Anesthesia depth analysis method based on deep neural network based on EEG singular spectrum analysis
Through the deep neural network method based on electroencephalogram analysis, combined with brainstem auditory induction and facial image data, the problem of low accuracy of anesthesia depth monitoring is solved, efficient and accurate prediction of anesthesia depth, and the risk of anesthesia surgery is reduced. Medical staff burden.
Patent Information
- Application Number
- CN202111264602.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-10-28
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2041-10-28
AI Technical Summary
The existing anesthesia depth monitoring methods have low prediction accuracy, resulting in a high incidence of excessive shallow or excessive depth of anesthesia, affecting patient safety and surgical effectiveness.
Using the deep neural network method based on electroencephalogram analysis, the deep learning neural network model is constructed to obtain brain waveforms and facial image data, combine brainstem auditory induction methods to mark anesthesia event, perform singular spectrum analysis and data processing, and build a deep learning neural network model to predict Depth of anesthesia.
It improves the accuracy of the prediction of anesthesia depth, reduces the workload of doctors, reduces the pressure of implementing anesthesia surgery, realizes low-cost and efficient anesthesia state analysis, and avoids labor costs and data errors.
Smart Images

Figure CN113974557B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of medical anesthesia technology, and in particular to a deep neural network anesthesia depth analysis method based on electroencephalogram singular spectrum analysis. Background Art
[0002] According to clinical statistics, only about 60% of patients can enjoy completely high-quality anesthesia services, about 14% of patients are over-anesthetized, 16% of patients are too shallowly anesthetized, and 10% of patients are in a state of alternating shallow and deep anesthesia. Among them, too deep anesthesia and overdose of drugs can cause breathing to slow down, even stop breathing, brain hypoxia, long-term brain hypoxia, cause the patient's heart to stop, and cause the patient's death. Too shallow anesthesia can cause intraoperative awareness, and the patient may have memories of the operation or even feel pain. Severe cases may also cause mental or sleep disorders. The "horrible memories" that may be triggered during the operation become another pain in postoperative life; intraoperative awareness is the recovery of consciousness during the operation under general anesthesia. In order to reduce the incidence of too shallow or too deep anesthesia, it is necessary to monitor the anesthesia state in real time to reduce the workload of medical staff and the pain of patients, and ensure safety and anesthesia effect.
[0003] The traditional judgment of anesthesia depth is achieved by observing the patient's vital signs, such as changes in body temperature, pulse, respiration, blood pressure, pupils, etc., which mainly reflects the state of autonomic nervous function during surgery. With the in-depth study of brain waves, anesthesia depth monitors based on brain wave signals have been developed, such as BIS index, Nacotend index, auditory evoked potential, etc., but both have their own advantages and disadvantages. The former is simple and easy, but it cannot be quantitatively measured at present; although the latter can be quantitatively detected, it is still affected by physiological factors, such as age, race, gender, low temperature, hypoglycemia, brain hypoxia, etc. Therefore, the prediction accuracy of the current anesthesia depth estimation method needs to be improved. Summary of the invention
[0004] The purpose of the present invention is to provide a deep neural network anesthesia depth analysis method based on EEG singular spectrum analysis, which can improve the accuracy of anesthesia depth prediction.
[0005] In order to achieve the above object, a deep neural network anesthesia depth analysis method based on EEG singular spectrum analysis is provided, comprising the following steps:
[0006] Data acquisition step: acquiring EEG waveform data and facial image data of patients during clinical surgery;
[0007] Anesthesia event labeling step: labeling each anesthesia event of the patient in an anesthesia state according to the EEG waveform data, facial image data and brainstem auditory evoked method, and obtaining the labeling results;
[0008] Data processing steps: The EEG waveform data is converted into a matrix through singular spectrum analysis and singular value decomposition is performed. The decomposed singular values are grouped and denoised, and then the original one-dimensional EEG information is reconstructed through matrix transformation;
[0009] Data analysis steps: filter the original one-dimensional EEG information to obtain EEG information within the normal frequency range of human brain activity, and then combine Fourier transform, wavelet transform, power spectrum analysis and sample entropy analysis to obtain the characteristic information of EEG information in the time domain and frequency domain;
[0010] Anesthesia depth prediction steps: construct a deep learning neural network model, input the annotation results and feature information into the neural network model, and output the patient's anesthesia depth prediction results.
