A method and system for intelligent identification of multiple abnormal EEG patterns
By combining fast Fourier transform, hemi-brain peak matching, Minirocket, RidgeClassifierCV, wavelet coefficient and XGBoost, multiple abnormal patterns in EEG are identified, and the identification efficiency and high false alarm rate in the existing technology is solved. Problem, achieving rapid and accurate EEG abnormal pattern recognition and report generation.
Patent Information
- Application Number
- CN202410315581.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-19
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-03-19
AI Technical Summary
The existing EEG abnormal pattern recognition method is time-consuming and labor-intensive, making it difficult to quickly identify multiple abnormal patterns, especially burst suppression in low burst states and periodic discharges with large individual differences. The existing automatic recognition algorithm runs slowly, and the data The processing is complex, the false positive rate is high, and the time difference between multiple modes cannot be effectively handled.
Fast Fourier transform and nearest neighbor threshold comparison method are used to identify burst suppression, half-brain peak matching recognizes periodic discharge, Minirocket and RidgeClassifierCV are used to identify spinous-spike waves, epilepsy is identified based on wavelet coefficients and XGBoost, and low voltage is identified in combination with threshold discrimination. And through data reorganization and fusion modules, the recognition speed is improved, the false alarm rate is reduced, and the EEG report in image form is output.
It realizes rapid identification of various abnormal EEG patterns, reduces the pressure of medical staff to read pictures, improves recognition efficiency, can quickly identify under small amounts of data, reduces computing resource requirements, and assists with image display and report output Clinical diagnosis.
Smart Images

Figure CN118203334B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of artificial intelligence and clinical medicine, and specifically to a method and system for intelligently identifying a variety of abnormal EEG patterns, which are used for identifying abnormal EEG patterns such as burst suppression, periodic discharge, spike waves, epilepsy, and low voltage. Background Art
[0002] As the most widely used technology for recording human brain electrical activity, EEG plays an important role in many fields. Especially in the clinical field, EEG plays a wide range of roles. It can be used to diagnose diseases such as epilepsy, encephalitis, brain tumors, and can also be used to identify the cause of coma and detect postoperative anesthesia. However, at present, clinicians need to read EEG manually when processing it. A person's 24-hour EEG data may take a doctor several hours to fully judge, which is very time-consuming and laborious. Therefore, it is of great significance to use artificial intelligence methods to automatically recognize EEG.
[0003] In medical EEG identification, epilepsy, burst suppression, periodic discharge, spike waves, and low voltage are relatively common. Epilepsy is a chronic disease caused by a disorder of the nervous system, which is mainly manifested by repeated epileptic seizures.
[0004] Burst suppression is an EEG pattern characterized by alternating bursts of medium-high amplitude activity and low voltage states, which is common in general anesthesia, hypothermia, hypoxia, etc. International standards believe that standard burst suppression should be an EEG pattern in which the proportion of the suppression period (<10uV) is greater than 50% of the recorded components, but due to individual differences, the EEG amplitude of some patients' bursts may be lower than 10uV.
[0005] Periodic discharge is a rhythmic, burst-like, recurring discharge. Burst suppression is a special form of periodic discharge, but the periodic discharge referred to in the present invention is limited to extensive periodic discharge without burst suppression, that is, repeated and similarly spaced discharges that occur synchronously in both cerebral hemispheres, which has a good predictive effect in the prognosis assessment of comatose patients.
[0006] Spike waves are a type of sudden EEG change that clearly stands out from the background waves and are a typical feature of epileptic discharges, but the presence of spike waves does not necessarily mean an epileptic seizure is present.
[0007] Low voltage refers to the EEG phenomenon in which the EEG voltage is continuously lower than a certain value. It is generally caused by extensive and severe damage to the cerebral cortex. Its existence indicates that the patient's brain function may be severely inhibited or basically lost.
[0008] Automatic identification of these abnormal patterns can help reduce the pressure on medical staff to read images, while improving the efficiency of reading images, which can speed up the time it takes to issue EEG reports, allowing doctors to understand the patient's condition more quickly and assist in diagnosis and treatment.
[0009] However, the existing abnormal EEG recognition methods have the following shortcomings:
[0010] 1. Artificial EEG recognition can achieve higher recognition accuracy, but it is time-consuming and labor-intensive.
[0011] 2. Most of the existing automatic recognition algorithms focus on the recognition of a single abnormal EEG pattern. In terms of burst suppression recognition, threshold discrimination is relatively common, but due to individual differences, the EEG amplitude of some patients' burst forms will also be lower than 10uV. It is difficult to identify patients in such low-burst states with simple threshold discrimination. However, more complex algorithms such as recursive graphs take a long time to run and cannot achieve rapid recognition. In terms of periodic discharge recognition, the traditional method is to manually extract features and then use machine learning algorithms for discrimination, but due to individual differences among patients, the periodicity of each patient is different, and it is difficult to produce unified features; as for deep learning algorithms such as neural networks, although they can automatically extract features, they require a large amount of data for training and run slowly. Epilepsy recognition and spike wave recognition are relatively mature, but the final result is greatly affected by the ratio of positive and negative samples in the data. There are few studies related to low voltage.
[0012] 3. Existing models that can classify multiple abnormal EEG patterns are limited by data sets and often input data of fixed length without considering the time differences between various abnormal EEG patterns.
[0013] The present invention designs an algorithm to quickly identify abnormal EEG patterns such as burst suppression, periodic discharge, spike waves, epilepsy, low voltage, etc., and integrates the algorithm into an abnormal EEG recognition system. After medical staff import EEG data, the system can automatically identify EEG, indicate the time point when abnormal EEG occurs, and display it in the form of images. At the same time, the EEG report is output in the format of the hospital's EEG report, reducing the pressure on medical staff to read the images. Summary of the invention
[0014] The present invention relates to a method and system for intelligently identifying a variety of abnormal EEG patterns. The purpose is to solve the problem that it is time-consuming and laborious for medical staff to read EEGs, and to use artificial intelligence technology to realize the rapid identification of abnormal EEG patterns such as burst suppression, periodic discharge, spike waves, epilepsy, low voltage, etc. that are common in clinical practice.
