WO2023098303A1 - Real-time epileptic seizure detecting and monitoring system for video electroencephalogram examination of epilepsy - Google Patents

Abstract

A real-time epileptic seizure detecting and monitoring system for video electroencephalogram examination of epilepsy, comprising a data processing module, a model selection module, an epileptic seizure alarm module, and an interactive collaborative annotation module. The data processing module is configured to process input electroencephalogram signal data into a data form required by an epileptic seizure detection model; the model selection module is configured to construct and select an epileptic seizure detection model for the current patient; the epileptic seizure alarm module is configured to detect epileptic seizure and give an alarm; and the interactive collaborative annotation module is used for interaction among the epileptic seizure alarm module, accompanying persons, and a monitoring room. According to the real-time epileptic seizure detecting and monitoring system for video electroencephalogram examination of epilepsy, for practical problems in clinical video electroencephalogram examination, functional modules are designed such that the feature learning ability and training speed of the epileptic seizure detection model and the accuracy of real-time epileptic seizure detection are improved, and the experience of accompanying persons and doctors is greatly improved.

Description

Real-time detection and monitoring system for epileptic seizures oriented to video EEG examination of epilepsy technical field

The invention relates to the field of medical care information, in particular to a real-time detection and monitoring system for epileptic seizures oriented to epilepsy video electroencephalogram examination.

Background technique

Epilepsy, one of the most common neurological disorders, is characterized by frequent sudden surges of abnormal electrical activity in parts or entire brain regions. An EEG is an electrical recording of brain activity and is an important basis for the diagnosis and analysis of epilepsy. Video EEG combines scalp EEG and high-definition video to simultaneously record brain electrical activity and patient behavior. It has the characteristics of high temporal resolution and non-invasiveness, and is currently the most commonly used epilepsy detection method. The video EEG examination lasts for a long time and usually needs to include the wake-sleep-wake process, usually more than 4 hours, and can be as long as 24 hours. Existing technology does not fully integrate and utilize the resources of accompanying personnel, systems and medical personnel. Faced with long-term video EEG data, doctors need to spend a lot of time on visual inspection. Generally, escorts are required to mark the patient's epileptic seizures so that they can quickly locate the seizure moment during visual inspection. The accompanying escorts bring a great burden to the accompanying staff, and are limited by the identification ability of the accompanying staff. In reality, there will be many problems such as missing labels, wrong labels, and large time delays. And the current pre-labeling by the machine through the algorithm has too many alarms and is of limited help to doctors.

At present, there are mainly two types of technical solutions for the detection of epileptic seizures. One is based on traditional machine learning methods, using signal processing methods to extract features, screening effective features, and classifying through traditional machine learning methods. It relies too much on artificial feature selection, which is difficult Adapt to patients with different types of epilepsy; the other uses deep learning to simply preprocess the EEG time signal (filtering, denoising, and dividing it into segments of a certain length of time), that is, through a deep neural network (especially a convolutional neural network) ) to train the classification model. The current technical solution directly adopts the time-series representation when using the deep neural network for epilepsy detection, and does not fully consider the electrode placement according to the international 10-20 system, and the inherent spatial position relationship between the electrodes. The channels of the EEG are arranged in a simple one-dimensional order, and part of the position information is not fully utilized, and the relative spatial relationship between the channels is not reflected, resulting in insufficient performance of the seizure detection model. Epilepsy has a wide range of diseases, individual differences are large, and there are many forms of EEG. Because there are various types of epilepsy and individual differences among patients, the existing technology relies on the training model of individual historical EEG signal data to achieve the detection effect (96% recall rate). Low (about 75%); Judging from the results of technical solutions for epilepsy detection in published journals, the current technical solutions need to be based on patient-specific models to achieve better results, and the generalization performance of the models is weak.

Contents of the invention

The purpose of the present invention is to address the deficiencies in the prior art, and propose a real-time detection and monitoring system for seizures oriented to epileptic video EEG examination, which is used to solve the problem of real-time detection of epileptic seizures during the video EEG examination process of epileptic patients. And the problem of alarming and recording the moment of the patient's seizure after detecting the patient's seizure.