[0011] Principles and advantages:
[0012] 1. The setting of the anesthesia event marking step can objectively and sensitively reflect the function of the central nervous system through the brainstem auditory evoked method, EEG waveform data, and facial image data, so as to facilitate the understanding of the patient's current anesthesia state and realize the reverse marking of anesthesia events. The reverse marking of anesthesia events can facilitate the analysis of the patient's current anesthesia state, thereby facilitating the analysis of the patient's anesthesia depth according to the anesthesia state, and thus providing data support for the prediction of the anesthesia depth in the later stage.
[0013] 2. The setting of data processing steps and data analysis steps. The data processing step converts the EEG waveform data into a matrix and performs singular value decomposition through singular spectrum analysis, and reconstructs the original one-dimensional EEG information through matrix transformation, so as to reduce the dimension of high-dimensional EEG waveform data to low-dimensionality, which greatly reduces the amount of data processing, thereby improving the efficiency of data analysis and processing. Then the original one-dimensional EEG information is reconstructed through matrix transformation; and the original one-dimensional EEG information is filtered, and combined with Fourier transform, wavelet transform, power spectrum analysis and sample entropy analysis, the invalid data in the EEG data can be fully eliminated, so as to obtain EEG information that is more deeply related to the depth of anesthesia. Finally, by constructing a deep learning neural network model and inputting the annotation results and feature information into it, the patient's anesthesia depth prediction results can be obtained quickly and accurately, which greatly reduces the workload of doctors and reduces the pressure of implementing anesthesia surgery.
[0014] Furthermore, the singular spectrum analysis method in the data processing step includes the following sub-steps:
[0015] Trajectory matrix construction sub-step: converting the original EEG signal in the EEG waveform data into a multi-dimensional trajectory matrix X;
[0016] Singular value decomposition sub-step: construct the covariance matrix, convert the asymmetric multi-dimensional trajectory matrix into a symmetric square matrix, and perform eigenvalue decomposition to obtain the eigenvalues in descending order and the corresponding eigenvectors.
[0017] For a symmetric square matrix, the eigenvalues and singular values are equal, so the eigenvalues and eigenvectors of the multidimensional trajectory matrix X can be quickly obtained through singular value decomposition, which facilitates the dimensionality reduction of the data to reduce the amount of data processing and thus improve the data processing efficiency.
[0018] Furthermore, the original EEG signal is a single-channel time series signal s=(s 1 ,s 2 ,……,s N ) T ,The formula of the multidimensional trajectory matrix X is as follows:
[0019]
[0020] N is the length of the original EEG signal, and L is the embedded time window of the trajectory matrix, where L <N,K=N-L+1。
[0021] Furthermore, the data processing step specifically includes the following sub-steps:
[0022] Grouping artifact removal and denoising sub-steps: grouping the eigenvalues and eigenvectors obtained from singular value decomposition according to the eigenvalue change rate formula; removing artifacts from the original EEG signal according to the artifact removal formula and the fixed threshold set in the artifact removal formula; removing noise from the original EEG signal according to the size of the singular value, and obtaining a grouping result with characteristic significance.
[0023] Furthermore, the formula for removing artifacts from the original EEG signal is as follows:
[0024]
[0025] When the maximum amplitude in the original EEG signal is greater than the fixed threshold V 0 When , RC1 and RC2 are classified as artifacts, otherwise RC1 is classified as an artifact; RC1 and RC2 are both composed of the eigenvectors corresponding to the eigenvalues of the group.
[0026] Furthermore, the data processing step specifically includes the following sub-steps:
[0027] Matrix reconstruction sub-step: Select a valid eigenvector group according to the eigenvalue change rate formula, and reconstruct the trajectory matrix according to the reconstruction formula.
[0028] Further, the characteristic value change rate formula is as follows:
[0029]
[0030] Among them, λ i is the eigenvalue, i <j<L。
[0031] Further, the reconstruction formula is as follows:
[0032]
[0033] in, X i is the trajectory matrix after reorganization of a specific eigenvector.
[0034] Furthermore, the anesthesia event marking step specifically includes the following steps:
[0035] Reverse annotation sub-step: extracting feature data according to the EEG data, and reversely annotating the anesthesia event according to the feature data to obtain a first annotation result; performing image recognition on the collected facial image data to obtain a second annotation result of each anesthesia event;
[0036] Neural network model labeling sub-step: associate the EEG data, feature data and the first labeling result and store them as the first cumulative data, and use the facial image data and the second labeling result as the second cumulative data, build, train and optimize the neural network model based on the first cumulative data and the second cumulative data, and input the facial image data of the patient to be tested into the neural network model for automatic judgment and labeling of anesthesia events, and obtain a third labeling result.