[0015] In order to achieve the above object, the present invention proposes a method and system for intelligent identification of multiple abnormal EEG patterns. The technical solution is as follows:
[0016] In a first aspect, the present invention provides a method for intelligently identifying multiple abnormal EEG patterns, which is divided into the following steps:
[0017] S1. EEG burst suppression pattern recognition based on fast Fourier transform and nearest neighbor threshold comparison method. Specifically, the original EEG is segmented according to a fixed interval t, which can generally be taken from 10 seconds to 1 minute, and it is mainly necessary to include at least one burst suppression cycle. After discrete wavelet filtering and outlier processing, the spectrum features after fast Fourier transform are extracted, and the nearest neighbor threshold comparison method is used to distinguish burst waves and suppression waves.
[0018] S2. For the recognition of periodic discharge patterns of hemi-brain synchronization, the original EEG is segmented according to a fixed interval t, which can generally be taken from 10 seconds to 1 minute, and the main thing is to include at least one discharge cycle. After discrete wavelet filtering and outlier processing, hemi-brain peak matching is performed. If the match is achieved, it is considered that periodic discharge exists.
[0019] S3. Spike wave recognition based on minirocket: resample the original data, filter it with Butterworth bandpass filter, extract EEG features with minirocket (see Dempster A's 2021 paper Minirocket: A very fast (almost) deterministic transform for time series classification published on KDD), and use RidgeClassifierCV for classification training (see D.Nhu's 2023 paper Automated Interictal Epileptiform Discharge Detection from Scalp EEG Using Scalable Time-series Classification Approaches). No additional cross-validation is required during training. Save the obtained spike wave recognition model.
[0020] S4. Epilepsy seizure identification based on wavelet coefficients: resample the original data, extract the wavelet coefficients after db4 decomposition after removing outliers, and use xgboost for classification training (see Tianqi Chen's 2016 paper XGBoost:AScalable Tree Boosting System for details). During training, due to the disparity in the ratio of positive and negative samples, negative sample data greater than a certain threshold are extracted for each patient in training, but all positive sample data are retained. The threshold should be less than the original number of negative samples, and its purpose is to reduce the number of negative sample data. At the same time, the upsampling method is used to increase the number of positive samples. Finally, the obtained epilepsy seizure model is retained.
[0021] S5. Low voltage discrimination based on threshold discrimination: most of the values in the EEG data that are less than the threshold are marked as suspected low voltage. The threshold is generally set according to the definition of low voltage, such as 20uV, but can be relaxed according to its own data, such as 25uV. The proportion of relevant data in the overall EEG data is counted, and it is considered to be low voltage if it is greater than the threshold.
[0022] S6. Fusion of multiple abnormal EEG recognition modes, comprehensive consideration of the correlation and differences between various EEGs, can greatly improve recognition speed while reducing false alarm rate; the specific steps are as follows:
[0023] S6.1. Cut the original EEG into non-repeating windows at a fixed interval t, which is the same as the value in the previous burst suppression identification. Check whether the channel of the data in each window is consistent with the channel used during training or algorithm design. If it does not meet the requirements, terminate directly, otherwise recombine the columns that meet the corresponding channels. The channel here refers to the potential difference between the two input terminals when the EEG data is collected. For example, Fp1-F3 represents the potential difference from Fp1 to F3. The specific channel depends mainly on the collected data. The combined data is abnormally transformed and resampled at the same time.
[0024] S6.2. In order to improve the recognition speed, reshape the data in the window. Specifically, divide the data in the window into each second, transpose each second, and then add it to a dataframe. Assuming that the original data is represented as t*fequency×channel, t*fequency is the number of EEG data in time t, frequency is the sampling frequency, and channel is the number of channels, then the reshaped data format is t*channel×frequency.
[0025] S6.3. For the reshaped data, call the spike wave recognition model to perform spike wave recognition, with 1 indicating the presence of a spike wave and 0 indicating the absence. The final result will be a 0l list of length t×channel.
[0026] S6.4. Integrate the above lists, with each channel as a group. If there is 1, it is marked as 1. Therefore, the original 01 list with a length of t×channel becomes a 01 list with a length of t.
[0027] S6.5. Check whether there are m consecutive 1s in the 01 list. If so, call the epilepsy recognition model to perform epilepsy detection on this part of the data. Otherwise, proceed to S6.6. If epilepsy is detected, mark the location where the epilepsy data exists.
[0028] S6.6. Use the data before reformatting to perform low voltage detection using a low voltage detection algorithm.
[0029] S6.7, perform wavelet decomposition filtering and abnormal data detection on the data, and then use the burst suppression and periodic discharge detection algorithm to perform burst suppression detection and periodic discharge detection. The detection results enter the corresponding result list, which retains data of a fixed length.
[0030] S6.8. If burst suppression and periodic discharge exist in the data segment at the same time, mark the final result as burst suppression.
[0031] S6.9. After all windows are tested, the low voltage, burst suppression and periodic discharge test results are counted and the respective threshold determination steps are entered. For specific thresholds, see the specific descriptions of each part.
[0032] S6.10. The degree of overlap between the epilepsy detection results and burst suppression and periodic discharges. If there is overlap, the priority of the final report is: burst suppression > periodic discharge > epilepsy.
[0033] The specific steps of the nearest neighbor threshold comparison method in S1 above are as follows:
[0034] S1.1. Set a flag to indicate the current state, initially 'end'.
[0035] S1.2. Find the beginning of the burst wave. The beginning of the burst wave is considered to be a significant change in the spectrum compared to the previous second. The formula is as follows:
[0036] (fft j / fft j-1 >start_threshold)∩(flag==′end′)
[0037] Among them, fftj Indicates the feature of the jth second in the current window. start_threshold indicates the threshold.
[0038] S1.3. Find the end time of the burst wave. In addition to the large change in frequency compared to the previous second, the frequency characteristics at the end must also be similar to those before the burst begins. The formula is as follows:
[0039] ((fft high / fft j > end_threshold1)∪(fft j / fft low <end_threshold2))∩(flag=='start')∩(time<=time_threshold)
[0040] Among them, fft high represents the maximum feature from the beginning of the burst wave to the jth second, fft low It indicates the characteristic value one second before the outbreak begins, time indicates the duration of the outbreak, end_threshold1, end_threshold2, and time_threshold are three thresholds that can be determined through experiments.