The purpose of the present invention is achieved through the following technical solutions: a real-time detection and monitoring system for epileptic seizures oriented to epilepsy video electroencephalogram examination, the system comprising:

Data processing module: used to input the patient's EEG signal data, divide the EEG signal data into several fixed-length data segments according to time, and perform wavelet decomposition on each channel of the segmented EEG signal data, according to the EEG electrode position , get 18 EEG signal channels, and arrange the 18 EEG signal channels according to the two-dimensional structure of 4 rows and 5 columns, and the missing 2 channels in the middle column are calculated according to the relative distance relationship between the channel positions; the output is two The decomposed data of the signal arranged in the dimensional structure;

Model selection module: build a model with z+3 layer structure, where the first z layer corresponds to the number of layers of wavelet decomposition in the data processing module, the input of the first layer of the model is the output of the last layer of wavelet decomposition; The input of the second layer is the output of the first layer of the model and the high frequency part decomposed by the penultimate layer of wavelet decomposition; the input of the third layer of the model is the output of the second layer of the model and the high frequency part of the penultimate layer of wavelet decomposition part; and so on, the input of the zth layer of the model is the output of the z-1st layer of the model and the high-frequency part decomposed by the first layer of wavelet decomposition; the z+1 layer, z+2 layer and z+3 layer are the volume Multilayer, maximum pooling layer and fully connected layer, the output is the probability of no epilepsy and the probability of seizure at the current moment; according to the data of the patient database in the hospital, the epilepsy types that have exceeded the threshold number of times in the diagnosis results are counted, and the epilepsy types are corresponding According to the EEG signal data of the current patient's video EEG examination, the corresponding epilepsy type is closest to the epilepsy type, and the corresponding seizure detection model is selected for detection.

Epilepsy alarm module: used to obtain the probability of epileptic seizures output by the epileptic seizure detection model in the model selection module in real time, smooth the seizure probability within a period of time, judge whether to alarm the accompanying staff according to the set threshold, and report to the alarm Stop working until the escort confirms that the epilepsy has stopped or resumes work after the set time has passed;

Interactive collaborative labeling module: used for the interaction between the epileptic seizure alarm module, the escort and the monitoring room; the escort judges the false positives and false negatives of the epileptic seizure alarm module, and reviews it through the monitoring room. If it is confirmed, the epileptic seizure alarm module is judged If there are false positives and false negatives, adjust the alarm threshold.

Further, the patient's EEG signal data input by the data processing module includes data used for model training from the patient database and multi-channel EEG signal data of real-time video EEG examination.

Further, according to the position of the EEG electrodes, 18 EEG signal channels are obtained as [Fp1-F7, F7-T3, T3-T5, T5-O1, Fp1-F3, F3-C3, C3-P3, P3-O1, Fz-Cz, Cz-Pz, Fp2-F4, F4-C4, C4-P4, P4-O2, Fp2-F8, F8-T4, T4-T6, T6-O2]; arrange 18 channels in 4 rows of 5 The two-dimensional structure of the column, the first column is [Fp1-F7,F7-T3,T3-T5,T5-O1], the second column is [Fp1-F3,F3-C3,C3-P3,P3-O1], The third column is [Fz-Cz,Cz-Pz], the fourth column is [Fp2-F4,F4-C4, C4-P4,P4-O2], and the fifth column is [Fp2-F8,F8-T4,T4 -T6,T6-O2], since there are only 2 channels in the middle column, add a channel before and after, denoted by N1 and N2, that is, the third column is [N1,Fz-Cz,Cz-Pz,N2], N1 and N2 It is determined according to the relative distance relationship between electrodes.

Further, the distance between Fp1-F3, Fp2-F4 and Fp1-F7, Fp2-F8 to N1 is estimated to be 1:2, considering the exponential attenuation of signal propagation, so the ratio of the two coefficients when calculating N1 is 4:1, and N2 is the same as The calculation formula of N1 and N2 is as follows:

Further, the convolution kernel of the z+1th convolutional layer in the model constructed by the model selection module is set according to the data form of the two-dimensional structure arrangement output by the data processing module, since the two-dimensional structure arrangement is 4*5 form, so the structure of the convolution kernel of the z+1th layer convolutional layer and the relevant part of the EEG signal channel is set to 2*2, and the pooling kernel of the z+2th layer maximum pooling layer neutralizes the EEG signal channel The structure of the relevant part is set to 3*2.

