the fusion layer obtains four progressively increasing fields of view at different scales, and extracts the local features of a single channel Sw of the sEEG under the current field of view by relying on the two convolution layers superimposed before each pooling with a convolution kernel of 1*3.
  • the encoding process of sEEG at different scales in the time domain is completed at the bottom of the U-shaped semantic structure; on the right side of the U-shaped semantic segmentation structure, the model relies on the same number of 4 deconvolution structures as the pooling layers for upsampling.
  • the morphological features on a single channel of sEEG encoded by the model are decoded, and the corresponding positional relationship between each feature and the original signal in the time domain is restored to realize the difference between the abnormal waveform and the background in each channel within the 1s time window.
  • the expression F (m) [F1, F2,..., Fw] T , w ⁇ [1, 19]; the feature expression obtained by the U-shaped semantic segmentation structure to the sampling point (that is, every moment) will be sent to Among the subsequent gated recurrent units, the gated recurrent unit makes decisions on the input in a chain of repeating network modules, where the signals from the 19 sensors are sequentially input to the gated recurrent unit, where each element in the chain structure is The output y t of a module is jointly decided by the input x t of the current module and the state h t-1 of the previous module through the reset gate rt and the update gate z t .
  • the specific calculation is as follows:
  • [] is the vector connection
  • * is the product of matrices
  • ⁇ ( ) is the activation function
  • W r , W z , W y are the weight matrices learned by the corresponding gates during training, respectively. That is, the final output of the gated recurrent unit is the comprehensive decision result after dynamic weighting of semantic-level morphological features obtained from 19 sensors.
  • the feature vector composed of the comprehensive decision results of 19 sensors in the time domain is input to the fully connected network with sigmod as the activation function, and the discrimination result of whether the samples in the current 1s time window contain abnormal electrical activity in the interictal epilepsy is obtained.
  • a 5-fold cross-validation across the population is performed by using a balanced training set, and the present invention and the existing end-to-end detection are compared and analyzed.
  • the performance of the model is to assess the usability and robustness of the present invention in a clinical context using an independent test set.
  • the quantitative evaluation results are as follows.
  • Table 1 Average detection results of 5-fold cross-validation for different models in the balanced training set
  • Table 1 shows the model evaluation results of the method of the present invention in cross-population 5-fold cross-validation, and lists the typical convolutional neural network model VGGNet under the same conditions, and compares the detection effect of IEDs with the method of the present invention. It is not difficult to see that each index of the method used in the present invention achieves the best results. Comparing VGGNet and UNet, U-shaped semantic network can better obtain multi-scale morphological information of IEDs in time domain than classical convolutional neural network. Similarly, comparing UNet and the method of the present invention, it can be obtained by gating loop The unit introduces simultaneous sensor information for analysis and decision-making, which greatly improves the ability of sEEG to distinguish between IEDs and background signals.
  • Table 2 shows the IEDs detection effect of the method of the present invention for the independent test set.
  • the performance of the model was evaluated with a 1s time window and a 50% overlap of the long-range sEEG complete input model for up to 1 hour per patient.
  • the ratio of IEDs to the background signal in each sample of the independent test set is seriously unbalanced, and the ratio of the two samples in the sample with the smallest gap also reaches 1:23, thus making the test accuracy and accuracy closely related.
  • the relevant F1-Score effect is lower than the result of 5-fold cross-validation.
  • the advantage of high sensitivity (ie recall) of the model can ensure that the IEDs in the sEEG of patients can be obtained as completely as possible, and to a great extent avoid a large number of backgrounds in the long-range sEEG, especially the artifacts that are easily confused with IEDs for clinical diagnosis.
  • the huge workload brought by it saves a lot of time and energy for doctors to interpret sEEG, and also provides high-quality sEEG interpretation methods for epilepsy patients in remote areas without high-quality medical resources.
  • the present invention has one or more of the following technical effects:
  • the superposition of multi-layer pooling layers provides a multi-scale observation perspective for the feature mapping of the model, so that for a single spine as small as 20ms in a 1s time window Whether the wave is large enough to fill the entire time window, the model can obtain the corresponding morphological features in the sEEG;
  • the jump cross-layer connection used by the model enables the model to ensure sufficient depth to mine the feature information of the IED in the sEEG At the same time, it will not be affected by factors such as gradient explosion, in which the connection method of shallow and deep features is spliced, and the network can adaptively select shallow and deep features according to the difference between IED and background sEEG features.
  • the proportion of sEEG in the model makes the model have a good ability to identify abnormal discharge waveforms in different shapes and modes in a single channel; finally, the model uses an adaptive convolution filling method to complete the automatic feature extraction method from sEEG encoding to decoding.
  • the original signal morphological features are preserved while preserving the intact timing and channel correlations in the signal, providing sufficient information for further decision-making.
  • the model creatively regards the features acquired by different sensors in the U-shaped semantic segmentation structure at the same time as sequence information representing the relationship between channels, and uses the gated recurrent unit to obtain all The EEG signals of the sensor space at each moment (corresponding to each sampling point in the 256 sampling points) in the 1s time window, so as to analyze and decide whether the detected morphological abnormalities come from abnormal discharges of the nervous system or are only caused by random disturbances Artifacts.
  • the dynamic decision-making of the update gate and the reset gate on the information of each element in the sequence of the present invention greatly reduces the influence of the bad channel in the sEEG on the overall decision-making; at the same time, since the output result is the synthesis of the overall sequence by the gated cyclic unit Therefore, the arrangement order of elements in the sequence will not have a significant impact on the final decision, making the model more applicable to sEEG collected in different environments.