[0041] S1.4. After the burst wave detection is completed, all segments between the burst waves are considered as candidate suppression segments and the suppression wave detection is performed. The specific requirements for the suppression wave are as follows:
[0042] (abs(Q 10 (Y j ))<5)∩(abs(Q 90 (Y j ))<5)∩(std_baofa j / 3>=std_yizhi j )∩(std_baofa j+1 / 3>
[0043] =std_yizhi j )∩(std_yizhi j <3)∩(((Q 99 (Y j )<Q 99 (X j ) / 2.5)∩(Q 99 (Y j )<Q 99 (X j+1 ) / 2.5))∪((abs(Q1(Y j ))<abs(Q1(X j)) / 2.5)∩(abs(Q1(Y j ))<abs(Q1(X j+1 )) / 2.5)))∩(Q1(Y j )>min(X j ))∩(Q1(Y j )>min(X j+1 ))
[0044] Where abs(X) represents the absolute value of X, min(X) represents the minimum value of X, and Q m (X) represents the m-quantile of X, std_baofa j Indicates the standard deviation of the j-th burst segment, std_yizhi j represents the standard deviation of the jth inhibition segment, X j represents the jth burst wave, Y i represents the j-th candidate suppression wave.
[0045] S1.5. Two burst waves and one suppression wave are recorded as a burst suppression, and the number of burst suppressions in each channel is counted in k consecutive windows. If it exceeds the threshold baoyi_thres, it is considered that there is burst suppression. The threshold can be obtained through experiments, there is no fixed range, and it is related to the data itself. The patient's data can be statistically analyzed first, and then the number of burst suppressions that most patients have can be selected.
[0046] The specific steps of hemi-brain peak matching in S2 above are as follows:
[0047] S2.1. Standardize the data of each channel and record it as a forward sequence, then negate it to form a reverse sequence.
[0048] S2.2, traverse each forward sequence, calculate the upper limit position of the background wave, the value is set as 3 times the upper quartile of the sequence and the maximum value of the upper 95%, then calculate the peak position greater than the background wave, add it to the peak list, and require that the distance between peaks is greater than 50 data points. The upper limit position of the background wave and the peak distance can be enlarged or reduced as needed.
[0049] S2.3. Traverse each reverse sequence and calculate the upper limit of the background wave. However, this value is set as the minimum of -3 times the lower quartile of the corresponding forward sequence and the lower 5% of the corresponding forward sequence. It can also be adjusted as needed. Peaks larger than the background wave are added to the trough list, and the distance between peaks is required to be greater than 50 data points. The range setting requirements for peaks and troughs are as follows:
[0050] peak>max(Q 75 (E j,m )*3,Q 95(E j,m ))
[0051] peak_down<min(Q 25 (E j,m )*3, Q5(E j,m ))
[0052] distance>=50
[0053] Among them, Q m (X) represents the m-quantile of X, and distance represents the minimum distance between adjacent peaks and adjacent troughs. j,m Represents the EEG data of the jth channel and the mth window, with the peak represented by peak and the trough represented by peak_down
[0054] S2.4, perform peak and trough matching in the left and right hemispheres. j,m Where j is an odd number for the left brain and j is an even number for the right brain. For each hemisphere, the number of peaks and troughs that can be matched is calculated. Spike matching is defined as the time when the spikes in the same hemisphere occur within 0.2s. This value can be adjusted as needed. If you want to reduce the false alarm rate, you can lower this value.
[0055] S2.5. Calculate the proportion of the peaks and troughs matched in each hemibrain in the original peak and trough list. If the percentage exceeds a certain threshold value pre_thres, it is considered that the current window may have periodic discharges. This threshold can be obtained through experiments. The peak and trough proportions of all patients are sorted from small to large, and the value is taken according to actual needs, such as 0.6. If there are periodic discharges in k consecutive windows, then it is determined that this person has periodic discharges.
[0056] The present invention also provides an abnormal EEG recognition system based on the above method, comprising:
[0057] EEG signal preprocessing module, used for window division, outlier processing, checking of original data channels and data resampling;
[0058] The burst suppression identification module extracts the spectrum features after fast Fourier transformation after discrete wavelet filtering, and uses the nearest neighbor threshold comparison method to distinguish burst waves from suppression waves;
[0059] The periodic discharge recognition module performs hemi-brain peak matching after discrete wavelet filtering. If a match is found, it is considered that periodic discharge may exist.
[0060] The spike wave recognition module uses a Butterworth bandpass filter, then uses minirocket to extract EEG features and inputs them into RidgeClassifierCV for discrimination;
[0061] The epilepsy recognition module extracts the wavelet coefficients after db4 decomposition and uses machine learning algorithms for identification;
[0062] The low voltage identification module uses a threshold discrimination method to identify low voltage.
[0063] The fusion module is mainly responsible for coordinating the calling of the burst inhibition recognition module, periodic discharge recognition module, spike wave recognition module, epilepsy recognition module, and low voltage recognition module, as well as the display priority of the later results according to the associations and differences between various abnormal EEG patterns, while improving the overall recognition speed.
[0064] The time labeling module mainly includes the extraction of EEG start time and subsequent real time mapping, projecting the time of the identified abnormality to the correct time point;
[0065] The report generation module is performed after the detection is completed, and mainly includes the capture of images corresponding to the abnormal time and the generation of report results.
[0066] Technical effects and advantages of the present invention:
[0067] 1. The burst suppression identification method provided by the present invention can also identify patients with low bursts by using fast Fourier transform and nearest neighbor threshold comparison method, which overcomes the influence of patient differences on the identification results to a certain extent. Moreover, the required computing resources are small, and fast identification can be performed;
[0068] 2. The periodic discharge recognition method provided by the present invention takes into account the synchronization of the hemi-brain and uses hemi-brain peak matching for recognition. Compared with the traditional method, it does not require a training set, can be recognized with a small amount of data, and has a fast recognition speed;
[0069] 3. The present invention comprehensively considers the similarities and differences between various abnormal EEG patterns, and uses steps such as data reorganization to speed up the recognition speed. The corresponding abnormal EEG system can automatically identify the EEG, indicate the time point when the abnormal EEG occurs, and display it in the form of an image. At the same time, the EEG report is output in the format of the hospital's EEG report, reducing the pressure on medical staff to read the images. BRIEF DESCRIPTION OF THE DRAWINGS
[0070] Figure 1 Identify flow charts for outbreak suppression;
[0071] Figure 2 Identify the flow chart for periodic discharge;
[0072] Figure 3 Flow chart for spike wave identification;
[0073] Figure 4 Identify a flow chart for epilepsy;
[0074] Figure 5 Integrate a flow chart for the algorithm;
[0075] Figure 6 Use a flow chart for your system. DETAILED DESCRIPTION
[0076] The method of the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The present invention mainly aims at the problem that it is time-consuming and laborious for medical staff to manually identify EEGs, and designs an artificial intelligence method to automatically identify five abnormal EEG patterns: burst suppression, periodic discharge, spike waves, epilepsy, and low voltage. Data preprocessing is performed by discrete wavelet transform, bandpass filtering and other methods, and then features are extracted by fast Fourier transform, wavelet decomposition, Minirocket and other methods. The final abnormal EEG recognition results and the time point of abnormal EEG occurrence and other information are obtained through the model to assist clinical diagnosis.