[0037] By extracting characteristic data from brain wave data, the function of the central nervous system can be reflected more objectively and sensitively, so as to facilitate understanding of the patient's anesthesia state at this time, and then realize the reverse marking of anesthesia events. Its advantage is that the marking of anesthesia events is more accurate, but because the waveform of brain waves is updated at a high frequency, it is necessary to arrange medical staff to watch the auditory evoked potential monitor in real time and record relevant brain wave data, which is easy to distract medical staff. When anesthetizing a patient, the doctor needs to perform relevant surgical operations on the patient's face. The facial image data of the patient during the operation can be obtained through the camera of the intelligent terminal, and the doctor's surgical operation can be collected. Since the surgical operation will be fed back to the patient's face, it is then recognized by image recognition. It is convenient to analyze the patient's anesthesia state and complete the marking of anesthesia events. Its advantage is that the marking of anesthesia events does not require real-time attention from medical staff, making it difficult for medical staff to be distracted and more convenient. Finally, a neural network model is constructed, trained and optimized according to the first cumulative data and the second cumulative data, and the facial image data of the patient to be tested is input into the neural network model for automatic judgment and marking of anesthesia events, and the third marking result is obtained. Among them, it is very convenient to obtain the facial data of the patient to be tested. Only a camera needs to be installed, which is low-cost and profitable. However, the cost of brain wave equipment is high, and the operation is relatively complicated when used. It is necessary to wear head-mounted sensors and other equipment, and perform equipment calibration operations, which is very inconvenient. This solution uses a neural network model to associate the data of the two, that is, to find the association between facial data and brain wave data, and quickly associate and analyze the corresponding brain wave data through facial data. The analyzed brain wave data is used to further confirm the depth of anesthesia, achieving low cost and high efficiency while ensuring the measurement accuracy of the depth of anesthesia, and complementing the advantages and disadvantages of the two. Thereby, the accuracy of the analysis of the anesthesia state can be further improved, and at the same time, there is no need to arrange medical staff to pay attention and record in real time, which saves labor costs and is more efficient. Thereby, the patient's anesthesia state can be automatically recorded and marked, avoiding errors caused by the recorder misreading the time, writing the wrong content by mistake, etc., resulting in the recorded data being unfavorable for subsequent clinical anesthesia research and application.
[0038] Furthermore, the data acquisition step includes: acquiring static data, physical data and surgical data of the patient before surgery; the static data includes height, weight, gender, age and medical history;
[0039] The method further comprises the following steps:
[0040] Data storage management steps: Screen and classify patients according to their static data, remove patients with brain or neurological abnormalities, and divide the scope of clinical cases.
[0041] By screening and classification, accidental errors caused by special cases can be eliminated, thus ensuring the universality of the data. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Figure 1 This is a flowchart of a method for analyzing anesthesia depth using a deep neural network based on EEG singular spectrum analysis in Embodiment 1 of the present invention;
[0043] Figure 2 for Figure 1 Flow chart of the data processing steps in . DETAILED DESCRIPTION
[0044] The following is further described in detail through specific implementation methods:
[0045] Embodiment 1
[0046] A deep neural network anesthesia depth analysis method based on EEG singular spectrum analysis, the equipment used includes a smart terminal, a Bluetooth headset and an auditory evoked potential monitor. The smart terminal is equipped with an APP, which is used for communication connection with the Bluetooth headset to send short sound stimulation information. In this embodiment, the smart terminal uses a tablet computer with a camera, and the tablet computer is fixed on the operating table by a bracket. The Bluetooth headset and the tablet computer are paired via Bluetooth to achieve remote wireless communication connection. Basically as shown in the attached Figure 1 , Figure 2 As shown, the following steps are included:
[0047] Data acquisition step: acquiring EEG waveform data and facial image data of the patient during clinical surgery; acquiring static data, physical sign data and surgical data of the patient before surgery; the static data includes height, weight, gender, age and medical history;
[0048] Patient screening and management steps: Screen and classify patients according to their static data, remove patients with brain or neurological abnormalities, and divide the range of clinical cases.
[0049] Anesthesia event labeling step: labeling each anesthesia event of the patient in an anesthesia state according to the EEG waveform data, facial image data and brainstem auditory evoked method, and obtaining the labeling results;
[0050] The anesthesia event marking step specifically includes the following steps:
[0051] S001. Before the operation, the patient should wear Bluetooth headphones and an auditory evoked potential monitor;
[0052] S002, equipment initialization, check whether the Bluetooth headset, auditory evoked potential monitor and camera are connected normally;
[0053] S101. Before the operation begins, the patient's identity information and configuration information of short sound stimulation information are input into the intelligent terminal, wherein the configuration information includes the content, volume and playing time interval of the stimulation information content. The stimulation information content is the patient's name and requires the patient to blink when hearing the call.