Landscapes

  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Signal Processing (AREA)
  • Physics & Mathematics (AREA)
  • Molecular Biology (AREA)
  • Public Health (AREA)
  • Pathology (AREA)
  • Physiology (AREA)
  • Biomedical Technology (AREA)
  • Heart & Thoracic Surgery (AREA)
  • Medical Informatics (AREA)
  • Artificial Intelligence (AREA)
  • Surgery (AREA)
  • Animal Behavior & Ethology (AREA)
  • General Health & Medical Sciences (AREA)
  • Biophysics (AREA)
  • Veterinary Medicine (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Psychiatry (AREA)
  • Neurology (AREA)
  • Neurosurgery (AREA)
  • Power Engineering (AREA)
  • Evolutionary Computation (AREA)
  • Fuzzy Systems (AREA)
  • Mathematical Physics (AREA)
  • Measurement And Recording Of Electrical Phenomena And Electrical Characteristics Of The Living Body (AREA)

Abstract

一种癫痫患者发作间期痫样电活动检测装置和方法，该检测装置包括：脑电信号传感器，用于采集患者的脑电信号；脑电信号预处理单元，用于对脑电信号滤波和归一化处理；脑电信号分割单元，用于对预处理后的脑电信号进行分割，生成m个w*256的输入样本矩阵S (m)；以及端到端的痫样电活动信号与背景信号二分类单元，用于根据输入的m个w*256的输入样本矩阵S (m)输出痫样电活动检测结果。该方法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。