[0077] The specific implementation plan is:
[0078] In order to solve the problem that it takes time and effort for medical staff to read EEG, the present invention proposes a method and system for intelligent recognition of multiple abnormal EEG patterns, which solves the recognition of abnormal EEG patterns such as burst suppression, periodic discharge, spike wave, epilepsy, low voltage, etc. The data used in the present invention conforms to the 10-20 international standard. For each abnormal EEG recognition, it basically includes data preprocessing, feature extraction, and model or algorithm design.
[0079] Specifically include the following:
[0080] (1) Burst suppression identification
[0081] ① Data preprocessing and feature extraction: For the original EEG data, first perform non-repetitive cutting at intervals of t minutes. For each window obtained by segmentation, first perform a db6 discrete wavelet decomposition, and then reconstruct the signal to remove some noise interference. Then perform an outlier detection again, and count the number of absolute values of the amplitude in each channel that are greater than max. If the number is greater than count, it is determined that the window may have large interference and is directly discarded, otherwise continue to identify. For the window that passes the outlier detection, obtain the value with the largest amplitude corresponding to the spectrum value after fast Fourier transform as the feature. In the experiment, t, max, and count are 1, 500, and 10 respectively.
[0082] ②Algorithm design: Different from the general threshold discrimination method, considering the differences in EEG between different patients, the nearest neighbor discrimination is used, that is, only the changes in features between the last few seconds are considered. This can adapt to the differences between different patients and the differences in EEG of the same patient before and after the change of time to a certain extent. Specifically, a flag is set to indicate the current state, which is initially 'end'. Then the burst wave detection begins. The start of the burst wave is determined as follows:
[0083] (fft j / fft j-1 >start_threshold)∩(flag==′end′)
[0084] Among them, fft j It represents the feature of the jth second in the current window, start_threshold represents the threshold set for the burst wave, and after experiments, bao_threshold of 2.5 can achieve good recognition results.
[0085] The end criteria of the burst wave are set as:
[0086] ((fft high / fft j > end_threshold1)∪(fft j / fft low <end_threshold2))∩(flag=='start')∩(time<=time_threshold)
[0087] Among them, fft high represents the maximum feature from the beginning of the burst wave to the jth second, fft low It represents the characteristic value one second before the outbreak begins, time represents the duration of the outbreak, end_threshold1, end_threshold2, and time_threshold are three thresholds. After experiments, it was found that 2.5, 2.5, and 5 respectively can achieve better experimental results.
[0088] Then we start to detect the existence of suppression waves. The data segment between each burst wave may be a suppression wave. Assuming that X j =[baofa_start j ,baofa_end j ] represents the jth burst wave, then Y j =[baofa_end j ,baofa_start j+1] is the wave band between the jth burst wave and the j+1th burst wave. This section of data needs to be tested for suppression waves. The standard setting for suppression waves is as follows:
[0089] (abs(Q 10 (Y j ))<5)∩(abs(Q 90 (Y j ))<5)∩(std_baofa j / 3>=std_yizhi j )∩(std_baofa j+1 / 3>
[0090] =std_yizhi j )∩(std_yizhi j <3)∩(((Q 99 (Y j )<Q 99 (X j ) / 2.5)∩(Q 99 (Y j )<Q 99 (X j+1 ) / 2.5))∪((abs(Q1(Y j ))<abs(Q1(X j )) / 2.5)∩(abs(Q1(Y j ))<abs(Q1(X j+1 )) / 2.5)))∩(Q1(Y j )>min(X i ))∩(Q1(Y j )>min(X j+1 ))
[0091] Where abs(X) represents the absolute value of X, min(X) represents the minimum value of X, and Q m (X) represents the m-quantile of X, std_baofa j Indicates the standard deviation of the j-th burst segment, std_yizhi j represents the standard deviation of the jth inhibition segment.
[0092] Two burst waves and one suppression wave are recorded as a burst suppression, and the number of burst suppressions in each channel is counted in k consecutive windows. If it exceeds the threshold baoyi_thres, it is considered that burst suppression exists. After experiments, k is set to 10 and baoyi_thres is set to 8 to achieve better results.
[0093] (2) Periodic discharge identification
[0094] ① Data preprocessing: It is similar to the preprocessing process of burst suppression identification, so it will not be repeated here.
[0095] ② Feature extraction and algorithm design: In the case of a small data set, it is often difficult to achieve good results using machine learning algorithms or deep learning algorithms for recognition. In addition, due to differences between patients, the periodicity of each patient is different, and it is difficult to start from the perspective of periodicity. Therefore, the synchronization of the hemisphere is considered during feature extraction. It mainly includes the following steps:
[0096] Step 1: Standardize the data of each channel, record it as a forward sequence, and then negate it to form a reverse sequence.
[0097] Step 2: traverse each forward sequence and calculate the upper limit position of the background wave, which is set as 3 times the upper quartile of the sequence and the maximum value of the upper 95%. Then calculate the peak position that is greater than the background wave and add it to the peak list. The distance between the peaks is required to be greater than 50 data points.
[0098] Step 3, traverse each reverse sequence and calculate the upper limit of the background wave, but the value is set as the minimum of -3 times the lower quartile of the corresponding forward sequence and the lower 5% of the corresponding forward sequence. Peaks larger than the background wave are added to the trough list, and the distance between peaks is required to be greater than 50 data points. The range setting requirements for peaks and troughs are as follows:
[0099] peak>max(Q 75 (E j,m )*3,Q 95 (E j,m ))
[0100] peak_down<min(Q 25 (E j,m )*3, Q5(E j,m ))
[0101] distance>=50
[0102] Among them, Q m (X) represents the m-quantile of X, and distance represents the minimum distance between adjacent peaks and adjacent troughs. j,m Represents the EEG data of the jth channel and the mth window, with the peak represented by peak and the trough represented by peak_down
[0103] Step 4: Match the peaks and troughs of the left and right hemispheres. j,m Where j is an odd number for the left hemisphere, and j is an even number for the right hemisphere. For each hemisphere, the number of peaks and troughs that can be matched is calculated. Spike matching is defined as the time when spikes occur within 0.2s in the same hemisphere.