[0054] S102, retrieve the patient's hearing sign data from the hospital's medical record database according to the patient's identity information, and optimize the volume setting in the configuration information according to the hearing sign data, so as to adjust the volume to a suitable level for subsequent anesthesia status analysis.
[0055] S1. After the operation begins, the camera of the smart terminal is pointed at the patient's face to obtain the patient's facial image data (ie, facial video image data) during the operation; after the operation is completed, the facial image data is saved.
[0056] S2. At the beginning of the operation, a short sound stimulation message is sent to the Bluetooth headset worn by the patient through the smart terminal, and the patient's brain wave data is recorded using the auditory evoked potential monitor worn by the patient; when the patient hears the stimulation message content in the anesthetized state, the patient will respond, so there will be waveform information (characteristic waves) corresponding to the cycle (there is a phase difference with the playback time interval of the stimulation message content) in the brain wave data. There are differences between the characteristic waves in the anesthetized state and the characteristic waves in the non-anesthetized state, which are specifically manifested in the frequency and amplitude of the characteristic waves.
[0057] S3, extracting feature data according to the brain wave data, and reversely annotating the anesthesia event according to the feature data to obtain a first annotation result; performing image recognition on the collected facial image data to obtain a second annotation result of each anesthesia event;
[0058] The step S3 further comprises the following steps:
[0059] S301, extracting characteristic data according to the brain wave data, wherein the characteristic data includes the amplitude, phase, and frequency of the brain wave;
[0060] S302. Combine the characteristic data with the characteristics of the frequency domain and the analysis of the power spectrum to reversely annotate the anesthesia event and obtain the first annotation result, which includes the time nodes of each anesthesia event; when the patient is in a non-anesthetized state, when the patient hears the stimulation information content, the response is relatively strong, the power is relatively large, the amplitude is relatively large, and the frequency can basically coincide with the playback frequency of the stimulation information content. In the anesthetized state, the patient's central nervous system is shielded, and when the patient hears the stimulation information content, the response is relatively weak, the power is relatively small, the amplitude is relatively small, the frequency will decrease, and the phase of the characteristic wave will change to a certain extent. The change in the strength of the shielding will cause corresponding changes in the power, amplitude, frequency and phase. Therefore, reverse annotation of anesthesia events can be performed.
[0061] S303, performing image recognition on the collected facial image data to obtain a second annotation result of each anesthesia event, wherein the second annotation result includes image data corresponding to each anesthesia event, and the image data includes image frame data from the start time point to the end time point of the anesthesia event. The image frame data contains corresponding time information. When the doctor performs a surgical operation related to the anesthesia event, the patient's feedback experience appears on the face, so by performing image recognition on the facial image data, the patient's anesthesia event can be indirectly annotated.
[0062] S304. According to the time nodes of each anesthesia event in the first annotation result, the key image frame data corresponding to each anesthesia event in the second annotation result is screened and matched. If there is matching key image frame data, the first annotation result and the second annotation result are marked as normal data. If there is no matching key image frame data or it is not completely matched, the first annotation result and the second annotation result are marked as abnormal data.
[0063] S4, the brain wave data, the feature data and the first annotation result are associated and stored as the first cumulative data, and the facial image data and the second annotation result are used as the second cumulative data, a neural network model is constructed, trained and optimized according to the first cumulative data and the second cumulative data, and the facial image data of the patient to be tested is input into the neural network model for automatic judgment and annotation of anesthesia events, and a third annotation result is obtained;
[0064] S401, when the first annotation result and the second annotation result are marked as abnormal data, the first accumulated data and the second accumulated data are sent to the doctor, the annotation data of the anesthesia event re-annotated by the doctor for the first annotation result and the second annotation result are obtained, and the anesthesia event marked in the first annotation result and / or the second annotation result is replaced according to the annotation data;
[0065] S402, randomly extracting first accumulated data and second accumulated data marked as normal data by the first and second annotation results, and sending the first and second accumulated data to a doctor, to obtain annotated data in which the doctor re-annotates anesthesia events for the first and second annotation results;
[0066] S403, accumulate a set amount of labeled data and the first cumulative data and the second cumulative data, and construct, train and optimize the neural network model based on the labeled data and the first cumulative data and the second cumulative data. In this embodiment, the neural network model is modeled using a deep neural network algorithm, using a structure of multiple hidden layers and multiple neural network units, and considering the problems of overfitting and underfitting. The deep neural network algorithm is a mature technology and will not be described in detail in this embodiment.