Description

癫痫患者发作间期痫样电活动检测装置和方法 技术领域
本发明涉及生物医学检测领域，具体涉及一种癫痫患者发作间期痫样电活动检测装置和方法，尤其是一种基于同时刻传感器空间脑电信号的特征分析的癫痫患者发作间期痫样电活动检测装置和方法。
背景技术
头皮脑电图(sEEG)是一种用于临床记录脑相关活动的无创信号采集方法。它具有较高的时间分辨率，在癫痫的检测、诊断、治疗、疗效评价和病理学研究中具有重要作用。目前基于头皮脑电图的癫痫诊断主要是癫痫发作的检测，然而癫痫发作并不常见。相比之下，近80％的癫痫患者在发作间期有神经元异常放电，这些在sEEG中可观测到的异常放电被称为发作间期痫样放电(Interictal Epileptiform Discharges,IEDs)。IED是癫痫诊断或癫痫发作风险评估的有力支持，也是确定癫痫起源潜在病灶的关键因素。因此，基于sEEG的IEDs检测对癫痫的诊断意义重大。
与癫痫发作不同，IEDs更为频繁并以各种形式出现在头皮脑电信号中，如棘波、尖波、棘慢复合波、多棘波复合波等，它们在sEEG中存在的时长不一(短的仅有20-70ms，长的持续1s以上)、极易与sEEG中无法通过滤波手段滤除的伪差(由咀嚼、传感器滑动、心跳等因素造成)混淆，使得sEEG中IEDs的检出具有极大的挑战性。
目前，临床上长程sEEG中IEDs的检测仍然需要由具备丰富经验的技师、医生进行人工标注。从长达数小时甚至超过一天的庞大信号中找寻毫秒级的异常片段，对于临床工作者而言耗时巨大，且其准确性难免会因为个人的主观经验及疲惫程度而产生较大的偏差。
在过去的几十年里，通过测量sEEG中感兴趣特征的相似性来自动描述和检测IED的方法被提出，如字典学习、模板匹配、子带分解等。 此外，也有研究者根据sEEG的时域、频域或非线性特征建立特征工程，用具有一个或多个特征的分类器实现IEDs的自动检测，如决策树、人工神经网络、梯度提升机等。虽然上述的IEDs自动检测方法取得了一定的效果，但在临床应用中还存在一些障碍。主要挑战如下：
(1)复杂的预处理。sEEG中含有大量无法被滤波算法滤除的伪差，利用上述传统方法很难准确地区分伪差与IEDs，因此传统方法中大多配有单独设计的伪差去除算法或人工伪差去除准则，工程量大且易损伤信号中包含的病理信息。
(2)不同患者中表现不佳。不同类型、不同年龄、不同性别的癫痫患者，IEDs的形态、时长等外在表现存在较大差异，不同医院之间，由于设备、参考电极、环境等因素的不同，临床记录可能存在较大差异。此外，某些患者的sEEG中可能存在一个或多个不良通道。这些原因导致传统方法中依靠信号处理经验设置的特征工程无法全面深入的挖掘sEEG特征，使得IEDs与背景信号的区分缺乏稳健性、难以应对实际临床应用中复杂多变的患者情况。
鉴于上述，本发明旨在提供一种痫患者发作间期痫样电活动检测装置和方法，来解决上述的一个或多个技术问题。
发明内容
为了解决现有技术中的一个或多个技术问题，根据本发明一方面，提供一种基于同时刻传感器空间脑电信号的特征分析的癫痫患者发作间期痫样电活动检测装置，其特征在于包括：
脑电信号传感器，用于采集患者的脑电信号，所述脑电信号为多导联头皮脑电信号；
脑电信号预处理单元，用于对所述脑电信号滤波和归一化处理；
脑电信号分割单元，用于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)＝[S1,S2,…,Sw] T，其中w为所述脑 电信号的传感器通道数，Sw为时长1s的单通道脑电信号片段构成的向量；以及
端到端的痫样电活动信号与背景信号二分类单元，用于根据输入的m个w*256的输入样本矩阵S (m)输出痫样电活动检测结果；
其中，所述端到端的痫样电活动信号与背景信号二分类单元包括：
U型语义分割结构单元，包括4个叠加的最大池化层和4个叠加的反卷积结构，每个最大池化层前叠加有两层1*3卷积层，每个反卷积结构后叠加有两层1*3卷积层，最上层的反卷积结构对应的两层1*3的卷积层之后叠加有1*1卷积层，该U型语义分割结构单元根据输入的m个w*256的输入样本矩阵S (m)输出特征向量F (m)＝[F1,F2,…,Fw] T；其中，Fw是第w个传感器通道的脑电信号经过U型语义分割结构映射后得到的特征；和
基于同时刻传感器空间脑电信号的特征分析单元，用于根据所述特征向量F (m)＝[F1,F2,…,Fw] T分析同时刻各传感器空间脑电信号的特征，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，进而判定1s时间窗的S (m)片段中是否含有痫样电活动，汇总m个片段的结果，实现整段脑电信号中痫样电活动的检测。
根据本发明又一方面，所述基于同时刻传感器空间脑电信号的特征分析单元包括：
门控循环单元，用于根据输入的w个传感器空间的同时刻的脑电信号的特征，以重复网络模块的链式形式对输入进行决策，输出各个 时刻相应的决策结果；
sigmod全连接网络，用于根据输入的所述各个时刻相应的决策结果，判定1s时间窗的S (m)片段中是否含有痫样电活动，汇总m个片段的结果，实现整段脑电信号中痫样电活动的检测。
根据本发明又一方面，所述脑电信号预处理单元包括：截止频率为0.5-45Hz的带通滤波器，用于对输入的所述脑电信号进行滤波；参考电极，设置于患者双侧耳垂处；归一化单元，用于对重参考后的脑电信号进行归一化处理。
根据本发明又一方面，所述脑电信号传感器的采样频率为256Hz。
根据本发明又一方面，所述脑电信号分割单元选择窗宽为1s且重叠度为50％的时间窗，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)。
根据本发明又一方面，还提供了一种基于同时刻传感器空间脑电信号的特征分析的癫痫患者发作间期痫样电活动检测方法，其特征在于包括以下步骤：
采集患者的脑电信号，所述脑电信号为多导联头皮脑电信号；
通过脑电信号预处理单元对所述脑电信号滤波和归一化处理；
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)＝[S1,S2,…,Sw] T，其中w为所述脑电信号的传感器通道数，Sw为时长1s的单通道脑电信号片段构成的向量；
通过端到端的痫样电活动信号与背景信号二分类单元将输入的m个w*256的输入样本矩阵S (m)进行分类并输出痫样电活动检测结果；
其中，所述通过端到端的痫样电活动信号与背景信号二分类单元将输入的m个w*256的输入样本矩阵S (m)进行分类并输出痫样电活动检测结果包括：
通过U型语义分割结构单元将输入的m个w*256的输入样本矩阵S (m)映射和输出为特征向量F (m)＝[F1,F2,…,Fw] T；其中，Fw是第w个传感器通道的脑电信号经过U型语义分割结构映射后得到的特征，所述U型语义分割结构单元包括4个叠加的最大池化层和4个叠加的反卷积结构，每个最大池化层前叠加有两层1*3卷积层，每个反卷积结构后叠加有两层1*3卷积层，最上层的反卷积结构对应的两层1*3的卷积层之后叠加有1*1卷积层；
根据所述特征向量F (m)＝[F1,F2,…,Fw] T分析同时刻各传感器空间脑电信号的特征，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，进而判定1s时间窗的S (m)片段中是否含有痫样电活动，汇总m个片段的结果，实现整段脑电信号中痫样电活动的检测。
根据本发明又一方面，通过门控循环单元将输入的w个传感器空间的同时刻的脑电信号的特征，以重复网络模块的链式形式对输入进行决策，输出各个时刻相应的决策结果；
利用sigmod全连接网络根据所述各个时刻相应的决策结果，判定1s时间窗的S (m)片段中是否含有痫样电活动，汇总m个片段的结果，实现整段脑电信号中痫样电活动的检测。
根据本发明又一方面，所述脑电信号预处理单元包括：截止频率为0.5-45Hz的带通滤波器，用于对输入的所述脑电信号进行滤波；参考电极，设置于患者双侧耳垂处；归一化单元，用于对重参考后的脑电信号进行归一化处理。
根据本发明又一方面，通过采样频率为256Hz的脑电信号传感器采集患者的脑电信号。
根据本发明又一方面，选择窗宽为1s且重叠度为50％的时间窗，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)。
与现有技术相比，本发明具有以下一个或多个技术效果：
(1)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。