[0104] Step 5: Calculate the proportion of the peaks and troughs in each hemibrain in the original peak and trough list. If the percentage exceeds a certain threshold pre_thres, it is considered that the current window may have periodic discharges. If there are periodic discharges in k consecutive windows, then it is determined that this person has periodic discharges. After experiments, setting pre_thres and k to 0.6 and 6 respectively can obtain better experimental results.
[0105] (3) Spike wave identification
[0106] ① Data preprocessing: Read the EDF EEG data according to the label file and resample it to 125Hz. Considering the possible interference in the EEG data, a Butterworth bandpass filter is applied to each acquired data to retain the components between 0.5Hz and 49Hz. This is a proven effective preprocessing method.
[0107] ② Feature extraction and model design: Use Minirocket to extract features from each set of data (1 second), and input the extracted features into RidgeClassifierCV for classification training. Since RidgeClassifierCV has built-in cross-validation, no additional cross-validation is required. During the experiment, the value of RidgeClassifierCV was set to (-3, 3, 10).
[0108] (4) Epilepsy identification:
[0109] ① Data preprocessing: In order to unify the data, the original data was resampled to 125Hz. In addition, since the original data recorded a single channel, such as FP1, but epileptic seizures were marked on dual channels such as Fp1-F3, the data needed to be further organized in a dual-channel format.
[0110] The next step is to divide the window. Through statistical analysis of epileptic seizure time, it can be found that about 99% of epileptic seizures last longer than 3.5 seconds, so 4 seconds is used as a window to divide the data set. Data less than 4 seconds are directly discarded. In addition, for the data of all channels, first remove the data that are all 0, and then replace the abnormal values such as nan and inf with 0.
[0111] ② Feature extraction: Perform a 5-level db4 wavelet transform, which has been proven to be effective in epilepsy identification. Get the final wavelet coefficients D1, D2, D3, D4, D5, and A6. Here D represents the detail coefficient, and A represents the approximate coefficient. This method does not consider the coefficients at level A, but only further extracts features from the D coefficients, including log sum, standard deviation of the coefficients in each subband, average value of the absolute value of the coefficients in each subband, average power of the wavelet coefficients in each subband, and ratio of the absolute average values of adjacent subbands. However, the final experiment shows that the average value of the absolute value of the coefficients in each subband can achieve better experimental results.
[0112] The final extracted features will be a 5×C matrix, where C represents the number of channels. The matrix needs to be flattened to form 1×5C data.
[0113] ②Model training and parameter setting: Considering that the similarity of epileptic seizures in the same patient may be large, if the method of randomly dividing the training set and the test set is adopted, the data of the same patient will be divided into both the training set and the test set, resulting in a false good result in the end. Therefore, this point needs to be considered when dividing the data set. However, since there is a lot of data after segmentation, and the ratio of positive and negative samples in the data is very different, in order to reduce the gap in the number of positive and negative samples in the data, the data set needs to be processed to a certain extent. For each patient in the training set, if it is shown that there is an epileptic seizure, then all the data of the epileptic seizure will be retained. If it is shown as background, then the number of data will be tested. If it exceeds part_num, then part_num pieces of data will be randomly extracted, otherwise all will be retained. The data retention condition can be seen in the following formula:
[0114] Data retention = (data_type == 'seiz') ∪ ((data_type == 'bckg') ∩ (length <
[0115] =part_num))∪(random_get(data_type=='bckg'), part_num)
[0116] Among them, data_type indicates the label represented by the current data, length indicates the length of all the data of the patient, and random_get(X, m) indicates randomly extracting m data from X. The part_num used in the experiment is 5.
[0117] In addition, the extracted data is oversampled again to increase the number of epileptic seizure samples. The classification model used in this method is xgboost. Before training, the data is standardized and converted to the data format required by xgboost. Then the model was trained. Experiments show that better results can be achieved when the model parameters learning_rate, n_estimators, max_depth, min_child_weight, subsample, and gamma are set to 0.227, 199, 9, 2, 0.688, and 0.373. Since the model outputs probability, 0.5 is used as the dividing point. Values greater than or equal to 0.5 are considered epileptic seizures, and values less than 0.5 are considered background.
[0118] (5) Low voltage identification: Low voltage identification is mainly based on the characteristics of low voltage EEG morphology. For the original EEG data, it is also divided into non-repetitive intervals of 1 minute at the beginning. For each window, all non-numeric data are converted to 0. Low voltage is defined as an EEG with an amplitude always lower than 20uV. Due to the possibility of some interference, the conditions are relaxed during the judgment. The judgment conditions are as follows:
[0119] Q 99.9 (X)<25
[0120] If the proportion of the entire EEG that is judged to be low voltage is greater than the threshold, then the person is considered to have low voltage. In the experiment, the threshold was set to 50%.
[0121] (6) Fusion of multiple abnormal EEG: The identification of the five abnormal EEGs, burst suppression, periodic discharge, spike wave, epilepsy, and low voltage, is done at different time granularities. Spike wave is identified C times per second, where C represents the number of channels. Epilepsy is identified once every 4 seconds, and each time for all data. Burst suppression, periodic discharge, and low voltage are identified once every 1 minute. If spike wave is identified once every 1 second, and then epilepsy is identified after 4 seconds, and then the remaining abnormal EEGs are identified after 1 minute, it will be very time-consuming.
[0122] The specific steps are as follows:
[0123] Step 1: Cut the original EEG into non-repeated windows at 1-minute intervals, and check whether the channels of the data in each window match the channels used in training. If not, terminate the training process directly. Otherwise, reassemble the columns that match the corresponding channels. The combined data is transformed abnormally and resampled to 125Hz.
[0124] Step 2: In order to improve the recognition speed, reshape the data in the window. Specifically, divide the data in the window into each second, transpose each second, and then add it to a dataframe. Assuming that the original data is represented as 7500×channel, 7500 is the number of EEG data in 1 minute, and channel represents the number of channels, then the reshaped data format is 60*channel×125.
[0125] Step 3: For the reshaped data, perform spike wave recognition, with 1 indicating the presence of a spike wave and 0 indicating the absence. The final result will be a 01 list with a length of 60×channel.
