Children esketamine anesthesia-oriented sedation level monitoring system and monitoring method

Abstract

The invention discloses a sedation level monitoring system and monitoring method for children esketamine anesthesia, and belongs to the technical field of medical anesthesia, the sedation level monitoring system comprises a power supply module, a sensor module, a preprocessing module, a memory module, a central processing module, a communication module and an upper computer module; the power supply module supplies power to the system; the sensor module is used for collecting electroencephalogram signals of a patient in real time; the preprocessing module is used for carrying out noise reduction processing on the collected electroencephalogram signals; the memory module is used for storing the static sedation level evaluation neural network model after offline training and various data generated during system operation; the central processing module is used for calling the stored neural network model to judge the sedation level of the patient; the communication module is used for sending a judgment result of the model on the sedation level of the patient to the upper computer module; the upper computer module is used for displaying the sedation level of the patient. According to the invention, real-time sedation level monitoring service can be provided for 3-18 years old children in an esketamine anesthesia scene.

Description

Technical Field

[0001] This invention relates to the field of medical anesthesia technology, and in particular to a sedation level monitoring system and method for esketamine anesthesia in children. Background Art

[0002] Sedation level assessment plays a crucial role in perioperative management, helping to prevent cognitive impairment due to excessive anesthesia and intraoperative awareness due to insufficient anesthesia. Although some EEG-based clinical anesthesia monitoring systems exist (such as the bispectral index), most are only applicable to γ-aminobutyric acid (GABA) receptor-mediated anesthetics (such as propofol) and not to N-methyl-D-aspartate receptor antagonists (such as esketamine). This is primarily because different pharmacological mechanisms produce different EEG characteristics. For example, propofol increases the power of low-frequency alpha and delta bands and inhibits gamma wave activity; esketamine, on the other hand, enhances gamma rhythm power and has less impact on low-frequency rhythms. This difference in EEG characteristics due to pharmacological mechanisms makes traditional monitoring systems unreliable in esketamine anesthesia.

[0003] Furthermore, the influence of neurodevelopmental factors on EEG characteristics cannot be ignored. For example, under propofol anesthesia, the power of the theta and gamma bands in children exhibits a unique age-related variation pattern, with peak power occurring in the 3-6 year old age group, and then gradually decreasing with age. This difference leads to an important clinical observation: children's bispectral index values ​​during loss of consciousness and recovery are significantly higher than those in adults, which may lead to misjudgment of sedation levels by anesthesiologists.

[0004] Therefore, there is an urgent need for a sedation level monitoring system suitable for children and esketamine anesthesia scenarios, which can provide real-time and reliable sedation level monitoring services to guide anesthesiologists in drug administration, ensure patient safety during anesthesia, and improve patient prognosis. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a sedation level monitoring system and method for children under esketamine anesthesia, so as to provide targeted real-time sedation level monitoring for children under esketamine anesthesia.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A sedation level monitoring system for children under esketamine anesthesia includes a power module, a sensor module, a preprocessing module, a storage module, a central processing module, a communication module, and a host computer module.

[0008] The power module is used to provide power to the system.

[0009] The sensor module is used to collect the patient's electroencephalogram (EEG) signals in real time.

[0010] The preprocessing module is used to perform noise reduction processing on the acquired EEG signals.

[0011] The storage module is used to store the static sedation level evaluation neural network model that has completed offline training, as well as the log data, monitoring data and event tagging data generated during system operation.

[0012] The central processing module is used to call the neural network model for assessing the sedation level of children stored in the storage module. The neural network model for assessing the sedation level of children judges the sedation level of the patient based on the input EEG.

[0013] The communication module is used to send the judgment results of the pediatric sedation level assessment neural network model on the patient's sedation level to the host computer module.

[0014] The host computer module is used to display the patient's sedation level.

[0015] A method for monitoring sedation levels in children under esketamine anesthesia includes the following steps:

[0016] S1. Power on the power module, and the system begins to run;

[0017] S2. Place the sensor module on the patient's forehead and use the sensor module to collect single-channel EEG signals from the patient's forehead.

[0018] S3. The preprocessing module is used to reduce noise in the acquired single-channel EEG signals.

[0019] S4. The central processing module calls the neural network model for assessing the sedation level of children stored in the storage module. The neural network model for assessing the sedation level of children judges the sedation level of patients based on the input EEG, including a multi-scale time-frequency feature extraction component, a feature encoding component, and a contrast classification component.

[0020] S5. Use the communication module to send the judgment results of the sedation level of the patient by the neural network model for assessing the sedation level of the child to the host computer module.