(2)对于本发明使用的U型语义分割结构：首先，多层池化层的叠加为模型的特征映射提供了多尺度的观测视角，使得对于1s时间窗内无论是小到20ms的单个棘波还是大到充满整个时间窗的棘慢复合波，模型都能在sEEG中获取对应的形态学特征；其次，模型使用的跳跃式跨层连接使得模型能够保证足够深度挖掘sEEG中IED的特征信息的同时，不会被梯度***等因素所影响，在这之中浅层与深层特征的连接方式使用拼接的方式，网络能够根据IED与背景sEEG的差异性特征自适应地选择浅层与深层特征的比重，使得模型对单通道中不同形态、模式下的异常放电波形均有较好的识别能力；最后模型使用自适应的卷积填充方式完成对sEEG编码到解码的自动特征提取方式，在获取原始信号形态学特征的同时保留了信号中完好的时序及通道关联性，为进一步决策提供充足的信息。
(3)对于本发明使用的门控循环单元：模型创造性的将同一时刻不同传感器处在U型语义分割结构中获取的特征视作表征通道间关联关系的序列信息，使用门控循环单元获取所有传感器空间在1s时间窗内每个时刻(对应256个采样点中的每一个采样点)的脑电信号，从而分析决策检测到的形态学异常来自神经***的异常放电还是仅为随机扰动造成的伪差。本发明更新门与重置门对序列中每个元素信息的动态决策，使得sEEG中的不良通道为整体决策造成的影响大幅度降 低；同时，由于输出结果为门控循环单元对整体序列的综合决策，因此序列中元素的排布顺序不会对最终的决策造成重大影响，使得模型对不同环境下采集的sEEG具有更高的适用性。
附图说明
为了能够理解本发明的上述特征的细节，可以参照实施例，得到对于简要概括于上的发明更详细的描述。附图涉及本发明的优选实施例，并描述如下：
图1是根据本发明一种优选实施例的癫痫患者发作间期痫样电活动检测装置采用的端到端的痫样电活动信号与背景信号二分类单元结构示意图；
图2是本发明中1s时间窗截取痫样电活动(IEDs)的示意图。
具体实施方式
现在将对于各种实施例进行详细说明，这些实施例的一个或更多个实例分别绘示于图中。各个实例以解释的方式来提供，而非意味作为限制。例如，作为一个实施例的一部分而被绘示或描述的特征，能够被使用于或结合任一其他实施例，以产生再一实施例。本发明意在包含这类修改和变化。
在以下对于附图的描述中，相同的参考标记指示相同或类似的结构。一般来说，只会对于个别实施例的不同之处进行描述。除非另有明确指明，否则对于一个实施例中的部分或方面的描述也能够应用到另一实施例中的对应部分或方面。
实施例1
参见图1-2，其中，图1是根据本发明一种优选实施例的癫痫患者发作间期痫样电活动检测装置采用的端到端的痫样电活动信号与背景 信号二分类单元结构示意图；图2是本发明中1s时间窗截取痫样电活动(IEDs)的示意图。
根据本发明一种优选实施方式，提供一种基于同时刻传感器空间脑电信号的特征分析的癫痫患者发作间期痫样电活动检测装置，其特征在于包括：
脑电信号传感器，用于采集患者的脑电信号，所述脑电信号为多导联头皮脑电信号；
脑电信号预处理单元，用于对所述脑电信号滤波和归一化处理；
脑电信号分割单元(输入信号分割单元)，用于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)＝[S1,S2,…,Sw] T，其中w为所述脑电信号的传感器通道数，Sw为时长1s的单通道脑电信号片段构成的向量；以及
端到端的痫样电活动信号与背景信号二分类单元，用于根据输入的m个w*256的输入样本矩阵S (m)输出痫样电活动检测结果。
优选地，参见图1，所述端到端的痫样电活动信号与背景信号二分类单元包括：
U型语义分割结构单元，包括4个叠加的最大池化层和4个叠加的反卷积结构，每个最大池化层前叠加有两层1*3卷积层，每个反卷积结构后叠加有两层1*3卷积层，最上层的反卷积结构对应的两层1*3的卷积层之后叠加有1*1卷积层，该U型语义分割结构单元根据输入的m个w*256的输入样本矩阵S (m)输出特征向量F (m)＝[F1,F2,…,Fw] T；其中，Fw是第w个传感器通道的脑电信号经过U型语义分割结构映射后得到的特征；和
基于同时刻传感器空间脑电信号的特征分析单元(基于同时刻传感器空间信息的特征分析单元)，用于根据所述特征向量F (m)＝[F1,F2,…,Fw] T分析同时刻各传感器空间脑电信号的特征，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，进而判定1s时间窗的S (m)片段中是否含有痫样电活动，汇总m个片段的结果，实现整段脑电信号中痫样电活动的检测。
有利地，本发明无需人工设计伪差滤除单元及特征工程即可完成IEDs的自动检测，并在在符合临床应用环境的跨人群、多中心背景下实现高准确率、低假阳性率的IED检出，为临床工作者长程sEEG的判读减轻负担。
可以理解的是，易与痫样电活动混淆的这些滤波单元难以滤除的伪差种类很多，除了仅仅发生在某个或某几个传感器空间的偶发性伪差，还有一部分是发生在大多、甚至全部传感器空间上的。例如心电导致的伪差，这种伪差会在所有传感器空间上产生一个同样大小的、与痫样电活动中的棘波形状类似的瞬态变化。经过研究发现，它与真正痫样电活动在同时刻传感器空间上体现出的关联性不同。也就是说，伪差在传感器空间上的关联性是不符合癫痫的疾病特性的，例如，真正痫样电活动在同时刻传感器空间上体现出第一关联性，而心电导致的伪差在同时刻传感器空间上体现出异于第一关联性的第二关联性。
需要说明的是，本发明能够区分脑电信号中易与痫样电活动混淆的伪差只是目的之一，还有一个目的是使用的这种链式结构，可以获取不同传感器之间的关联性信息，并降低脑电中不良通道以及元素排布的影响，使得***拥有应用于不同场景及人群的潜力，即***在不同医疗中心中具有较高的稳健性。
根据本发明又一优选实施方式，所述基于同时刻传感器空间脑电 信号的特征分析单元包括：
门控循环单元，用于根据输入的w个传感器空间的同时刻的脑电信号的特征，以重复网络模块的链式形式对输入进行决策，输出各个时刻相应的决策结果；
sigmod全连接网络，用于根据输入的所述各个时刻相应的决策结果，判定1s时间窗的S (m)片段中是否含有痫样电活动，汇总m个片段的结果，实现整段脑电信号中痫样电活动的检测。
根据本发明又一优选实施方式，所述脑电信号预处理单元包括：截止频率为0.5-45Hz的带通滤波器，用于对输入的所述脑电信号进行滤波；参考电极，设置于患者双侧耳垂处；归一化单元，用于对重参考后的脑电信号进行归一化处理。
根据本发明又一优选实施方式，所述脑电信号传感器的采样频率为256Hz。
根据本发明又一优选实施方式，所述脑电信号分割单元选择窗宽为1s且重叠度为50％的时间窗，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)。
根据本发明又一优选实施方式，还提供了一种基于同时刻传感器空间脑电信号的特征分析的癫痫患者发作间期痫样电活动检测方法，其特征在于包括以下步骤：
采集患者的脑电信号，所述脑电信号为多导联头皮脑电信号；
通过脑电信号预处理单元对所述脑电信号滤波和归一化处理；
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)＝[S1,S2,…,Sw] T，其中w为所述脑电信号的传感器通道数，Sw为时长1s的单通道脑电信号片段构成的向量；
通过端到端的痫样电活动信号与背景信号二分类单元将输入的m个w*256的输入样本矩阵S (m)进行分类并输出痫样电活动检测结果；
其中，所述通过端到端的痫样电活动信号与背景信号二分类单元将输入的m个w*256的输入样本矩阵S (m)进行分类并输出痫样电活动检测结果包括：
通过U型语义分割结构单元将输入的m个w*256的输入样本矩阵S (m)映射和输出为特征向量F (m)＝[F1,F2,…,Fw] T；其中，Fw是第w个传感器通道的脑电信号经过U型语义分割结构映射后得到的特征，所述U型语义分割结构单元包括4个叠加的最大池化层和4个叠加的反卷积结构，每个最大池化层前叠加有两层1*3卷积层，每个反卷积结构后叠加有两层1*3卷积层，最上层的反卷积结构对应的两层1*3的卷积层之后叠加有1*1卷积层；
根据所述特征向量F (m)＝[F1,F2,…,Fw] T分析同时刻各传感器空间脑电信号的特征，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，进而判定1s时间窗的S (m)片段中是否含有痫样电活动，汇总m个片段的结果，实现整段脑电信号中痫样电活动的检测。