Further, according to the data of the patient database in the hospital, calculate the variance of the signal data of each frequency band of the 18-channel EEG signal data of each patient in the interictal period, where each frequency band corresponds to the frequency band of the wavelet decomposition input by the front z layer of the model , get a matrix representing the characteristics of EEG signals, and use the diagnosed epilepsy type as a label; when the current patient starts to undergo video EEG examination, calculate the signal data of each frequency band of the 18 channels of EEG signal data during the interictal period Variance, to obtain a matrix representing the characteristics of the EEG signal of the current patient, use the k-nearest neighbor algorithm to determine the closest epilepsy type, and select the corresponding seizure detection model.

Further, in the k-nearest neighbor algorithm, the calculation formula of the distance between the current patient and the patient in the patient bank in the hospital is as follows:

where d _i represents the distance corresponding to the i-th channel.

Further, the epileptic seizure alarm module smoothes the τ seizure probability values within a period of time, and sets the threshold to 0.9, and if it exceeds the threshold, an alarm is sent to the accompanying personnel; the specific implementation process is:

Among them, P(t) represents the smoothing result of the seizure alarm module at time t, p(t-j) represents the probability of seizures output by the seizure detection model at time t-j, t represents the current moment, and j represents the time interval.

Further, the interactive collaborative labeling module raises the alarm threshold when the epileptic seizure alarm module makes a false alarm, and the specific implementation process is:

r'=min(0.95, r+5(1-r)(1-e ^rP(t) )+0.01)

When the epileptic seizure alarm module fails to report, the alarm threshold is lowered; the specific implementation process is:

r'=max(0.8, r+5(1-r)(1-e ^rP(t) )-0.01)

Among them, r' represents the adjusted alarm threshold, min() represents the function of taking the minimum value, r represents the alarm threshold before adjustment, P(t) represents the smoothing result of the seizure alarm module at time t, and t represents the current time.

Further, when the interactive collaborative labeling module confirms that the total number of false positives or false positives of the seizure alarm module exceeds 3 times, the current seizure detection model is abandoned, and the process is performed again according to the EEG signal data of the patient's interictal period in the most recent period of time. Model selection.

Beneficial effects of the present invention:

1. Arrange the EEG channels in a two-dimensional plane according to their relative positions in the physical space, improve the efficiency of training and testing, reduce the time during training and testing, and meet the needs of real-time detection.

2. According to the historical data, the EEG type of the patient is divided into several categories, and the models are trained separately. When a new patient needs to detect epileptic seizures in real time, the model with the closest category of EEG signals is used as the epileptic seizure detection model to improve epilepsy. The accuracy of seizure detection improves the experience of doctors and accompanying staff.

3. Construct a convolutional neural network model based on wavelet decomposition, share parameters between different channels, and learn from the last layer of wavelet decomposition in the training process to fully learn the characteristics of different scales of EEG signals and improve seizure detection Model performance.

4. Form a collaborative marking between the escort, the detection system and the monitoring room. On the basis of the automatic detection and alarm of epileptic seizures, introduce the interactive collaborative marking of the accompanying staff and the monitoring room to improve the quality of epileptic seizure marking and improve the doctor's efficiency. Visual reading efficiency.

Description of drawings

Fig. 1 is the flow chart of the real-time detection and monitoring system facing epilepsy video EEG examination;

Figure 2 is a schematic diagram of the electrode position of the international 10-20 system;

Fig. 3 is the schematic diagram of wavelet decomposition process;

Figure 4 is a schematic diagram of the first five layers of the model structure;

Fig. 5 is the working flow chart of the interactive collaborative labeling module after the automatic alarm of the epileptic seizure alarm module;

Fig. 6 is a workflow diagram of the interactive collaborative labeling module after the accompanying staff actively alarms.

Detailed ways

The specific embodiments of the present invention will be described in further detail below in conjunction with the accompanying drawings.