[0126] Step 4: Integrate the above lists, with each channel as a group. If there is 1, it is marked as 1. Therefore, the original 01 list with a length of 60×channel becomes a 01 list with a length of 60.
[0127] Step 5: Check whether there are 4 consecutive 1s in the 01 list. If so, perform epilepsy detection on this part of the data, otherwise proceed to step 6. If epilepsy is detected, mark the location where the epilepsy data exists.
[0128] Step six, use the data before reformatting to detect low voltage.
[0129] Step 7: Perform wavelet decomposition filtering and abnormal data detection on the data, and then perform burst suppression detection and periodic discharge detection. The detection results enter the corresponding result list, which retains data of fixed length.
[0130] Step eight: if the data segment has both burst suppression and periodic discharge, the final result is marked as burst suppression.
[0131] Step nine, after all windows are tested, the low voltage, burst suppression and periodic discharge test results are counted and the respective threshold determination steps are entered. For specific thresholds, see the specific description of each part.
[0132] Step 10: Detect the degree of overlap between epilepsy results and burst suppression and periodic discharges. If there is overlap, the priority of the final report is: burst suppression > periodic discharge > epilepsy.
[0133] The present invention also provides an abnormal EEG recognition system, which mainly includes the following modules:
[0134] EEG signal preprocessing module, used for window division, outlier processing, checking of original data channels and data resampling;
[0135] The burst suppression identification module extracts the spectrum features after fast Fourier transformation after discrete wavelet filtering, and uses the nearest neighbor threshold comparison method to distinguish between burst waves and suppression waves; the module executes the steps and methods in the burst suppression identification algorithm, which will not be repeated here.
[0136] The periodic discharge recognition module performs hemi-brain peak matching after discrete wavelet filtering. If a match is found, it is considered that periodic discharge may exist. The module executes the steps corresponding to the periodic discharge recognition algorithm, so the specific content will not be repeated here.
[0137] The spike wave recognition module uses a Butterworth bandpass filter to retain the components between 0.5Hz and 49Hz, then uses minirocket to extract EEG features and inputs them into RidgeClassifierCV for discrimination; the module executes the corresponding steps of the spike wave recognition algorithm, which will not be repeated here.
[0138] The epilepsy recognition module extracts the wavelet coefficients after db4 decomposition and uses a machine learning algorithm for discrimination; the module executes the corresponding steps of the epilepsy recognition algorithm, which will not be repeated here.
[0139] The low voltage identification module uses a threshold discrimination method to identify low voltage. The module executes the corresponding steps of the low voltage identification algorithm, which will not be repeated here.
[0140] The fusion module is mainly responsible for coordinating the calling of the burst inhibition recognition module, periodic discharge recognition module, spike wave recognition module, epilepsy recognition module, and low voltage recognition module, as well as the display priority of the later results according to the associations and differences between various abnormal EEG patterns, while improving the overall recognition speed.
[0141] The time labeling module mainly includes the extraction of EEG start time and subsequent real time mapping, projecting the time of the identified abnormality to the correct time point;
[0142] The report generation module is performed after the detection is completed. It mainly includes the capture of images corresponding to the abnormal time and the generation of report results. When capturing images, the first 10s EEG of each abnormality is captured. If the total number is less than 5, random images are captured. There are two types of report results. One is a text result indicating what abnormalities there are, and the other is a docx document that imitates the hospital report.
[0143] The system receives the entire EEG data, and for each EEG data segment, starts the model for identification, determines the time point and duration of various abnormal EEG, and gives a recognition result. Then the entire EEG is plotted, and the various abnormal EEG times are marked in the original EEG. In addition, considering the need for medical staff to produce reports, when the recognition result is given, a report template is generated based on the results of the hospital EEG report, and medical staff are allowed to make further modifications. At the same time, the current recognition result is saved, allowing the user to view it in the history record. See the flowchart of the system for details. Figure 6 .
[0144] The abnormal EEG that can be identified by this system is not limited to the above five types. If you need to identify a new abnormal EEG, you first need to use wavelet transform, filtering and other methods to pre-process the data of this EEG, then extract features, and then use the extracted features for identification to obtain the trained model or algorithm. In addition, it is necessary to consider the correlation between this abnormal EEG and the existing abnormal EEG. If the correlation is very low, it can be directly integrated into the system. Otherwise, in order to improve the recognition speed, it is necessary to identify it under certain conditions. Taking spike waves as an example, if the newly added abnormal EEG has a large number of spike waves or basically no spike waves, it can be performed when the spike wave existence time or non-existence time exceeds a certain threshold.
[0145] Figure 1 This is the implementation process of burst suppression identification. Taking 1 minute of EEG data as an example, the data is first subjected to discrete wavelet decomposition filtering, and then analyzed second by second. For each second of EEG data, outlier detection is performed, and the outlier value is set to 500. The number of outliers is calculated. If it exceeds the threshold, it is discarded, otherwise continue to distinguish. Then perform fast Fourier transform on each channel to extract features. Combine the features of adjacent time to perform burst wave detection. For the specific detection process, see the specific steps in the scheme. If the burst wave detection is passed, then the suppression wave detection is performed. If the suppression wave detection is passed, the minute is recorded as 1, otherwise it is recorded as 0.
[0146] Figure 2 This is the algorithm flow for periodic discharge identification. For the original EEG data, the window is first divided into non-repeated intervals of 1 minute. Then, for each minute of data, wavelet decomposition filtering and outlier detection are performed, and then the data is standardized. After extracting the peaks and troughs that meet the conditions, the peaks and troughs of the left and right brains are matched. The specific extraction conditions and matching requirements can be seen in the specific steps of the solution. If the matching percentage exceeds a certain threshold, the minute can be marked as 1.
[0147] Figure 3This is the process of spike wave identification. Taking 1s EEG data as an example, resampling is first performed to 125Hz to keep the data uniform, and then a bandpass filter is used to retain the data between 0.5Hz and 49Hz. Then Minirocket feature extraction is performed on the data, and after the feature extraction, ridge classification is used for the next step of classification.
[0148] Figure 4 This is the process of epilepsy identification. The first step is data preprocessing. The original EEG is divided into 4s intervals. Then, outlier processing is performed to replace non-numeric types such as nan and inf with 0. Then feature extraction is performed. First, db4 wavelet transform is performed to extract the relevant features of the transformed detail coefficients. When training the model, considering the imbalance of the data set, the data set is extracted and SMOTE operations are performed, and then the features are input into xgboost for classification.