[0067] S5. Output the second annotation result and associate it with the patient.
[0068] Data processing steps: The EEG waveform data is converted into a matrix through singular spectrum analysis and singular value decomposition is performed. The decomposed singular values are grouped and denoised, and then the original one-dimensional EEG information is reconstructed through matrix transformation. Compared with traditional filtering methods, this method reconstructs the original one-dimensional EEG information, which provides more representative time-domain EEG information than the original EEG waveform and can more accurately analyze the patient's anesthesia status.
[0069] The singular spectrum analysis method in the data processing step includes the following sub-steps:
[0070] Trajectory matrix construction sub-step: converting the original EEG signal in the EEG waveform data into a multi-dimensional trajectory matrix X;
[0071] The original EEG signal is a single-channel time series signal s=(s 1 ,s 2 ,……,s N ) T ,The formula of the multidimensional trajectory matrix X is as follows:
[0072]
[0073] N is the length of the original EEG signal, and L is the embedded time window of the trajectory matrix, where L <N,K=N-L+1。
[0074] Singular value decomposition sub-step: construct the covariance matrix C = XX T , convert the asymmetric multidimensional trajectory matrix into a symmetric square matrix, and perform eigenvalue decomposition to obtain the eigenvalues λ in descending order i (λ 1 ≥λ 2 ≥…≥λ L ≥0), and the corresponding eigenvector V i The calculation of singular value decomposition (SVD) has been fully disclosed in the prior art and will not be described in detail in this embodiment.
[0075] Grouping denoising and artifact removal sub-step: perform waveform component analysis and extraction based on the eigenvalues and eigenvectors obtained after singular value decomposition. The waveform is a composite wave, and the main components include three partial features, namely artifacts, oscillations, and noise. The oscillation feature needs to be retained, while artifacts and noise need to be removed.
[0076] Artifacts in the original EEG signal are removed according to the artifact removal formula and the fixed threshold set in the artifact removal formula; noise in the original waveform is removed according to the size of the singular value, and then grouping is performed according to the eigenvalue change rate formula to obtain eigenvalue grouping results with characteristic significance. The eigenvalue change rate of the eigenvalues corresponding to the previous waveform and the next waveform in the same time window is less than 5%, indicating that the previous waveform and the next waveform have little change and can be considered to be the same waveform. If the eigenvalue change rate is greater than or equal to 5%, it means that the previous waveform and the next waveform are two different groups of waveforms, thereby achieving grouping.
[0077] The formula for removing artifacts from the original EEG signal is as follows:
[0078]
[0079] When the maximum amplitude in the original EEG signal is greater than the fixed threshold V 0 , RC1 and RC2 are classified as artifacts, and vice versa, RC1 is classified as artifact; RC is a construction component, and the RC division standard is similar eigenvalues, that is, eigenvalues with small changes before and after are classified as the same RC, and RC is composed of eigenvectors corresponding to this group of eigenvalues. The common EEG signal size is less than 100μV. In this embodiment, the fixed threshold V 0 is 200μV.
[0080] Matrix reconstruction sub-step: Select a valid eigenvector group according to the eigenvalue change rate formula, and reconstruct the trajectory matrix according to the reconstruction formula.
[0081] The characteristic value change rate formula is as follows:
[0082]
[0083] Among them, λ i is the eigenvalue, i <j<L。
[0084] The reconstruction formula is as follows:
[0085]
[0086] in, X i is the trajectory matrix after reorganization of a specific eigenvector.
[0087] Data analysis steps: filter the original one-dimensional EEG information to obtain EEG information within the frequency range of normal human brain activity. EEG information includes several sinusoidal waveform data, such as sinusoidal waveforms in the α, β, θ, and δ frequency bands. Fourier transform is performed on the time domain information of EEG information to achieve time domain frequency domain transformation, and frequency domain information is extracted for analysis. Considering the problems caused by the non-steady-state waveform of EEG in the transformation process, wavelet transform is further used to obtain frequency domain information while taking into account time domain information. Combined with power spectrum analysis and sample entropy, characteristic information of EEG information in the time domain and frequency domain is obtained; for example, the ratio of high frequency to low frequency in EEG information, and the energy value of sinusoidal waveforms in the α, β, θ, and δ frequency bands.