[0021] S6. Use the host computer module to display the patient's sedation level for real-time monitoring.

[0022] A further improvement to the technical solution of the present invention is that S3 specifically includes the following steps:

[0023] S301. The EEG signal is amplified by a preamplifier circuit, with an amplification factor of 1000 times;

[0024] S302. Low-frequency noise in EEG signals is filtered out based on a hardware high-pass filter, with a cutoff frequency of 0.1Hz.

[0025] S303. High-frequency components in EEG signals are filtered out based on a hardware low-pass filter with a cutoff frequency of 60Hz.

[0026] S304. Converting EEG signals into digital signals based on analog-to-digital conversion circuits;

[0027] S305. Power frequency noise in EEG signals is filtered out based on a digital notch filter with a cutoff frequency of 49-51Hz.

[0028] S306. Based on the downsampling algorithm, the EEG signal is downsampled to a specified frequency;

[0029] The specified downsampling frequency is 100Hz;

[0030] S307. Save the raw EEG and the preprocessed EEG to the storage module.

[0031] A further improvement to the technical solution of this invention lies in: In S4, the neural network model for assessing the sedation level of children determines the patient's sedation level based on the input EEG, specifically including the following steps:

[0032] S401. Based on a multi-scale time-frequency feature extraction component, extract time-domain and frequency-domain features at different scales from the EEG samples input to the neural network model for assessing pediatric sedation levels, and obtain the local time-domain feature T. local Temporal global features T global Frequency domain local features F local and frequency domain global features T global ;

[0033] S402, Based on the feature encoding component, from the input T local T global F local and T global Extracting local time-frequency features O local Time-frequency global features O global and sample feature representation D;

[0034] S403, Based on the contrastive classification component, from the input O local O global Obtain the time-frequency local feature projection Z from D. local Time-frequency global feature projection Z global And the probability output Y, where Y contains the predicted probabilities of 5 different sedation levels;

[0035] During the offline training phase of the neural network model for assessing sedation levels in children, based on Z... local Z global And Y calculate the total loss L of the neural network model for assessing children's sedation levels. total And use the error backpropagation algorithm to L total Minimize the direction of adjustment of neural network model parameters for assessing children's sedation levels;

[0036] When the system uses a neural network model to assess pediatric sedation levels online, the final assessment of the patient's sedation level is determined by the sedation level with the highest probability in Y.

[0037] A further improvement of the technical solution of the present invention is that: in S401, the multi-scale time-frequency feature extraction component is composed of a time-domain multi-scale feature extractor and a frequency-domain multi-scale feature extractor;

[0038] S401 specifically includes the following steps:

[0039] S401a. Obtain the latest preprocessed EEG x1 with a time length of t and a sampling rate of fs from the storage module, and perform a fast Fourier transform on x1 to obtain its frequency domain representation x2.

[0040] S401b, Adjust the shapes of x1 and x2 to obtain the time-domain input X1 of the multi-scale time-frequency feature extraction component. and frequency domain input

[0041] S401c, Input X1 into the temporal multi-scale feature extractor to obtain the temporal local feature T local and temporal global features T global ;

[0042] S401d: Input X2 into the frequency domain multi-scale feature extractor to obtain the local frequency domain feature F. local and frequency domain global features T global ;

[0043] The time-domain multi-scale feature extractor and the frequency-domain multi-scale feature extractor have the same structure, both consisting of local feature extraction branches and global feature extraction branches;

[0044] The local feature extraction branch consists of two convolutional layers with 1×3 kernels, while the global feature extraction branch consists of two convolutional layers with 1×11 kernels.

[0045] A further improvement to the technical solution of the present invention is that, in S402, the feature encoding component consists of a local multi-head attention block and MHA. local Global Multi-Head Attention Block (MHA) globalIt consists of a multi-scale attention block (MSA) and a feature reduction block (Down).

[0046] S402 specifically includes the following steps:

[0047] S402a, T local and F local Enter MHA local T local As the key vector and value vector of this multi-head attention block, F local As the query vector for this multi-head attention block, the output FT is obtained. local , FT local With F local By concatenating along the first dimension, we obtain the time-frequency local feature O. local ;

[0048] S402b, F global and T global Enter MHA global F global As the key vector and value vector of this multi-head attention block, T global As the query vector for this multi-head attention block, the output FT is obtained. global , FT global With T global By concatenating along the first dimension, we obtain the time-frequency global feature O. global ;

[0049] S402c, O local and O global O is obtained by concatenating the second dimension. O then passes through a batch normalization layer, a ReLU activation function, a multi-scale attention block (MSA), and a feature reduction block (Down) to obtain the sample feature representation D.