根据本发明又一优选实施方式，通过门控循环单元将输入的w个传感器空间的同时刻的脑电信号的特征，以重复网络模块的链式形式对输入进行决策，输出各个时刻相应的决策结果；
利用sigmod全连接网络根据所述各个时刻相应的决策结果，判定1s时间窗的S (m)片段中是否含有痫样电活动，汇总m个片段的结果，实现整段脑电信号中痫样电活动的检测。
根据本发明又一优选实施方式，所述脑电信号预处理单元包括：截止频率为0.5-45Hz的带通滤波器，用于对输入的所述脑电信号进行滤波；参考电极，设置于患者双侧耳垂处；归一化单元，用于对重参 考后的脑电信号进行归一化处理。
根据本发明又一优选实施方式，通过采样频率为256Hz的脑电信号传感器采集患者的脑电信号。
根据本发明又一优选实施方式，选择窗宽为1s且重叠度为50％的时间窗，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)。
根据本发明又一优选实施方式，还提供了一种头皮脑电中融合多尺度时域信息与传感器信息的癫痫患者发作间期痫样电活动检测方法，主要包含如下步骤：
步骤1.选择截止频率为0.5-45Hz的带通滤波器，对输入的原始多导联头皮脑电信号进行滤波，然后将患者双侧耳垂处电极(“A1”、“A2”)设为参考电极，计算重参考后的sEEG，并利用最大最小值法对其进行归一化处理；
步骤2.通过计算，调整sEEG采样频率至256Hz，然后选择窗宽为1s重叠度为50％的时间窗，对预处理后的sEEG信号进行分割，生成m个w*256的输入样本矩阵S (m)＝[S1,S2,…,Sw] T，其中w为重参考后的sEEG信号的通道数。Sw为时长1s的单通道脑电信号片段构成的向量；
步骤3.利用U型语义分割网络同步提取S (m)中单通道信号中IED的形态学特征，输出与之同尺寸的特征向量F (m)＝[F1,F2,…,Fw] T，然后利用门控循环单元分析F (m)中单通道sEEG在时间尺度上的IED特征并输出每个通道的决策结果，接着对相连的脑电通道使用门控循环单元挖掘传感器层面与IED相关的空间信息，以此最终判定1s时长的S (m)片段中是否含有IED事件，最后汇总m个片段的结果，即可实现整段 长程sEEG中IED事件的标注，从而减轻临床工作者的工作负担。
优选地，本发明的癫痫患者发作间期痫样电活动检测模型是一种端到端的实现sEEG中IEDs与背景信号二分类的模型，其结构图如图1所示。令输入模型的m个sEEG信号为S (m)＝[S1,S2,…,Sw] T，w∈[1，19]，U型语义分割结构的左侧，模型依靠4个叠加的最大池化层获得了逐级递增的4个不同尺度下的视野，并依靠每次池化前叠加的两层卷积核为1*3的卷积层提取当前视野下sEEG单个通道Sw的局部特征，并最终在U型语义结构的底端完成时域上不同尺度下sEEG的编码过程；U型语义分割结构的右侧，模型依靠与池化层数量相同的4个反卷积结构进行上采样，以此解码被模型编码后的sEEG单通道上的形态学特征，并恢复每个特征与原始信号在时域上的对应位置关系，实现1s时间窗内每个通道中的异常波形与背景的差异化表达F (m)＝[F1,F2,…,Fw] T，w∈[1，19]；U型语义分割结构获取的精细到采样点(即每一时刻)的特征表达会被送入之后的门控循环单元之中，门控循环单元以一种重复网络模块的链式形式对输入进行决策，此处来自19个传感器的信号依次输入门控循环单元，其中链式结构中的每一个模块的输出y t由当前模块的输入x t和之前模块的状态h t-1经由重置门r t与更新门z t计算后共同决策，具体计算如下：
r t＝σ(W r·[h t-1，x t])    (2)
z t＝σ(W z·[h t-1，x t])     (3)
Figure PCTCN2021140341-appb-000001
Figure PCTCN2021140341-appb-000002
y t＝σ(W y·h t)      (6)
其中[]为向量相接，*为矩阵的乘积，σ(·)为激活函数，W r，W z，
Figure PCTCN2021140341-appb-000003
W y分别为对应门在训练中学习到的权值矩阵。即门控循环单元最终的输出是19个传感器处获取的语义级形态特征动态加权后的综合决策结果。最终，时间域上19个传感器的综合决策结果构成的特征向量输入以sigmod为激活函数的全连接网络，得到当前1s时间窗内的样本是否含有发作间期痫样异常电活动的判别结果。
优选地，通过将本发明的检测方法与现有比较模型效果进行了量化评估，其中，一方面是利用平衡训练集进行跨人群的5折交叉验证，对比分析本发明与已有端到端检测模型的性能，另一方面是利用独立测试集，评估临床背景下本发明的可用性及稳健性。其中，量化评估结果如下。
表1不同模型在平衡训练集中5折交叉验证的平均检测结果
Figure PCTCN2021140341-appb-000004
表1中所展示的为本发明方法在跨人群的5折交叉验证中的模型评估结果，并列举了同样条件下典型卷积神经网络模型VGGNet，和本发明方法的IEDs检测效果进行对比。不难看出，本发明所使用的方法各项指标均取得最好的效果。对比VGGNet与UNet可得U型语义网络相较于经典卷积神经网络能够更好的获取时域中IEDs的多尺度形态学信息；同理，对比UNet与本发明方法可得，通过门控循环单元引入同时刻传感器信息进行分析决策，大幅提高了sEEG中区分IEDs与背景 信号的能力。
表2本发明方法在独立测试集中的IEDs检测效果
Figure PCTCN2021140341-appb-000005
表2所示为本发明方法用于独立测试集的IEDs检测效果。以1s时间窗及50％的重叠度将每例患者长达1小时的长程sEEG完整输入模型进行性能评估。与平衡训练集不同，独立测试集的每例样本中IEDs与背景信号的比例均出现严重失衡，差距最小的一例样本二者的比例也达到1：23，因此使得测试得到的精度和与精度密切相关的F1-Score效果低于5折交叉验证的结果。然而，模型高敏感度(即查全率)的优势能够保证尽可能完整的获取患者sEEG中IEDs，极大程度的避免了长程sEEG中大量的背景尤其是易与IEDs混淆的伪差给临床诊断带来的巨大工作量，为医生判读sEEG节省了大量时间及精力，也为没有优质医疗资源的偏远地区提供了优质的癫痫患者sEEG判读手段。
与现有技术相比，本发明具有以下一个或多个技术效果：
(1)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。
(2)对于本发明使用的U型语义分割结构：首先，多层池化层的叠加为模型的特征映射提供了多尺度的观测视角，使得对于1s时间窗内无论是小到20ms的单个棘波还是大到充满整个时间窗的棘慢复合波，模型都能在sEEG中获取对应的形态学特征；其次，模型使用的跳 跃式跨层连接使得模型能够保证足够深度挖掘sEEG中IED的特征信息的同时，不会被梯度***等因素所影响，在这之中浅层与深层特征的连接方式使用拼接的方式，网络能够根据IED与背景sEEG的差异性特征自适应地选择浅层与深层特征的比重，使得模型对单通道中不同形态、模式下的异常放电波形均有较好的识别能力；最后模型使用自适应的卷积填充方式完成对sEEG编码到解码的自动特征提取方式，在获取原始信号形态学特征的同时保留了信号中完好的时序及通道关联性，为进一步决策提供充足的信息。
(3)对于本发明使用的门控循环单元：模型创造性的将同一时刻不同传感器处在U型语义分割结构中获取的特征视作表征通道间关联关系的序列信息，使用门控循环单元获取所有传感器空间在1s时间窗内每个时刻(对应256个采样点中的每一个采样点)的脑电信号，从而分析决策检测到的形态学异常来自神经***的异常放电还是仅为随机扰动造成的伪差。本发明更新门与重置门对序列中每个元素信息的动态决策，使得sEEG中的不良通道为整体决策造成的影响大幅度降低；同时，由于输出结果为门控循环单元对整体序列的综合决策，因此序列中元素的排布顺序不会对最终的决策造成重大影响，使得模型对不同环境下采集的sEEG具有更高的适用性。
虽然前述内容是关于本发明的实施例，但可在不背离本发明的基本范围的情况下，设计出本发明其他和更进一步的实施例，本发明的范围由权利要求书确定。
上述实施例仅为本发明的较佳实施例而已，并不用以限制本发明，这些实施例中不互相违背的技术特征可彼此结合。凡在本发明的精神和原则之内所作的任何修改、等同替换和改进等，均应包含在本发明的保护范围之内。