As shown in Figure 1, a kind of real-time detection and monitoring system of epileptic seizures oriented to epilepsy video EEG examination provided by the present invention, the system includes: a data processing module, a model selection module, an epileptic seizure alarm module and an interactive collaborative labeling module; details as follows:

1. Data processing module, the original input data comes from the patient database or the multi-channel EEG signal of real-time video EEG examination, the data processing module is used to divide the data into 1s length segments, decompose the data by discrete wavelet transform and Arrange the EEG signal channels according to the two-dimensional structure; specifically:

The present invention adopts the EEG electrode placement standard stipulated by the international 10-20 system, including Fp1, F7, T3, T5, O1, F3, C3, P3, Fz, Cz, Pz, Fp2, F4, C4, P4, O2, F8 , T4, T6 have a total of 19 electrode positions, as shown in Figure 2, using the potential difference signals of adjacent double electrodes, we can get [Fp1-F7, F7-T3, T3-T5, T5-O1, Fp1-F3, F3- C3,C3-P3,P3-O1,Fz-Cz,Cz-Pz,Fp2-F4,F4-C4,C4-P4,P4-O2,Fp2-F8,F8-T4,T4-T6,T6-O2] 18 channel information.

1.1 Data Segmentation

Segment the EEG signal with a moving window with a time length of 1s, and the signal sampling frequency is 256Hz to obtain 256*18 two-dimensional signal segments.

1.2 Wavelet decomposition of signal

For the EEG signal x of each channel, the wavelet function and scale function of the discrete wavelet transform are used to decompose the signal, and the input EEG signal is decomposed into two parts: high-frequency details and low-frequency approximation, and continuously The low-frequency approximation part is repeatedly operated, subdivided into the fifth layer, and the length of the EEG signal data obtained by each decomposition is half of the original. For example, the length of the EEG signal data is changed from 256 to 128 in the first decomposition. As shown in Figure 3, the initial EEG signal of each channel is decomposed to obtain D1, A1, D2, A2, D3, A3, D4, A4, D5, A5, respectively representing 64-128Hz, 0-64Hz, 32-64Hz, 0-32Hz, 16-32Hz, 0-16Hz, 8-16Hz, 0-8Hz, 4-8Hz, 0-4Hz a group of signals (x, D1, A1, D2, A2, D3, A3, D4, A4, D5, A5), the data lengths are (256, 128, 128, 64, 64, 32, 32, 16, 16, 8, 8) respectively.

1.3 Two-dimensional arrangement of channels

According to the currently commonly used method, the channels of the EEG are arranged one-dimensionally in a simple order to form 18 signal channels, namely [Fp1-F7, F7-T3, T3-T5, T5-O1, Fp1-F3, F3 -C3,C3-P3,P3-O1,Fz-Cz,Cz-Pz,Fp2-F4,F4-C4,C4-P4,P4-O2,Fp2-F8,F8-T4,T4-T6,T6-O2 ]. In the current application, the channel position is not processed, but recorded in a fixed order. In order to fully reflect the spatial relative positional relationship of the channels, the present invention arranges 18 channels into a two-dimensional structure of 4 rows and 5 columns according to the form of Table 1, and the first column is [Fp1-F7, F7-T3, T3-T5, T5 -O1], the second column is [Fp1-F3,F3-C3,C3-P3,P3-O1], the third column is [Fz-Cz,Cz-Pz], the fourth column is [Fp2-F4,F4 -C4,C4-P4,P4-O2], the fifth column is [Fp2-F8,F8-T4,T4-T6,T6-O2], because there are only 2 channels in the middle column, add a channel before and after, use N1 and N2, that is, the third column is [N1, Fz-Cz, Cz-Pz, N2], N1 and N2 are calculated according to the adjacent electrodes respectively, as the filling value of the vacancy, the calculation formula of N1 and N2 is as follows, Convert 18 linearly recorded signal channels into a two-dimensional channel arrangement structure of 4 rows and 5 columns.

As shown in Figure 2, considering the relative distance relationship between electrodes, the distance between Fp1-F3, Fp2-F4 and Fp1-F7, Fp2-F8 to the middle vacant channel is about 1:2, considering the exponential attenuation of signal propagation , so the ratio of the two coefficients is 4:1 when calculating N1, and the same is true for N2.

The present invention adopts the two-dimensional arrangement of EEG signal channels, which can make full use of the relative positional relationship between the channels, combined with the convolutional neural network, can reduce model training parameters, improve the speed of model training, and effectively solve the problem of the present invention based on different Seizure types train the seizure detection model, the corresponding seizure training samples are reduced, and it is necessary to train multiple models in reality.