[0050] A further improvement of the technical solution of the present invention is that the multi-scale attention block (MSA) consists of four branches;

[0051] The first branch consists of a convolutional layer with a 1×1 kernel, a batch normalization layer, a ReLU activation function, and an adaptive weight.

[0052] The second branch consists of a convolutional layer with a 3×3 kernel, a batch normalization layer, a ReLU activation function, and an adaptive weight.

[0053] The third branch consists of a convolutional layer with a 3×3 kernel, a batch normalization layer, a ReLU activation function, a skip connection from the input of this branch to the ReLU activation function, and an adaptive weight.

[0054] The fourth branch consists of an adaptive pooling, a convolutional layer with a 1×1 kernel, a ReLU activation function, another convolutional layer with a 1×1 kernel, a Sigmoid activation function, a skip connection from the input of this branch to the Sigmoid activation function, and an adaptive weight.

[0055] The outputs of the four branches are weighted and summed based on four adaptive weights, and finally the output of the multi-scale attention block is generated by passing the sigmoid activation function.

[0056] A further improvement of the technical solution of the present invention is that: the feature dimensionality reduction block Down is composed of two serially connected feature dimensionality reduction units, and the feature dimensionality reduction unit is composed of two dimensionality reduction branches;

[0057] The first dimension reduction branch consists of a convolutional layer with a 1×1 kernel and a batch normalization layer.

[0058] The second dimensionality reduction branch consists of a convolutional layer with a 3×3 kernel, a batch normalization layer, a ReLU activation function, another convolutional layer with a 3×3 kernel, and a batch normalization layer.

[0059] The outputs of the two dimensionality reduction branches are summed, and finally the ReLU activation function is used to generate the final output of the feature dimensionality reduction block unit.

[0060] A further improvement of the technical solution of the present invention is that: in S403, the comparison and classification component is composed of a local feature projection head, a global feature projection head, and a classifier;

[0061] The local feature projection head and the global feature projection head have the same structure, both consisting of an adaptive pooling layer, a flattening layer, a linear connection layer, a batch normalization layer, a ReLU activation function, and a linear connection layer in sequence.

[0062] The classifier consists of an adaptive pooling layer, a flattening layer, a linear connection layer, a ReLU activation function, a Dropout layer, and a linear connection layer, in sequence.

[0063] S403 specifically includes the following steps:

[0064] S403a, O local Input the local feature projection head to obtain the time-frequency local feature projection Z. local ;

[0065] S403b, O global Input a local feature projection head to obtain the time-frequency global feature projection Z. global ;

[0066] S403c: Input D into the classifier to obtain the probability output Y of the pediatric sedation level assessment neural network model regarding the patient's sedation level.

[0067] A further improvement to the technical solution of the present invention lies in: during the offline training phase of the neural network model for assessing the child's sedation level, the overall loss L of the neural network model for assessing the child's sedation level is... total Classification loss L CE Supervised local contrast loss L local Compared with supervised global comparison loss L global The proportions of these three losses in the total loss are controlled by the designed balance coefficient β.

[0068] The technological advancements achieved by this invention due to the adoption of the above technical solutions are as follows:

[0069] 1. This invention addresses the current clinical dilemma of lacking a reliable monitoring system specifically for esketamine anesthesia in children. It innovatively develops a deep learning-based sedation level monitoring system for esketamine anesthesia in children. This system can effectively assist anesthesiologists in judging the depth of sedation in children, providing a basis for individualized medication decisions, thereby significantly reducing the risk of excessive anesthesia or intraoperative awareness.

[0070] 2. Thanks to its unique modular design, the sedation level monitoring model developed in this invention outperforms existing models in performance: On a dataset of 53 children (3-18 years old) under esketamine anesthesia obtained from Peking University First Hospital, the accuracy reached 88.72%; on the same dataset, the general-purpose model ResNet18 achieved only 71.61% accuracy, and the general-purpose model EEGNet only 68.15% accuracy. Furthermore, ablation experiments showed that removing the multi-scale time-frequency feature extraction component reduced the performance of the sedation level monitoring model to 73.24%; removing the multi-scale attention block MSA reduced the performance to 75.02%; removing the feature dimensionality reduction block Down reduced the performance to 75.54%; and removing the feature projection head and the designed loss function reduced the performance to 74.32%. These experimental results demonstrate the technical effectiveness and advancement of the components of the model designed in this invention.