[0149] Figure 5 This is a flowchart for algorithm integration. For the original EEG data, the window is divided at intervals of 1 minute, and then the data is tested. Here, the main purpose is to test whether the channel of the input data is consistent with the channel used during training. Then, after the abnormal data is converted, the spike wave is identified. Since the identification of spike waves is more than 1 second, it will be very slow if each one is carried out according to the identification process. Therefore, a data reorganization operation is added to reorganize all the EEG data of 1 minute into a new matrix, so that the filtering and feature extraction of the new data only need to call the model once, which speeds up the recognition speed. Then, epilepsy is identified based on the identification results of spike waves, and it will only be performed when spike waves appear for 4 consecutive seconds. In addition, low voltage identification, burst suppression identification and periodic discharge identification are also required after abnormal data conversion. The identification results of burst suppression, periodic discharge and epilepsy will interfere with each other. At the same time point, the priority of the report is: burst suppression> periodic discharge> epilepsy. In order to reduce false alarms, after the 1-minute recognition result comes out, it is necessary to combine the recognition results of the previous few minutes to perform a threshold judgment. Only when the proportion of people identified as abnormal for several consecutive minutes exceeds the threshold will the person be reported as abnormal.
[0150] Figure 6 Use a flowchart for the system. First, you need to export the EEG data from the EEG machine (in txt or xlsx format), then import the exported EEG data into the recognition system, which will store the original data in the database according to the patient. Then, the abnormal EEG pattern will be identified, and the recognition results and the time points of each abnormal EEG will be marked in the EEG. At the same time, the relevant results will be stored in the database to facilitate subsequent historical record queries, and an EEG report will be automatically generated according to the hospital report format, saving the time of the reader.
Claims
1. A method for intelligent identification of multiple abnormal EEG patterns, characterized in that: The method is divided into the following steps: S1. EEG burst suppression pattern recognition based on fast Fourier transform and nearest neighbor threshold comparison method. Specifically, the original EEG is segmented according to a fixed interval t, and the spectrum features after fast Fourier transform are extracted after discrete wavelet filtering and outlier processing. The nearest neighbor threshold comparison method is used to distinguish burst waves and suppression waves. S2. Recognition of periodic discharge patterns for hemispheric synchronization: the original EEG is segmented according to a fixed interval t, and hemispheric peak matching is performed after discrete wavelet filtering and outlier processing. If a match is found, it is considered that periodic discharge exists. S3, spike wave recognition based on minirocket, resample the original data, then filter it with Butterworth bandpass filter, then use minirocket to extract EEG features, and use RidgeClassifierCV for classification training; during training, no additional cross-validation is required; save the obtained spike wave recognition model; S4. Epilepsy seizure recognition based on wavelet coefficients: resample the original data, extract the wavelet coefficients after db4 decomposition after removing outliers, and use xgboost for classification training; during training, due to the large disparity between the positive and negative sample ratios, negative sample data greater than a certain threshold are extracted for each patient in training, but all positive sample data are retained; at the same time, the upsampling method is used to increase the number of positive samples; finally, the obtained epilepsy seizure recognition model is retained; S5. Low voltage discrimination based on threshold discrimination: 99.9% of the EEG data that is less than the threshold is marked as suspected low voltage, and the proportion of relevant data to the overall EEG data is counted. If it is greater than the proportion, it is considered to be low voltage; S6. Fusion of multiple abnormal EEG recognition modes. The specific steps are as follows: S6.
1. Cut the original EEG into non-repeating windows at a fixed interval t, which is the same as the value in the previous burst suppression identification. Check whether the channel of the data in each window is consistent with the channel used in training or algorithm design. If not, terminate directly. Otherwise, regroup the columns that meet the corresponding channels. Perform abnormal transformation on the combined data and resample at the same time. S6.
2. In order to improve the recognition speed, the data in the window is reorganized: Specifically, the data in the window is divided into each second, and each second is transposed and then added to a dataframe; the original data is expressed as t*frequency×channel, t*frequency is the number of EEG data in time t, frequency is the sampling frequency, and channel is the number of channels, so the reorganized data format is t*channel×frequency; S6.
3. For the reshaped data, the spike wave recognition model is called to perform spike wave recognition, with 1 indicating the presence of a spike wave and 0 indicating the absence of a spike wave. The final result is a 01 list of length t×channel. S6.
4. Integrate the above lists, with each channel as a group, and mark the existence of 1 as 1, so that the original 01 list with a length of t×channel becomes a 01 list with a length of t; S6.5, check whether there are m consecutive 1s in the 01 list. If so, call the epilepsy recognition model to perform epilepsy detection on this part of the data, otherwise proceed to S6.6; if epilepsy is detected, mark the location where the epilepsy data exists; S6.6, using the data before reformatting to perform low voltage detection using a low voltage detection algorithm; S6.7, performing wavelet decomposition filtering and abnormal data detection on the data, and then using the burst suppression and periodic discharge detection algorithm to perform burst suppression detection and periodic discharge detection; the detection results enter the corresponding result list, which retains data of a fixed length; S6.
8. If burst suppression and periodic discharge exist in the data segment at the same time, mark the final result as burst suppression; S6.9, after all windows are tested, the low voltage, burst suppression and periodic discharge test results are counted and the respective threshold determination steps are entered; S6.
10. The degree of overlap between the epilepsy detection results and burst suppression and periodic discharges. If there is overlap, the priority of the final report is: burst suppression > periodic discharge > epilepsy.
2. A method for intelligent identification of multiple abnormal EEG patterns according to claim 1, characterized in that: The specific steps of the nearest neighbor threshold comparison method in S1 are as follows: S1.
1. Set a flag to indicate the current state, initially 'end'; S1.
2. Find the beginning of the burst wave. The beginning of the burst wave is considered to be a significant change in the spectrum compared to the previous second. The formula is as follows: (fft j / fft j-1 >start_threshold)∩(flag==′end′) Among them, fft j Indicates the feature of the jth second in the current window; start_threshold indicates the threshold; S1.
3. Find the end time of the burst wave. In addition to the large change in frequency compared to the previous second, the frequency characteristics at the end must also be similar to those before the burst begins. The formula is as follows: ((fft high / fft j >end_threshold1)∪(fft j / fft low <end_threshold2))∩(flag==′start′) ∩(time<=time_threshold) Among them, fft high represents the maximum feature from the beginning of the burst wave to the jth second, fft low It indicates the characteristic value one second before the outbreak begins, time indicates the duration of the outbreak, end_threshold1, end_threshold2, and time_threshold are three thresholds; S1.