Table 1 Spatial arrangement table of signal channel of the present invention

Fp1-F7	Fp1-F3	N1	Fp2-F4	Fp2-F8
F7-T3	F3-C3	Fz-Cz	F4-C4	F8-T4
T3-T5	C3-P3	Cz-Pz	C4-P4	T4-T6
T5-O1	P3-O1	N2	P4-O2	T6-O2

2. Model selection module, design a model structure suitable for a series of data after wavelet decomposition, train the seizure detection model for the patients with the highest frequency of epileptic seizures to form a seizure detection model set, and according to the EEG signal of the patient Signal characteristics, using the k-nearest neighbor algorithm to select a seizure detection model suitable for the current patient from the model set;

2.1 Model Construction

The model construction is based on the EEG signal data processed by wavelet decomposition and channel two-dimensional arrangement. The construction process is shown in Figure 4, specifically:

1) The input of the model is based on the output of the data processing module, with the high-frequency details D1, D2, D3, D4, D5 of each order wavelet and the low-frequency approximation A5 of the last order wavelet as the input of the model;

2) The input of the first layer of the model is D5, A5, the data length is 8, and the signal channel is a two-dimensional structure of 4*5, so the structure of the input data of the first layer is (2*8)*4*5, volume The product parameter matrix is (2*2)*1*1@4, where @ indicates that the subsequent number is the number of features extracted by this layer, the step size is (1, 2, 1, 1), and the structure of the calculation result is ( 1*4)*4*5*4, keep the channel structure unchanged, adjust the signal structure, as the output C5 of the first layer, and also part of the input of the next layer, its structure is (1*16)*4 *5;

3) The input of the second layer of the model is the input C5 and D4 of the first layer, the data length is 16, and the signal channel is still a two-dimensional structure of 4*5, so the structure of the input data of the second layer is (2*16) *4*5, the convolution parameter matrix is (2*2)*4*5@4, the step size is (1, 2, 1, 1), and the structure of the calculation result is (1*8)*4*5* 4. Keep the channel structure unchanged, adjust the signal structure, as the output C4 of the second layer, and also part of the input of the next layer, its structure is (1*32)*4*5;

4) The structure of the third to fifth layers of the model can be analogized according to the second layer above, and the output of the third output C3, the fourth output C2, and the fifth output C1 are respectively (1*64)*4*5 , (1*128)*4*5, (1*256)*4*5;

5) The input of the sixth layer of the model is the output C1 of the fifth layer, (1*256)*4*5, and the dimension with a length of 1 is removed, which is 256*4*5, and the convolution parameter matrix is 8*2* 2@4, the step size is set to (8,1,1), and the output structure is 16*3*4*4;

6) The seventh layer of the model is the maximum pooling layer of (1,3,2,1), whose input is the output of the sixth layer, and the structure of the calculation result is (16*1*2*4), which is further converted to A one-dimensional vector with a length of 128, input to the last layer;

7) The last layer of the model is a fully connected layer, the input is the output of the previous layer, and the output is a vector with a length of 2, which is normalized by the softmax function, which respectively represent the probability of no epilepsy and the probability of seizure at the current moment;

8) This model uses the modified linear unit function ReLU as the activation function, and the last layer uses the softmax function to normalize the output. The training process uses cross-entropy as the loss function, as shown in the following formula.

Where L represents the loss function, where M represents the number of samples for each training, i represents the i-th sample, and y ⁱ represents the true category of the i-th sample,

Indicates the probability that the i-th sample is predicted to be a positive class.

2.2 Model selection

According to the existing patient database data in the hospital, including the EEG signal data and EEG report of the patient's previous visits, the epilepsy types that have exceeded the threshold number of times in the statistical diagnosis results are trained separately according to the patient's EEG signal data corresponding to the epilepsy type Corresponding models, where the threshold value is determined according to the model accuracy obtained through training, and generally takes a value of 15 to form a set of models. Calculate 6 frequency band signals of 0-4Hz, 4-8Hz, 8-16Hz, 16-32Hz, 32-64Hz, 64-128Hz of 18 channels of EEG signal data of each patient in this group of models A 18*6 matrix is obtained to represent the EEG signal features, and the epilepsy type diagnosed in the EEG report is used as its label to calculate the mean and variance of the EEG signal data in each frequency band of each channel.

Among them, x[i] represents the value of the i-th sampling point of the EEG signal data of each frequency band of each channel, n represents the length of each frequency band EEG signal data x of each channel, and avr represents the value of each channel The mean value of the EEG signal data of each frequency band, and var represents the variance of the EEG signal data of each frequency band of each channel.