[0071] 3. The neural network model for assessing pediatric sedation levels in the monitoring system of this invention is trained based on real pediatric clinical anesthesia data obtained from hospitals. It has the ability to accurately distinguish five sedation levels, and this grading ability is sufficient to provide reliable guidance for anesthesiologists in drug administration.

[0072] 4. The monitoring system of this invention is suitable for children and adolescents aged 3 to 18, covering the developmental stages from infancy to puberty, and has excellent clinical applicability. Attached Figure Description

[0073] Figure 1 This is a schematic diagram of the deep learning-based sedation level monitoring system for pediatric esketamine anesthesia, based on the present invention.

[0074] Figure 2 This is a flowchart illustrating the operation of the monitoring system of the present invention;

[0075] Figure 3 This is a schematic diagram of the preprocessing module of the monitoring system of the present invention performing noise reduction processing on the collected electroencephalogram (EEG) signals;

[0076] Figure 4 This is a schematic diagram of the process by which the monitoring system of the present invention judges the sedation level of a patient based on the input EEG. Detailed Implementation

[0077] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments:

[0078] like Figure 1 As shown, a deep learning-based sedation level monitoring system for pediatric esketamine anesthesia includes a power module, a sensor module, a preprocessing module, a storage module, a central processing module, a communication module, and a host computer module.

[0079] The power module is used to provide power to the system;

[0080] The sensor module is used to collect the patient's electroencephalogram (EEG) signals in real time.

[0081] The preprocessing module is used to reduce noise in the acquired EEG signals;

[0082] The storage module is used to save the static sedation level assessment neural network model that has been trained offline, as well as log data, monitoring data and event tagging data generated during system operation;

[0083] The central processing module is used to call the neural network model for assessing pediatric sedation levels stored in the storage module. The model determines the patient's sedation level based on the input EEG. The sedation level refers to five predefined sedation levels during the model's offline training phase.

[0084] Awake: The patient is fully conscious before any medication has been administered;

[0085] Deep anesthesia: A state in which there is no response to strong stimuli and surgery can be performed;

[0086] Moderate anesthesia: a state in which there is a slight response to strong stimuli (such as body movement), and surgery can be performed;

[0087] Superficial anesthesia: The patient's consciousness is about to return, a transitional state between moderate anesthesia and the return of consciousness;

[0088] Recovery of consciousness: The patient responds to verbal commands and has regained consciousness.

[0089] The communication module is used to send the model's judgment result on the patient's sedation level to the host computer module;

[0090] The host computer module is used to display the patient's sedation level.

[0091] like Figure 2 As shown, a deep learning-based method for monitoring sedation levels in children undergoing esketamine anesthesia includes the following steps:

[0092] S1. Power on the power module, and the system begins to run;

[0093] S2. Place the sensor module on the patient's forehead and use the sensor module to collect single-channel EEG signals from the patient's forehead. The sampling rate of the EEG signals must be greater than 120Hz.

[0094] In this embodiment, the dataset used is the EEG data of 53 children under anesthesia obtained from the Children's Center of Peking University First Hospital, with a sampling rate of 250Hz, and the sedative drug used is esketamine. During the EEG collection process, the anesthesiologist will observe the patient's vital signs (such as blood pressure, heart rate, etc.) and the patient's response to stimuli (such as tapping, calling name, etc.) to determine the patient's sedation level and record it (such as determination time, determination result, etc.).

[0095] S3. The preprocessing module is used to perform noise reduction processing on the acquired single-channel EEG signals, such as... Figure 3 As shown, step S3 specifically includes the following steps:

[0096] S301. The EEG signal is amplified by a preamplifier circuit, with an amplification factor of 1000 times;

[0097] S302. Low-frequency noise in EEG signals is filtered out based on a hardware high-pass filter, with a cutoff frequency of 0.1Hz.

[0098] S303. High-frequency components in EEG signals are filtered out based on a hardware low-pass filter with a cutoff frequency of 60Hz.

[0099] S304. Converting EEG signals into digital signals based on analog-to-digital conversion circuits;

[0100] S305. Power frequency noise in EEG signals is filtered out based on a digital notch filter with a cutoff frequency of 49-51Hz.

[0101] S306. Based on the downsampling algorithm, the EEG signal is downsampled to a specified frequency;

[0102] In this exemplary embodiment, the specified frequency for downsampling is 100Hz;

[0103] S307. Save the raw EEG and the preprocessed EEG to the storage module.

[0104] S4. The central processing module calls the neural network model for assessing the sedation level of children stored in the storage module. The model judges the patient's sedation level based on the input EEG. The neural network model for assessing the sedation level of children includes a multi-scale time-frequency feature extraction component, a feature encoding component, and a contrast classification component.