[0088] Fourier transform is used to transform signals between time domain and frequency domain. Its purpose is to obtain the characteristic expression of time domain information in frequency domain through Fourier transform. Its Fourier transform only acts on steady-state waveforms and has a long window time. Discrete Fourier transform is used for frequency domain conversion. For the sequence {x[n]}0≤n≤N, its discrete Fourier transform is:
[0089]
[0090] Wavelet transform refers to an analysis method that uses the oscillating waveform of a mother wavelet of finite length or rapid decay to identify the time domain and frequency domain characteristics of the original signal. Wavelet transform can reflect the transformed frequency domain information while retaining the original time domain information. Wavelet transform acts on non-steady-state waveforms with a short window time, providing a non-steady-state variable-duration analysis method. Discrete wavelet transform is used to convert and extract the time-frequency characteristics of the original signal. The discrete wavelet transform formula is as follows:
[0091]
[0092] Wavelet transform can be regarded as calculating the convolution of the mother wavelet and the original signal in the time domain at multiple scales (frequencies), obtaining the similarity coefficient between the wavelet and the original signal at each time point and each scale, and thereby analyzing the frequency domain composition of the original signal at each time point.
[0093] Power spectrum analysis is based on signal frequency domain classification, and the power spectra corresponding to the α, β, θ, and δ frequency bands are calculated respectively. The original one-dimensional EEG information of length N is divided into L segments, and the length of each data segment is M, that is, N = LM. The window function w is used on each signal segment to find the power spectrum of each signal segment. Finally, the power spectrum of each signal segment is averaged to obtain the power spectrum of the entire signal segment. The power spectrum density calculation formula is as follows:
[0094]
[0095] Sample entropy is an algorithm for analyzing the complexity of time series. The larger the sample entropy, the greater the complexity, and vice versa. The calculation method of sample entropy is as follows;
[0096] Suppose a time series N consisting of N data, N = {x 1 , x 2 , x 3 , ……, x N};
[0097] (1) Divide N into multiple subsets with a step size of m. The formula is as follows: X m (i) = {x i , x i+1 , x i+2 , ……, x i+m-1};
[0098] (2) Calculate the Chebyshev distance between subsets respectively, d[X m (i), X m (j)] (i ≠ j). The Chebyshev distance is the absolute value of the maximum value of the differences at corresponding positions in two sequences. The formula is as follows:
[0099] max k=0,…,m-1 (|x(i + k) - x(j + k)|)
[0100] (3) Calculate the sample entropy. The formula is as follows:
[0101]
[0102] where A is the total number of all i that satisfy d[X m+1 (i), X m+1 (j)] < r; B is the total number of all i that satisfy d[X m (i), X m (j)] < r. Fourier transform, wavelet transform, power spectrum analysis, and sample entropy are relatively mature data processing techniques. In this embodiment, no specific improvements are involved, so they will not be elaborated too much.
[0103] Anesthesia depth prediction step: construct a deep learning neural network model, input the annotation results and feature information into the neural network model, and output the patient's anesthesia depth prediction result. In this embodiment, a deep neural network (DNN) is used, and the deep neural network uses multiple hidden layers (specifically 5 hidden layers), the number of neural network units is 2048, 1024, 512, 256, 128, and the model parameters such as dropout and learning rate are set. The epoch of the neural network model is set to 100000, the initial Learning Rate is set to 0.001, and the dropout ratio is 0.3 (that is, 30% of the neurons will be randomly shielded during the training process, which is regarded as a model penalty term to prevent the model from overfitting), and the loss function is selected as the log-likelihood loss function.
[0104] Embodiment 2
[0105] At present, when patients undergo surgery, they must fill out a surgery notification or informed consent form, which must be signed by their family members. First, some patients go to the hospital for surgery without the knowledge of their family members. In this case, there is no one to take care of the patients, and if problems arise during or after the operation, there is no family member to take care of them, which can easily cause problems. Second, the patient's family members also have the right to informed consent, and medical staff have the obligation and responsibility to communicate with the patient's family members about the patient's condition to gain their support and trust. Third, after the patient loses consciousness after general anesthesia, if there are unpredictable risks, it is necessary to communicate with the patient's family members, and only after the patient's family members sign and confirm can the next step of diagnosis and treatment be carried out.
[0106] The difference between the second embodiment and the first embodiment is that the second embodiment further includes the following steps:
[0107] The first family member data is retrieved from the hospital's medical record database according to the patient's identity information; the first family member data includes the family member's name, relationship with the patient and contact information.
[0108] Video surveillance data is retrieved from the hospital based on the patient's facial data, and facial recognition is performed based on the video surveillance data to obtain data on personnel who accompany the patient to the hospital for treatment; the personnel data includes name and contact information.