When a new patient needs to start a video EEG examination, after the normal recording of the EEG signal, take the 5-minute data of the interval between seizures, calculate the variance of the 6 frequency band signals of each channel, and obtain a representation of the current patient's brain. The matrix of electrical signal features uses the k-nearest neighbor algorithm (the k value of the present invention is set to 8) to determine which EEG data of the current patient is closer to the EEG data of the patient with epilepsy type, and corresponds to the closest class , as a seizure detection model for the current patient. The formula for calculating the distance between two patients in the k-nearest neighbor algorithm is:

Among them, d _i represents the distance corresponding to the i-th channel, var1 represents the variance of the 0-4Hz frequency band of the current patient, var2 represents the variance of the 4-8Hz frequency band of the current patient, and var6 represents the variance of the 64-128Hz frequency band of the current patient. var1' to var6' represent the variance of the corresponding frequency bands of the patients in the patient database used for comparison. distance is the distance used for k-nearest neighbor algorithm comparison.

3. Epilepsy alarm module, used for the current patient's epileptic seizure, the system of the present invention promptly alarms the monitoring room and the accompanying personnel, and can also be initiated by the accompanying personnel to send an alarm to the monitoring room;

The epileptic seizure detection model outputs in real time the possibility that the current second is judged to be an epileptic seizure according to the patient's EEG. The present invention comprehensively considers the sensitivity of epileptic seizure detection, the false alarm rate and real-time performance, and smoothes the probability value output for 3 consecutive seconds. The threshold is set to 0.9. If the threshold is exceeded, the system will alarm the accompanying personnel and the monitoring room. After the alarm is issued, the alarm module will enter a 3-minute immunity time, and the alarm will not be repeated during this period.

Among them, P(t) represents the smoothing result of the epileptic seizure alarm module at time t, which is the possibility of epileptic seizure events. The larger P(t) is, the greater the possibility of epileptic seizure events is, p(t) represents The seizure probability output by the seizure detection model at time t.

4. The interactive collaborative labeling module is used for the human-computer interaction between the epileptic seizure alarm module, the accompanying personnel and the monitoring room in the system of the present invention, involving the coordination of the system's automatic alarm and manual confirmation, and the feedback of manual active marking to system detection adjust;

The interactive collaborative labeling module is mainly the human-computer interaction between the escort and the monitoring room and the seizure alarm module. The main body is the interaction between the escort and the seizure alarm module, and the monitoring room plays the role of confirmation and security.

As shown in Figure 5, when the epileptic seizure alarm module detects an epileptic seizure, it will automatically send an alarm to the accompanying personnel and the duty room, prompting the accompanying personnel to keep attention within the next 3 minutes, and at the same time, among all the monitoring objects in the monitoring room, Prompt that the patient is in an epileptic state, keep attention. If there is indeed an epileptic seizure, record the symptoms of the patient's epileptic seizure; if no epileptic seizure is observed, a false alarm message will be sent to the monitoring room, and the monitoring room will record and confirm the situation 3 minutes before checking. Then through the interactive collaborative labeling module, the alarm threshold of the epileptic seizure alarm module is raised, and this false alarm is recorded. The specific implementation process of raising the alarm threshold is as follows:

r'=min(0.95, r+5(1-r)(1-e ^rP(t) )+0.01)

Among them, r' represents the adjusted alarm threshold, min() represents the function of taking the minimum value, r represents the alarm threshold before adjustment, P(t) represents the smoothing result of the seizure alarm module at time t, and t represents the current time.

If the monitoring room confirms that there is indeed an epileptic seizure, the seizure time will be recorded. If the monitoring room cannot confirm it, it will be temporarily marked as suspicious, pending the judgment of the doctor. If the epileptic seizure lasts longer or is more serious, the accompanying staff can ask for help from the monitoring room through the system, and the monitoring room can see the situation at the scene through the video at the first time and provide guidance and help.