[0105] like Figure 4 As shown, the neural network model for assessing sedation levels in children determines the patient's sedation level based on the input EEG, specifically including the following steps:

[0106] S401. Based on the multi-scale time-frequency feature extraction component, extract time-domain and frequency-domain features of different scales from the EEG samples of the input model to obtain the local time-domain features T. local Temporal global features T global Frequency domain local features F local and frequency domain global features T global ;

[0107] The multi-scale time-frequency feature extraction component consists of a time-domain multi-scale feature extractor and a frequency-domain multi-scale feature extractor. The time-domain multi-scale feature extractor and the frequency-domain multi-scale feature extractor have the same structure, both consisting of a local feature extraction branch and a global feature extraction branch. The local feature extraction branch consists of two convolutional layers with 1×3 kernels, and the global feature extraction branch consists of two convolutional layers with 1×11 kernels.

[0108] S401 specifically includes the following steps:

[0109] S401a. Obtain the latest preprocessed EEG x1 with a time length of t and a sampling rate of fs from the storage module, and perform a fast Fourier transform on x1 to obtain its frequency domain representation x2.

[0110] In this exemplary embodiment, the time duration is 10 seconds, and the sampling rate of the preprocessed EEG x1 is 100 Hz;

[0111] S401b, Adjust the shapes of x1 and x2 to obtain the time-domain input X1 of the multi-scale time-frequency feature extraction component. and frequency domain input

[0112] S401c, Input X1 into the temporal multi-scale feature extractor to obtain the temporal local feature T local and temporal global features T global The formula is shown below:

[0113]

[0114] Where σ represents the ReLU activation function, Conv represents the convolution operation of the convolutional layer, and k represents the kernel size of the convolutional layer;

[0115] S401d: Input X2 into the frequency domain multi-scale feature extractor to obtain the local frequency domain feature F. local and frequency domain global features T global The formula is shown below:

[0116]

[0117] S402, Based on the feature encoding component, from the input T local T global F local and T global Extracting local time-frequency features O local Time-frequency global features O global The sample feature representation is D; where the feature encoding component consists of a local multi-head attention block and MHA. local Global Multi-Head Attention Block (MHA) global It consists of a multi-scale attention block (MSA) and a feature reduction block (Down).

[0118] The multi-scale attention block (MSA) consists of four branches:

[0119] The first branch consists of a convolutional layer with a 1×1 kernel, a batch normalization layer, a ReLU activation function, and an adaptive weight.

[0120] The second branch consists of a convolutional layer with a 3×3 kernel, a batch normalization layer, a ReLU activation function, and an adaptive weight.

[0121] The third branch consists of a convolutional layer with a 3×3 kernel, a batch normalization layer, a ReLU activation function, a skip connection from the input of this branch to the ReLU activation function, and an adaptive weight.

[0122] The fourth branch consists of an adaptive pooling, a convolutional layer with a 1×1 kernel, a ReLU activation function, another convolutional layer with a 1×1 kernel, a Sigmoid activation function, a skip connection from the input of this branch to the Sigmoid activation function, and an adaptive weight.

[0123] The outputs of the four branches are weighted and summed based on four adaptive weights, and finally the output of the multi-scale attention block is generated by passing through the Sigmoid activation function.

[0124] The feature reduction block Down consists of two cascaded feature reduction units, and each feature reduction unit consists of two branches:

[0125] The first dimension reduction branch consists of a convolutional layer with a 1×1 kernel and a batch normalization layer.

[0126] The second dimensionality reduction branch consists of a convolutional layer with a 3×3 kernel, a batch normalization layer, a ReLU activation function, another convolutional layer with a 3×3 kernel, and a batch normalization layer.

[0127] The outputs of the two branches are summed, and finally the ReLU activation function is used to generate the final output of the feature reduction block unit.