[0109] The second family member data of the patient is obtained according to the surgery notification form or informed consent form; the second family member data includes the family member's name, relationship with the patient and contact information.
[0110] Different weight scores are set for the first family member data, personnel data, and second family member data, and the reference score of the relationship progress is set to comprehensively calculate the total score of the relationship progress of each person. Then, the total score of the relationship progress of each person is ranked in descending order, and the person ranked first is defined as the closest person. Among them, when the first family member data or the personnel data is empty, the data is analyzed and calculated as zero value.
[0111] Voice communication is conducted with the closest person according to their contact information, and the voice information set by them is obtained through voice communication. For example, the information content of the voice information is the patient's name, and the patient is required to blink when hearing the call. The voice information is then edited, and the playback interval and cycle are set to obtain short sound stimulation information.
[0112] In this embodiment, the closest persons are analyzed by obtaining the patient's family data, close personnel data, etc., so as to obtain the voice information of the closest persons. The voice information will be produced into short sound stimulation information of auditory evoked potential technology, which will make the patient feel more intimate and more likely to respond after the patient is anesthetized, thereby making the amplitude of the response signal in the brain wave larger, which is more convenient for analyzing and predicting the depth of anesthesia.
[0113] The above is only an embodiment of the present invention. The common sense such as the known specific structure and characteristics in the scheme is described too much here. The ordinary technicians in the relevant field know all the common technical knowledge in the technical field of the invention before the application date or priority date, can know all the existing technologies in the field, and have the ability to apply the conventional experimental means before that date. The ordinary technicians in the relevant field can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for ordinary technicians in the relevant field to implement this application. It should be pointed out that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the protection scope of the present invention, which will not affect the effect of the implementation of the present invention and the practicality of the patent. The protection scope required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.
Claims
1. A deep neural network anesthesia depth analysis method based on EEG singular spectrum analysis, characterized in that: The following steps are involved: Data acquisition step: acquiring EEG waveform data and facial image data of patients during clinical surgery; Anesthesia event labeling step: labeling each anesthesia event of the patient in an anesthesia state according to the EEG waveform data, facial image data and brainstem auditory evoked method, and obtaining the labeling results; Data processing steps: The EEG waveform data is converted into a matrix through singular spectrum analysis and singular value decomposition is performed. The decomposed singular values are grouped and denoised, and then the original one-dimensional EEG information is reconstructed through matrix transformation; Data analysis steps: filter the original one-dimensional EEG information to obtain EEG information within the normal frequency range of human brain activity, and then combine Fourier transform, wavelet transform, power spectrum analysis and sample entropy analysis to obtain the characteristic information of EEG information in the time domain and frequency domain; Anesthesia depth prediction step: construct a deep learning neural network model, input the annotation results and feature information into the neural network model, and output the patient's anesthesia depth prediction result; The steps of anesthesia event marking include: S3, extracting feature data according to the brain wave data, and reversely annotating the anesthesia event according to the feature data to obtain a first annotation result; performing image recognition on the collected facial image data to obtain a second annotation result of each anesthesia event; The step S3 further comprises the following steps: S301, extracting characteristic data according to the brain wave data, wherein the characteristic data includes the amplitude, phase, and frequency of the brain wave; S302, combining the characteristic data with the characteristics of the frequency domain and the analysis of the power spectrum to reversely annotate the anesthesia event, and obtain a first annotation result, wherein the first annotation result includes the time nodes of each anesthesia event; S303, performing image recognition on the collected facial image data to obtain a second annotation result of each anesthesia event, wherein the second annotation result includes image data corresponding to each anesthesia event, and the image data includes image frame data from a start time point to an end time point of the anesthesia event; S304, according to the time nodes of each anesthesia event in the first annotation result, filter and match the key image frame data corresponding to the time node from the image data corresponding to each anesthesia event in the second annotation result, if there is a matching key image frame data, mark the first annotation result and the second annotation result as normal data, if there is no matching key image frame data or it is not completely matched, mark the first annotation result and the second annotation result as abnormal data; S4, the brain wave data, the feature data and the first annotation result are associated and stored as the first cumulative data, and the facial image data and the second annotation result are used as the second cumulative data, a neural network model is constructed, trained and optimized according to the first cumulative data and the second cumulative data, and the facial image data of the patient to be tested is input into the neural network model for automatic judgment and annotation of anesthesia events, and a third