4. After the burst wave detection is completed, all segments between the burst waves are regarded as candidate suppression segments and the suppression wave detection is performed; the specific requirements for the suppression wave are as follows: (abs(Q 10 (Y j ))<5)∩(abs(Q 90 (Y j ))<5)∩(std_baofa j / 3>=std_yizhi j )∩(std_baofa j+1 / 3>=std_yizhi j )∩(std_yizhi j <3)∩(((Q 99 (Y j ) 99 (X j ) / 2.5)∩(Q 99 (Y j ) <Q 99 (X j+1 ) / 2.5))∪((abs(Q1(Y j ))<abs(Q1(X j )) / 2.5)∩(abs(Q1(Y j )) <abs(Q1(X j+1 )) / 2.5)))∩(Q1(Y j )>min(X j ))∩(Q1(Y j )>min(X j+1 )) Where abs(X) represents the absolute value of X, min(X) represents the minimum value of X, Q m (X) represents the m-quantile of X, std_baofa j Indicates the standard deviation of the j-th burst segment, std_yizhi j represents the standard deviation of the jth inhibition segment, X j represents the jth burst wave, Y j represents the j-th segment candidate suppression wave; S1.
5. Two burst waves and one suppression wave are recorded as one burst suppression, and the number of burst suppressions in each channel is counted in k consecutive windows. If it exceeds the threshold baoyi_thres, it is considered that burst suppression exists.
3. A method for intelligent identification of multiple abnormal EEG patterns according to claim 1, characterized in that: The specific steps of hemi-brain peak matching in S2 are as follows: S2.1, standardize the data of each channel, record it as a forward sequence, and then negate it to form a reverse sequence; S2.2, traverse each forward sequence, calculate the upper limit position of the background wave, the upper limit position is set as the maximum value between 3 times the upper quartile of the sequence and the upper 95%, then calculate the peak position greater than the background wave, add it to the peak list, and require that the distance between the peaks is greater than 50 data points; S2.
3. Traverse each reverse sequence and calculate the upper limit position of the background wave, but the upper limit position is set as the minimum value between -3 times the lower quartile of the corresponding forward sequence and the lower 5% of the corresponding forward sequence; the peaks larger than the background wave are added to the trough list, and the distance between the peaks is required to be greater than 50 data points; the range setting requirements for peaks and troughs are as follows: peak>max(Q 75 (E j,m )*3,Q 95 (E j,m )) peak_down<min(Q 25 (E j,m )*3,Q5(E j,m )) distance>=50 Among them, Q m (X) represents the m-quantile of X, distance represents the minimum distance between adjacent peaks and adjacent troughs; E j,m Represents the EEG data of the mth window of the jth channel, with the peak represented by peak and the trough represented by peak_down; S2.4, the left and right hemispheres are matched with peaks and troughs respectively; j,m Where j is an odd number for the left brain, and j is an even number for the right brain; for each hemisphere, the number of peaks and troughs that can be matched is calculated; spike matching is defined as the time when spikes occur in the same hemisphere within 0.2 s; S2.
5. Calculate the proportion of matching peaks and troughs in each hemibrain in the original peak and trough list. If the percentage exceeds a certain threshold pre_thres, it is considered that the current window may have periodic discharges; if periodic discharges exist in k consecutive windows, then it is determined that this person has periodic discharges.
4. A method for intelligent identification of multiple abnormal EEG patterns according to claim 2, characterized in that: In S1.5, the threshold is obtained through experiments, has no fixed range, and is related to the data itself. The patient data is first statistically analyzed, and then the number of burst suppressions that most patients have is selected.
5. A method for intelligent identification of multiple abnormal EEG patterns according to claim 3, characterized in that: In S2.5, the threshold is obtained through experiments, and the peak and trough proportions of all patients are sorted from small to large, and the value is determined according to actual needs.
6. A method for intelligent identification of multiple abnormal EEG patterns according to claim 5, characterized in that: In S2.5, the threshold is 0.
6.
7. A method for intelligent identification of multiple abnormal EEG patterns according to claim 3, characterized in that: In S2.5, 6≤k≤10.
8. A method for intelligently identifying multiple abnormal EEG patterns according to any one of claims 1 to 3, characterized in that: In S1, the fixed interval t is 10 seconds to 1 minute and should contain at least one burst suppression period.
9. A method for intelligently identifying multiple abnormal EEG patterns according to any one of claims 1 to 3, characterized in that: In S2, the fixed interval t is 10 seconds to 1 minute and should include at least one discharge cycle.
10. A method for intelligently identifying multiple abnormal EEG patterns according to any one of claims 1 to 3, characterized in that: In S4, the threshold should be smaller than the number of original negative samples.
11. A method for intelligently identifying multiple abnormal EEG patterns according to any one of claims 1 to 3, characterized in that: In S5, the threshold is between 20uV-25uV.
12. A method for intelligently identifying multiple abnormal EEG patterns according to claim 1, characterized in that: In S5, the proportion is 50%.
13. An abnormal EEG recognition system based on the method according to any one of claims 1 to 7, characterized in that: include: EEG signal preprocessing module, used for window division, outlier processing, checking of original data channels and data resampling; The burst suppression identification module extracts the spectrum features after fast Fourier transformation after discrete wavelet filtering, and uses the nearest neighbor threshold comparison method to distinguish burst waves from suppression waves; The periodic discharge recognition module performs hemi-brain peak matching after discrete wavelet filtering. If a match is found, it is considered that periodic discharge may exist. The spike wave recognition module uses a Butterworth bandpass filter, then uses minirocket to extract EEG features and inputs them into RidgeClassifierCV for discrimination; The epilepsy recognition module extracts the wavelet coefficients after db4 decomposition and uses machine learning algorithms for identification; The low voltage identification module uses the threshold discrimination method to identify low voltage; Fusion module, which is mainly responsible for coordinating the calling of burst inhibition recognition module, periodic discharge recognition module, spike wave recognition module, epilepsy recognition module, low voltage recognition module and the display priority of the later results according to the association and difference between various abnormal EEG patterns, and improving the overall recognition speed; The time labeling module mainly includes the extraction of EEG start time and subsequent real time mapping, projecting the time of the identified abnormality to the correct time point; The report generation module is performed after the detection is completed, and mainly includes the capture of images corresponding to the abnormal time and the generation of report results.