Claims (10)

  1. 一种基于同时刻传感器空间脑电信号的特征分析的癫痫患者发作间期痫样电活动检测装置，其特征在于包括：
    脑电信号传感器，用于采集患者的脑电信号，所述脑电信号为多导联头皮脑电信号；
    脑电信号预处理单元，用于对所述脑电信号滤波和归一化处理；
    脑电信号分割单元，用于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)＝[S1,S2,…,Sw] T，其中w为所述脑电信号的传感器通道数，Sw为时长1s的单通道脑电信号片段构成的向量；以及
    端到端的痫样电活动信号与背景信号二分类单元，用于根据输入的m个w*256的输入样本矩阵S (m)输出痫样电活动检测结果；
    其中，所述端到端的痫样电活动信号与背景信号二分类单元包括：
    U型语义分割结构单元，包括4个叠加的最大池化层和4个叠加的反卷积结构，每个最大池化层前叠加有两层1*3卷积层，每个反卷积结构后叠加有两层1*3卷积层，最上层的反卷积结构对应的两层1*3的卷积层之后叠加有1*1卷积层，该U型语义分割结构单元根据输入的m个w*256的输入样本矩阵S (m)输出特征向量F (m)＝[F1,F2,…,Fw] T；其中，Fw是第w个传感器通道的脑电信号经过U型语义分割结构映射后得到的特征；和
    基于同时刻传感器空间脑电信号的特征分析单元，用于根据所述特征向量F (m)＝[F1,F2,…,Fw] T分析同时刻各传感器空间脑电信号的特征，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，进而判定1s时间窗的S (m)片段中是否含有痫样电活动，汇总m个片段的结果，实现整段脑电信号中痫样电活动的检测。
  2. 根据权利要求1所述的检测装置，其特征在于所述基于同时刻传感器 空间脑电信号的特征分析单元包括：
    门控循环单元，用于根据输入的w个传感器空间的同时刻的脑电信号的特征，以重复网络模块的链式形式对输入进行决策，输出各个时刻相应的决策结果；
    sigmod全连接网络，用于根据输入的所述各个时刻相应的决策结果，判定1s时间窗的S (m)片段中是否含有痫样电活动，汇总m个片段的结果，实现整段脑电信号中痫样电活动的检测。
  3. 根据权利要求2所述的检测装置，其特征在于所述脑电信号预处理单元包括：截止频率为0.5-45Hz的带通滤波器，用于对输入的所述脑电信号进行滤波；参考电极，设置于患者双侧耳垂处；归一化单元，用于对重参考后的脑电信号进行归一化处理。
  4. 根据权利要求1-3任一项所述的检测装置，其特征在于：所述脑电信号传感器的采样频率为256Hz。
  5. 根据权利要求4所述的检测装置，其特征在于：所述脑电信号分割单元选择窗宽为1s且重叠度为50％的时间窗，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)。
  6. 一种基于同时刻传感器空间脑电信号的特征分析的癫痫患者发作间期痫样电活动检测方法，其特征在于包括以下步骤：
    采集患者的脑电信号，所述脑电信号为多导联头皮脑电信号；
    通过脑电信号预处理单元对所述脑电信号滤波和归一化处理；
    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)＝[S1,S2,…,Sw] T，其中w为所述脑电信号的传感器通道数，Sw为时长1s的单通道脑电信号片段构成的向量；
    通过端到端的痫样电活动信号与背景信号二分类单元将输入的m个w*256的输入样本矩阵S (m)进行分类并输出痫样电活动检测结果；
    其中，所述通过端到端的痫样电活动信号与背景信号二分类单元将输入的m个w*256的输入样本矩阵S (m)进行分类并输出痫样电活动检测结果包括：
    通过U型语义分割结构单元将输入的m个w*256的输入样本矩阵S (m)映射和输出为特征向量F (m)＝[F1,F2,…,Fw] T；其中，Fw是第w个传感器通道的脑电信号经过U型语义分割结构映射后得到的特征，所述U型语义分割结构单元包括4个叠加的最大池化层和4个叠加的反卷积结构，每个最大池化层前叠加有两层1*3卷积层，每个反卷积结构后叠加有两层1*3卷积层，最上层的反卷积结构对应的两层1*3的卷积层之后叠加有1*1卷积层；
    根据所述特征向量F (m)＝[F1,F2,…,Fw] T分析同时刻各传感器空间脑电信号的特征，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，进而判定1s时间窗的S (m)片段中是否含有痫样电活动，汇总m个片段的结果，实现整段脑电信号中痫样电活动的检测。
  7. 根据权利要求6所述的检测方法，其特征在于通过门控循环单元将输入的w个传感器空间的同时刻的脑电信号的特征，以重复网络模块的链式形式对输入进行决策，输出各个时刻相应的决策结果；
    利用sigmod全连接网络根据所述各个时刻相应的决策结果，判定1s时间窗的S (m)片段中是否含有痫样电活动，汇总m个片段的结果，实现整段脑电信号中痫样电活动的检测。
  8. 根据权利要求7所述的检测方法，其特征在于所述脑电信号预处理单元包括：截止频率为0.5-45Hz的带通滤波器，用于对输入的所述脑电信号进行滤波；参考电极，设置于患者双侧耳垂处；归一化单元，用于对重参考后的脑电信号进行归一化处理。
  9. 根据权利要求6-8任一项所述的检测方法，其特征在于通过采样频率为256Hz的脑电信号传感器采集患者的脑电信号。
  10. 根据权利要求9所述的检测方法，其特征在于选择窗宽为1s且重叠度为50％的时间窗，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)。
PCT/CN2021/140341 2020-12-22 2021-12-22 癫痫患者发作间期痫样电活动检测装置和方法 WO2022135449A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202011529814.6 2020-12-22
CN202011529814.6A CN112674782B (zh) 2020-12-22 2020-12-22 癫痫患者发作间期痫样电活动检测装置和方法