As shown in Figure 6, when the escort finds the patient's epileptic seizure during the escorting process, but the seizure alarm module does not alarm, the escort can manually call the police and send the missing report information to the monitoring room. Analyze and form a preliminary judgment. If the monitoring room confirms that the escort's alarm is a false alarm caused by non-specific actions, it will feed back to the accompanying staff; if the monitoring room cannot confirm the alarm, it will be marked as suspicious and will be analyzed when the doctor visually reads the picture; If the monitoring room confirms an epileptic seizure, it will feed back to the epileptic seizure alarm module, lower the threshold, and record a missed report by the epileptic seizure alarm module. The specific implementation process of lowering the alarm threshold is as follows:

r'=max(0.8, r+5(1-r)(1-e ^rP(t) )-0.01)

Among them, r' represents the adjusted alarm threshold, min() represents the function of taking the maximum value, r represents the alarm threshold before adjustment, P(t) represents the smoothing result of the seizure alarm module at time t, and t represents the current time.

If the automatic detection system has more than 3 false positives or false negatives in 1 hour, the current model will be abandoned, and the signal variance of each channel will be recalculated according to the interval EEG within 1 hour, and the model selection module will be returned to the model selection again .

The above-mentioned embodiments are used to illustrate the present invention, rather than to limit the present invention. Within the spirit of the present invention and the protection scope of the claims, any modification and change made to the present invention will fall into the protection scope of the present invention.

Claims (8)

1. A real-time detection and monitoring system for epileptic seizures oriented to epilepsy video electroencephalogram examination, characterized in that the system includes:

   Data processing module: used to input the patient's EEG signal data, divide the EEG signal data into several fixed-length data segments according to time, and perform wavelet decomposition on each channel of the segmented EEG signal data, according to the EEG electrode position , get 18 EEG signal channels, and arrange the 18 EEG signal channels according to the two-dimensional structure of 4 rows and 5 columns, and the missing 2 channels in the middle column are calculated according to the relative distance relationship between the channel positions; the output is two The decomposed data of the signal arranged in three-dimensional structure; the specific arrangement is as follows: 18 EEG signal channels are [Fp1-F7, F7-T3, T3-T5, T5-O1, Fp1-F3, F3-C3, C3 -P3,P3-O1,Fz-Cz,Cz-Pz,Fp2-F4,F4-C4,C4-P4,P4-O2,Fp2-F8,F8-T4,T4-T6,T6-O2]; Channels are arranged in a two-dimensional structure of 4 rows and 5 columns, the first column is [Fp1-F7, F7-T3, T3-T5, T5-O1], and the second column is [Fp1-F3, F3-C3, C3- P3,P3-O1], the third column is [Fz-Cz,Cz-Pz], the fourth column is [Fp2-F4,F4-C4,C4-P4,P4-O2], the fifth column is [Fp2- F8, F8-T4, T4-T6, T6-O2], add a channel before and after the third column [Fz-Cz, Cz-Pz], respectively represented by N1 and N2, that is, the third column is [N1, Fz-Cz, Cz-Pz, N2], N1 and N2 are determined according to the relative distance relationship between the channels; the distance between Fp1-F3, Fp2-F4 and Fp1-F7, Fp2-F8 to N1 is estimated to be 1:2, considering The exponential attenuation of signal propagation, so the coefficient ratio between the two is 4:1 when calculating N1, and the same is true for N2; the calculation formulas of N1 and N2 are as follows:

   

   

   Model selection module: build a model with z+3 layer structure, where the first z layer corresponds to the number of layers of wavelet decomposition in the data processing module, the input of the first layer of the model is the output of the last layer of wavelet decomposition; The input of the second layer is the output of the first layer of the model and the high frequency part decomposed by the penultimate layer of wavelet decomposition; the input of the third layer of the model is the output of the second layer of the model and the high frequency part of the penultimate layer of wavelet decomposition part; and so on, the input of the zth layer of the model is the output of the z-1st layer of the model and the high-frequency part decomposed by the first layer of wavelet decomposition; the z+1 layer, z+2 layer and z+3 layer are the volume The cumulative layer, the maximum pooling layer and the fully connected layer, the output is the probability of no seizure and the probability of seizure at the current moment; according to the data of the patient database in the hospital, the epilepsy types that exceed the threshold number of seizures in the diagnosis results are counted, and the epilepsy The EEG signal data of patients corresponding to the type respectively trains the corresponding seizure detection model, and selects the corresponding seizure detection model for detection according to the epilepsy type closest to the EEG signal data of the current patient's video EEG examination;