[0128] S402 specifically includes the following steps:

[0129] S402a, T local and F local Enter MHA local T local As the key vector and value vector of this multi-head attention block, F local As the query vector for this multi-head attention block, the output FT is obtained. local , FT local With F local By concatenating along the first dimension, we obtain the time-frequency local feature O. local The formula is shown below:

[0130]

[0131] Where Q represents the query vector, K represents the key vector, V represents the value vector, and Concat represents the concatenation operation;

[0132] S402b, F global and T global Enter MHA global F global As the key vector and value vector of this multi-head attention block, T global As the query vector for this multi-head attention block, the output FT is obtained. global , FT global With T global By concatenating along the first dimension, we obtain the time-frequency global feature O. global The formula is shown below:

[0133]

[0134] S402c, O local and O global O is obtained by concatenating the elements in the second dimension, as shown in the formula below:

[0135]

[0136] Where σ represents the ReLU activation function and Concat represents the concatenation operation;

[0137] Afterwards, O passes through a batch normalization layer and the ReLU activation function in sequence, as shown in the following formula:

[0138]

[0139] Where σ represents the ReLU activation function and BN represents the batch normalization layer;

[0140] Afterwards, O passes through a multi-scale attention block (MSA), as shown in the following formula:

[0141]

[0142] Where σ represents the ReLU activation function, Branch i represents the i-th branch of the multi-scale attention block, ⊙ represents the dot product operation, and α represents the adaptive weights of the four branches.

[0143] Finally, O is processed by the feature reduction block Down to obtain the sample feature representation D, as shown in the following formula:

[0144]

[0145] S403, Based on the contrastive classification component, from the input O local O global Obtain the time-frequency local feature projection Z from D. local Time-frequency global feature projection Z global And the probability output Y, where Y contains the predicted probabilities of 5 different sedation levels;

[0146] The local feature projection head and the global feature projection head have the same structure, both consisting of an adaptive pooling layer, a flattening layer, a linear connection layer, a batch normalization layer, a ReLU activation function, and a linear connection layer in sequence.

[0147] The classifier consists of an adaptive pooling layer, a flattening layer, a linear connection layer, a ReLU activation function, a Dropout layer, and a linear connection layer, in sequence.

[0148] S403 specifically includes the following steps:

[0149] S403a, O local Input the local feature projection head to obtain the time-frequency local feature projection Z. local ;

[0150] S403b, O global Input the global feature projection head to obtain the time-frequency global feature projection Z. global ;

[0151] S403c: Input D into the classifier to obtain the model's probability output Y regarding the patient's sedation level.

[0152] During the offline training phase of the neural network model for assessing sedation levels in children, based on Z... local Z global The overall loss L of the Y calculation model total And use the error backpropagation algorithm to L total Minimize the direction to adjust the model parameters; where the overall loss of the model is L. total Classification loss L CE Supervised local contrast loss L local Compared with supervised global comparison loss L global composition;

[0153] The classification loss L CE The formula is shown below:

[0154]

[0155] Where N represents the number of samples in a batch during training, y i,c p represents the true sedation level of the sample as expressed in one-hot code form. i,c The predicted sedation level of the sample is represented in the form of probability output, where i represents the i-th sample in the batch;

[0156] The supervised local contrast loss L local The formula is shown below:

[0157]

[0158] Where N represents the number of samples in a batch during training, P(i) represents the set of samples in the batch that have the same calming level label (positive sample) as anchor sample i, τ is a parameter controlling the sharpness of the similarity distribution, and z i ·z p Represents anchor sample z i With positive sample z p The cosine similarity between them, and z i ·z p Represents anchor sample z iCosine similarity with all other samples in the batch;

[0159] Supervised global comparison loss L global The calculation method and the supervised local comparison loss L local The only difference is that the supervised local contrast loss is calculated based on the projection of time-frequency local features. local The supervised global contrastive loss is calculated based on the projection of time-frequency global features. local ;

[0160] The proportion of these three types of losses in the total loss is controlled by the design balance coefficient β, as shown in the following formula:

[0161]

[0162] During offline training, β gradually increases during the iteration process, and the model that has completed offline training is saved in the storage module; in this exemplary embodiment, β gradually increases from 0.1 to 0.9 during the iteration process.

[0163] When the system uses a neural network model to assess pediatric sedation levels online, the final assessment of the patient's sedation level is determined by the sedation level with the highest probability in Y.

[0164] S5. Use the communication module to send the model's judgment result on the patient's sedation level to the host computer module.

[0165] S6. The host computer module displays the patient's sedation level for real-time monitoring. The host computer module also has historical data display and event annotation functions.

[0166] When the system operates continuously for extended periods, the model outputs sedation level assessment results in real time at 1-second intervals and displays them on the host computer. The host computer can display the sedation level based on historical data; the changes in EEG and other data; EEG data, sedation level monitoring data, system work logs (such as start time, termination time, and error reports), and the doctor's annotations of events during surgery (such as the start of drug administration, the start of surgery, and the end of surgery) can all be sent to the central processing module via the communication module and saved in encrypted CSV format to the storage module.