Publications (1)

Publication Number Publication Date
WO2022135449A1 true WO2022135449A1 (zh) 2022-06-30

Family

ID=75450670

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/140341 WO2022135449A1 (zh) 2020-12-22 2021-12-22 癫痫患者发作间期痫样电活动检测装置和方法

Country Status (2)

Country Link
CN (2) CN113662560B (zh)
WO (1) WO2022135449A1 (zh)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113662560B (zh) * 2020-12-22 2022-05-24 北京航空航天大学 发作间期痫样放电的检测方法、存储介质和设备
CN113647962B (zh) * 2021-08-20 2023-09-22 天津大学 基于深度学习集成模型的癫痫定位与发作预测方法
CN114595731B (zh) * 2022-05-09 2022-08-19 南京信息工程大学 基于持续学习的非线性医疗传感器数据的语义分割方法

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8483816B1 (en) * 2010-02-03 2013-07-09 Hrl Laboratories, Llc Systems, methods, and apparatus for neuro-robotic tracking point selection
CN106821376A (zh) * 2017-03-28 2017-06-13 南京医科大学 一种基于深度学习算法的癫痫发作预警***及方法
CN109480833A (zh) * 2018-08-30 2019-03-19 北京航空航天大学 基于人工智能的癫痫患者脑电信号的预处理和识别方法
US10588561B1 (en) * 2017-08-24 2020-03-17 University Of South Florida Noninvasive system and method for mapping epileptic networks and surgical planning
CN111080575A (zh) * 2019-11-22 2020-04-28 东南大学 一种基于残差密集u形网络模型的丘脑分割方法
CN111643075A (zh) * 2020-01-17 2020-09-11 天津大学 基于微状态分析方法的eeg实时检测分析平台
CN111772585A (zh) * 2020-07-13 2020-10-16 清华大学 间期颅内脑电信号中癫痫异常群体活动的数据提取方法
CN112674782A (zh) * 2020-12-22 2021-04-20 北京航空航天大学 癫痫患者发作间期痫样电活动检测装置和方法