annotation result is obtained; The step S4 further comprises the following steps: S401, when the first annotation result and the second annotation result are marked as abnormal data, the first accumulated data and the second accumulated data are sent to the doctor, the annotation data of the anesthesia event re-annotated by the doctor for the first annotation result and the second annotation result are obtained, and the anesthesia event marked in the first annotation result and / or the second annotation result is replaced according to the annotation data; S402, randomly extracting first accumulated data and second accumulated data marked as normal data by the first and second annotation results, and sending the first and second accumulated data to a doctor, to obtain annotated data in which the doctor re-annotates anesthesia events for the first and second annotation results; S403, accumulating a set amount of labeled data and the first accumulated data and the second accumulated data, and constructing, training and optimizing a neural network model based on the labeled data and the first accumulated data and the second accumulated data; S5, output the second annotation result and associate it with the patient; The method further includes: retrieving the first family member data of the patient from the medical record database of the hospital according to the patient's identity information; the first family member data includes the family member's name, relationship with the patient and contact information; Retrieving video surveillance data from the hospital based on the patient's facial data, and performing facial recognition based on the video surveillance data to obtain data on personnel who accompanied the patient to the hospital for treatment; the personnel data includes name and contact information; Obtain the patient's second family member data based on the surgery notification or informed consent form; the second family member data includes the family member's name, relationship with the patient, and contact information; Different weight scores are set for the first family member data, personnel data and the second family member data, and the relationship progress reference score is set to comprehensively calculate the total relationship progress score of each person; then the total relationship progress scores of each person are ranked in descending order, and the person ranked first is defined as the closest person; wherein, when the first family member data or the personnel data is empty, the data is analyzed and calculated as zero value; The closest person is contacted by voice according to their contact information, and the voice information set by them is obtained through voice communication.
2. The deep neural network anesthesia depth analysis method based on EEG singular spectrum analysis according to claim 1 is characterized in that: The singular spectrum analysis method in the data processing step includes the following sub-steps: Trajectory matrix construction sub-step: converting the original EEG signal in the EEG waveform data into a multi-dimensional trajectory matrix X; Singular value decomposition sub-step: construct the covariance matrix, convert the asymmetric multidimensional trajectory matrix into a symmetric square matrix, and perform eigenvalue decomposition to obtain the eigenvalues in descending order , and the corresponding eigenvector .
3. The deep neural network anesthesia depth analysis method based on EEG singular spectrum analysis according to claim 2 is characterized in that: The original EEG signal is a single-channel time series signal ,The formula of the multidimensional trajectory matrix X is as follows: N is the length of the original EEG signal, and L is the embedded time window of the trajectory matrix, where L <N,K=N-L+1。 4. The deep neural network anesthesia depth analysis method based on EEG singular spectrum analysis according to claim 3 is characterized in that: The data processing step specifically includes the following sub-steps: Grouping artifact removal and denoising sub-steps: grouping the eigenvalues and eigenvectors obtained from singular value decomposition according to the eigenvalue change rate formula; removing artifacts from the original EEG signal according to the artifact removal formula and the fixed threshold set in the artifact removal formula; removing noise from the original EEG signal according to the size of the singular value, and obtaining a grouping result with characteristic significance.
5. The deep neural network anesthesia depth analysis method based on EEG singular spectrum analysis according to claim 4 is characterized in that: The formula for removing artifacts from the original EEG signal is as follows: When the maximum amplitude in the raw EEG signal is greater than a fixed threshold When , RC1 and RC2 are classified as artifacts, otherwise RC1 is classified as an artifact; RC1 and RC2 are both composed of the eigenvectors corresponding to the eigenvalues of the group.
6. The deep neural network anesthesia depth analysis method based on EEG singular spectrum analysis according to claim 5 is characterized in that: The data processing step specifically includes the following sub-steps: Matrix reconstruction sub-step: Select a valid eigenvector group according to the eigenvalue change rate formula, and reconstruct the trajectory matrix according to the reconstruction formula.
7. The method for deep neural network anesthesia depth analysis based on EEG singular spectrum analysis according to claim 6, characterized in that: The characteristic value change rate formula is as follows: , , in, is the characteristic value, .
8. The method for deep neural network anesthesia depth analysis based on EEG singular spectrum analysis according to claim 6, characterized in that: The reconstruction formula is as follows: in, , is the trajectory matrix after reorganization of a specific eigenvector.
9. The deep neural network anesthesia depth analysis method based on EEG singular spectrum analysis according to claim 1 is characterized in that: The data acquisition step includes: acquiring static data of the patient before surgery; the static data includes height, weight, gender, age and medical history; The method further comprises the following steps: Patient screening and management steps: Screen and classify patients according to their static data, remove patients with brain or neurological abnormalities, and divide the range of clinical cases.