[0167] When tested on a dataset of 53 children (aged 3-18) under esketamine anesthesia obtained from the hospital, the system achieved an accuracy rate of 88.72%; among them, the accuracy rate was 83.11% in the 3-6 year old group, 91.59% in the 6-12 year old group, and 91.07% in the 12-118 year old group, which meets clinical needs.

[0168] In summary, this invention can provide real-time sedation level monitoring services for children aged 3-18 years under esketamine anesthesia.

Claims

1. A sedation level monitoring system for children under esketamine anesthesia, characterized in that: It includes a power module, a sensor module, a preprocessing module, a storage module, a central processing module, a communication module, and a host computer module; The power module is used to provide power to the system. The sensor module is used to collect the patient's electroencephalogram (EEG) signals in real time. The preprocessing module is used to perform noise reduction processing on the acquired EEG signals. The storage module is used to store the static sedation level evaluation neural network model that has completed offline training, as well as the log data, monitoring data and event tagging data generated during system operation. The central processing module is used to call the neural network model for assessing the sedation level of children stored in the storage module. The neural network model for assessing the sedation level of children judges the sedation level of the patient based on the input EEG. The communication module is used to send the judgment results of the pediatric sedation level assessment neural network model on the patient's sedation level to the host computer module. The host computer module is used to display the patient's sedation level.

2. A method for monitoring sedation levels in children under esketamine anesthesia, characterized in that: The application of the sedation level monitoring system for pediatric esketamine anesthesia as described in claim 1 includes the following steps: S1. Power on the power module, and the system begins to run; S2. Place the sensor module on the patient's forehead and use the sensor module to collect single-channel EEG signals from the patient's forehead. S3. The preprocessing module is used to reduce noise in the acquired single-channel EEG signals. S4. The central processing module calls the neural network model for assessing the sedation level of children stored in the storage module. The neural network model for assessing the sedation level of children judges the sedation level of patients based on the input EEG, including a multi-scale time-frequency feature extraction component, a feature encoding component, and a contrast classification component. S5. Use the communication module to send the judgment results of the sedation level of the patient by the neural network model for assessing the sedation level of the child to the host computer module. S6. Use the host computer module to display the patient's sedation level for real-time monitoring.

3. The method for monitoring sedation levels in children under esketamine anesthesia according to claim 2, characterized in that: S3 specifically includes the following steps: S301. The EEG signal is amplified by a preamplifier circuit, with an amplification factor of 1000 times; S302. Low-frequency noise in EEG signals is filtered out based on a hardware high-pass filter, with a cutoff frequency of 0.1Hz. S303. High-frequency components in EEG signals are filtered out based on a hardware low-pass filter with a cutoff frequency of 60Hz. S304. Converting EEG signals into digital signals based on analog-to-digital conversion circuits; S305. Power frequency noise in EEG signals is filtered out based on a digital notch filter with a cutoff frequency of 49-51Hz. S306. Based on the downsampling algorithm, the EEG signal is downsampled to a specified frequency; The specified downsampling frequency is 100Hz; S307. Save the raw EEG and the preprocessed EEG to the storage module.

4. The method for monitoring sedation levels in children under esketamine anesthesia according to claim 2, characterized in that: In S4, the neural network model for assessing pediatric sedation levels determines the patient's sedation level based on the input EEG, specifically including the following steps: S401. Based on a multi-scale time-frequency feature extraction component, extract time-domain and frequency-domain features at different scales from the EEG samples input to the neural network model for assessing pediatric sedation levels, and obtain the local time-domain feature T. local Temporal global features T global Frequency domain local features F local and frequency domain global features T global ; S402, Based on the feature encoding component, from the input T local T global F local and T global Extracting local time-frequency features O local Time-frequency global features O global and sample feature representation D; S403, Based on the contrastive classification component, from the input O local O global Obtain the time-frequency local feature projection Z from D. local Time-frequency global feature projection Z global And the probability output Y, where Y contains the predicted probabilities of 5 different sedation levels; During the offline training phase of the neural network model for assessing sedation levels in children, based on Z... local Z global And Y calculate the total loss L of the neural network model for assessing children's sedation levels. total And use the error backpropagation algorithm to L total Minimize the direction of adjustment of neural network model parameters for assessing sedation levels in children; When the system uses a neural network model to assess pediatric sedation levels online, the final assessment of the patient's sedation level is determined by the sedation level with the highest probability in Y.