Family Cites Families (16)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7502650B2 (en) * 2003-09-22 2009-03-10 Cvrx, Inc. Baroreceptor activation for epilepsy control
CN102488513B (zh) * 2011-12-05 2013-08-28 中国人民解放军第四军医大学 基于偶极（电荷）源时空模型的脑电源定位装置
CN102613971B (zh) * 2012-03-30 2013-10-30 北京品驰医疗设备有限公司 基于脑电的癫痫检测与干预装置
CN103932796A (zh) * 2014-04-13 2014-07-23 北京师范大学 一种基于多模态医学影像数据融合的颅内电极个体化定位的方法
CN105708452B (zh) * 2016-01-01 2018-12-04 孟艳林 头皮脑电的确定电极放置位置的装置及方法
US10589096B2 (en) * 2016-04-25 2020-03-17 The Board Of Regents Of The University Of Texas System Systems and methods for network based neurostimulation of cognitive processes
CN109645990B (zh) * 2018-08-30 2020-11-27 北京航空航天大学 一种癫痫患者脑电信号的计算机模式识别方法
AU2020219705A1 (en) * 2019-02-08 2021-09-16 Nanyang Technological University Method and system for seizure detection
CN210228524U (zh) * 2019-05-30 2020-04-03 首都医科大学宣武医院 一种癫痫患者专用床
CN110338786B (zh) * 2019-06-28 2020-10-02 北京师范大学 一种癫痫样放电的识别与分类方法、***、装置和介质
CN111134666B (zh) * 2020-01-09 2021-05-11 中国科学院软件研究所 一种多通道脑电数据的情绪识别方法及电子装置
CN111387974B (zh) * 2020-02-19 2022-12-02 杭州电子科技大学 基于深度自编码的脑电特征优化与癫痫发作检测方法
CN111150393B (zh) * 2020-02-19 2023-03-28 杭州电子科技大学 基于lstm多通道的脑电癫痫尖棘波放电联合检测方法
CN111419222B (zh) * 2020-03-08 2023-04-14 杭州电子科技大学 基于bni的癫痫发作信号的检测方法
CN111557662B (zh) * 2020-05-15 2023-04-18 灵犀医学科技(北京)有限公司 癫痫间期脑电信号的处理方法、装置、存储介质及设备
CN111568414B (zh) * 2020-05-22 2022-05-10 中国人民解放军东部战区总医院 基于功能影像学的癫痫活动检测方法及***

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US8483816B1 (en) * 2010-02-03 2013-07-09 Hrl Laboratories, Llc Systems, methods, and apparatus for neuro-robotic tracking point selection
CN106821376A (zh) * 2017-03-28 2017-06-13 南京医科大学 一种基于深度学习算法的癫痫发作预警***及方法
US10588561B1 (en) * 2017-08-24 2020-03-17 University Of South Florida Noninvasive system and method for mapping epileptic networks and surgical planning
CN109480833A (zh) * 2018-08-30 2019-03-19 北京航空航天大学 基于人工智能的癫痫患者脑电信号的预处理和识别方法
CN111080575A (zh) * 2019-11-22 2020-04-28 东南大学 一种基于残差密集u形网络模型的丘脑分割方法
CN111643075A (zh) * 2020-01-17 2020-09-11 天津大学 基于微状态分析方法的eeg实时检测分析平台
CN111772585A (zh) * 2020-07-13 2020-10-16 清华大学 间期颅内脑电信号中癫痫异常群体活动的数据提取方法
CN112674782A (zh) * 2020-12-22 2021-04-20 北京航空航天大学 癫痫患者发作间期痫样电活动检测装置和方法
CN113662560A (zh) * 2020-12-22 2021-11-19 北京航空航天大学 发作间期痫样放电的检测方法、存储介质和设备

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
FAHOUM FIRAS, ZELMANN RINA, TYVAERT LOUISE, DUBEAU FRANÇOIS, GOTMAN JEAN: "Epileptic Discharges Affect the Default Mode Network – fMRI and Intracerebral EEG Evidence", PLOS ONE, vol. 8, no. 6, 1 June 2016 (2016-06-01), pages 1 - 12, XP055946550, DOI: 10.1371/journal.pone.0068038 *

Also Published As

Publication number Publication date
CN113662560B (zh) 2022-05-24
CN113662560A (zh) 2021-11-19
CN112674782B (zh) 2021-11-26
CN112674782A (zh) 2021-04-20

Similar Documents

Publication Publication Date Title
WO2019100565A1 (zh) 基于人工智能自学习的动态心电图分析方法和装置
Liang et al. Scalp EEG epileptogenic zone recognition and localization based on long-term recurrent convolutional network
WO2022135449A1 (zh) 癫痫患者发作间期痫样电活动检测装置和方法
Wang et al. Arrhythmia classification algorithm based on multi-head self-attention mechanism
WO2019100560A1 (zh) 基于人工智能自学习的心电图自动分析方法和装置
CN113786204A (zh) 基于深度卷积注意力网络的癫痫颅内脑电信号预警方法
EP3698709A1 (en) Electrocardiogram information processing method and electrocardiogram workstation system
Wu et al. A novel method to detect multiple arrhythmias based on time-frequency analysis and convolutional neural networks
EP3698708A1 (en) Ecg information processing method and ecg workstation
Lopes et al. Automatic electroencephalogram artifact removal using deep convolutional neural networks
Zacharaki et al. Spike pattern recognition by supervised classification in low dimensional embedding space
do Vale Madeiro et al. Evaluation of mathematical models for QRS feature extraction and QRS morphology classification in ECG signals
Wang et al. A novel multi-scale dilated 3D CNN for epileptic seizure prediction
Migliorelli et al. SGM: a novel time-frequency algorithm based on unsupervised learning improves high-frequency oscillation detection in epilepsy
Li et al. An improved sparse representation over learned dictionary method for seizure detection
CA3065208C (en) Long qt syndrome diagnosis and classification
Subramanyan et al. A novel deep neural network for detection of Atrial Fibrillation using ECG signals
Yang et al. Automated intra-patient and inter-patient coronary artery disease and congestive heart failure detection using EFAP-Net
CN113208613A (zh) 基于fhls特征选择的多模态bci时序优化方法
Yun-Mei et al. The abnormal detection of electroencephalogram with three-dimensional deep convolutional neural networks
CN110200624A (zh) 基于卷积神经网络-循环神经网络-支持向量机混合模型的疾病识别算法
CN113598792B (zh) 基于有监督特征融合算法的癫痫脑电分类方法
Zhang et al. Cross-patient automatic epileptic seizure detection using patient-adversarial neural networks with spatio-temporal EEG augmentation
CN114532994A (zh) 一种基于卷积变分自编码器的无监督脑电高频振荡信号自动检测方法
Kovalchuk et al. A novel feature vector for ECG classification using deep learning

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 21909445

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21909445

Country of ref document: EP

Kind code of ref document: A1