   Epilepsy alarm module: used to obtain the probability of epileptic seizure output by the epileptic seizure detection model in the model selection module in real time, smooth the seizure probability within a period of time, judge whether to alarm the accompanying staff according to the set alarm threshold, and Stop working after calling the police, and resume work until the escort confirms that the epilepsy has stopped or the set time has passed;

   Interactive collaborative labeling module: used for the interaction between the epileptic seizure alarm module, the escort and the monitoring room; the escort judges the false positives and false negatives of the epileptic seizure alarm module, and reviews it through the monitoring room. If it is confirmed, the epileptic seizure alarm module is judged If there are false positives and false negatives, adjust the alarm threshold.
2. A kind of real-time detection and monitoring system for epileptic seizures oriented to epilepsy video EEG inspection according to claim 1, characterized in that, the patient's EEG signal data input by the data processing module includes data from the patient database for model training data and real-time video EEG examination EEG signal data.
3. A kind of real-time detection and monitoring system for epileptic seizures oriented to epilepsy video EEG inspection according to claim 1, wherein the convolution kernel of the z+1th convolutional layer in the model constructed by the model selection module is processed according to the data The data form of the two-dimensional structure arrangement output by the module is set. Since the two-dimensional structure is arranged in the form of 4*5, the structure of the convolution kernel of the z+1th convolutional layer and the relevant part of the EEG signal channel It is set to 2*2, and the structure of the pooling kernel and the part of the EEG signal channel in the maximum pooling layer of the z+2 layer is set to 3*2.
4. A real-time detection and monitoring system for epileptic seizures oriented to epilepsy video EEG examination according to claim 1, characterized in that, according to the data of the patient database in the hospital, 18 channels of EEG signal data are calculated for each patient between seizures The variance of the signal data of each frequency band, where each frequency band corresponds to the frequency band of the wavelet decomposition input in the front z layer of the model, and a matrix representing the characteristics of the EEG signal is obtained, and the diagnosed epilepsy type is used as the label; the current patient starts to perform video During the EEG examination, calculate the variance of the signal data of each frequency band of the 18 channels of EEG signal data during the interictal period, and obtain a matrix representing the characteristics of the current patient's EEG signal, and use the k-nearest neighbor algorithm to determine the closest epilepsy type, select Corresponding seizure detection model.
5. A kind of real-time detection and monitoring system for epileptic seizures oriented to epilepsy video EEG examination according to claim 4, wherein the calculation formula of the distance distance between the current patient and the patients in the patient bank in the hospital in the k-nearest neighbor algorithm is as follows :

   

   where d _i represents the distance corresponding to the i-th channel.
6. A kind of epileptic seizure real-time detection and monitoring system oriented to epilepsy video electroencephalogram examination according to claim 1, characterized in that, the epileptic seizure alarm module smoothes the τ seizure probability values within a period of time, and Set the alarm threshold to 0.9, and if it exceeds the alarm threshold, an alarm will be sent to the escort; the specific implementation process is as follows:

   

   Among them, P(t) represents the smoothing result of the seizure alarm module at time t, p(t-j) represents the probability of seizures output by the seizure detection model at time t-j, t represents the current moment, and j represents the time interval.
7. A real-time detection and monitoring system for epileptic seizures oriented to epilepsy video EEG inspection according to claim 1, wherein the interactive collaborative labeling module raises the alarm threshold when the epileptic seizure alarm module makes a false alarm, specifically implementing The process is:

   r'=min(0.95, r+5(1-r)(1-e ^rP(t) )+0.01)

   When the epileptic seizure alarm module fails to report, the alarm threshold is lowered; the specific implementation process is:

   r'=max(0.8, r+5(1-r)(1-e ^rP(t) )-0.01)

   Among them, r' represents the adjusted alarm threshold, min() represents the function of taking the minimum value, r represents the alarm threshold before adjustment, P(t) represents the smoothing result of the seizure alarm module at time t, and t represents the current time.
8. A real-time detection and monitoring system for epileptic seizures oriented to epilepsy video EEG examination according to claim 1, wherein the interactive collaborative labeling module confirms that the total number of false positives or false negatives of the seizure alarm module exceeds 3 times, Then the current epileptic seizure detection model is abandoned, and the model selection is re-selected according to the EEG signal data of the patient's interictal period in the most recent period.

Images