5. The method for monitoring sedation levels in children under esketamine anesthesia according to claim 4, characterized in that: In S401, the multi-scale time-frequency feature extraction component consists of a time-domain multi-scale feature extractor and a frequency-domain multi-scale feature extractor; S401 specifically includes the following steps: S401a. Obtain the latest preprocessed EEG x1 with a time length of t and a sampling rate of fs from the storage module, and perform a fast Fourier transform on x1 to obtain its frequency domain representation x2. S401b: Adjust the shapes of x1 and x2 to obtain the time-domain input of the multi-scale time-frequency feature extraction component. and frequency domain input S401c, Input X1 into the temporal multi-scale feature extractor to obtain the temporal local feature T local and temporal global features T global ; S401d: Input X2 into the frequency domain multi-scale feature extractor to obtain the local frequency domain feature F. local and frequency domain global features T global ; The time-domain multi-scale feature extractor and the frequency-domain multi-scale feature extractor have the same structure, both consisting of local feature extraction branches and global feature extraction branches; The local feature extraction branch consists of two convolutional layers with 1×3 kernels, while the global feature extraction branch consists of two convolutional layers with 1×11 kernels.

6. The method for monitoring sedation levels in children under esketamine anesthesia according to claim 4, characterized in that: In S402, the feature encoding component consists of a local multi-head attention block and MHA. local Global Multi-Head Attention Block (MHA) global It consists of a multi-scale attention block (MSA) and a feature reduction block (Down). S402 specifically includes the following steps: S402a, T local and F local Enter MHA local T local As the key vector and value vector of this multi-head attention block, F local As the query vector for this multi-head attention block, the output FT is obtained. local , FT local With F local By concatenating along the first dimension, we obtain the time-frequency local feature O. local ; S402b, F global and T global Enter MHA global F global As the key vector and value vector of this multi-head attention block, T global As the query vector for this multi-head attention block, the output FT is obtained. global , FT global With T global By concatenating along the first dimension, we obtain the time-frequency global feature O. global ; S402c, O local and O global O is obtained by concatenating the second dimension. O then passes through a batch normalization layer, a ReLU activation function, a multi-scale attention block (MSA), and a feature reduction block (Down) to obtain the sample feature representation D.

7. The method for monitoring sedation levels in children under esketamine anesthesia according to claim 6, characterized in that: The multi-scale attention block (MSA) consists of four branches; The first branch consists of a convolutional layer with a 1×1 kernel, a batch normalization layer, a ReLU activation function, and an adaptive weight. The second branch consists of a convolutional layer with a 3×3 kernel, a batch normalization layer, a ReLU activation function, and an adaptive weight. The third branch consists of a convolutional layer with a 3×3 kernel, a batch normalization layer, a ReLU activation function, a skip connection from the input of this branch to the ReLU activation function, and an adaptive weight. The fourth branch consists of an adaptive pooling, a convolutional layer with a 1×1 kernel, a ReLU activation function, another convolutional layer with a 1×1 kernel, a Sigmoid activation function, a skip connection from the input of this branch to the Sigmoid activation function, and an adaptive weight. The outputs of the four branches are weighted and summed based on four adaptive weights, and finally the output of the multi-scale attention block is generated by passing the sigmoid activation function.

8. The method for monitoring sedation levels in children under esketamine anesthesia according to claim 6, characterized in that: The feature reduction block Down consists of two cascaded feature reduction units, and each feature reduction unit consists of two reduction branches. The first dimension reduction branch consists of a convolutional layer with a 1×1 kernel and a batch normalization layer. The second dimensionality reduction branch consists of a convolutional layer with a 3×3 kernel, a batch normalization layer, a ReLU activation function, another convolutional layer with a 3×3 kernel, and a batch normalization layer. The outputs of the two dimensionality reduction branches are summed, and finally the ReLU activation function is used to generate the final output of the feature dimensionality reduction block unit.

9. The method for monitoring sedation levels in children under esketamine anesthesia according to claim 4, characterized in that: In S403, the comparison and classification component consists of a local feature projection head, a global feature projection head, and a classifier; The local feature projection head and the global feature projection head have the same structure, both consisting of an adaptive pooling layer, a flattening layer, a linear connection layer, a batch normalization layer, a ReLU activation function, and a linear connection layer in sequence. The classifier consists of an adaptive pooling layer, a flattening layer, a linear connection layer, a ReLU activation function, a Dropout layer, and a linear connection layer, in sequence. S403 specifically includes the following steps: S403a, O local Input the local feature projection head to obtain the time-frequency local feature projection Z. local ; S403b, O global Input a local feature projection head to obtain the time-frequency global feature projection Z. global ; S403c: Input D into the classifier to obtain the probability output Y of the pediatric sedation level assessment neural network model regarding the patient's sedation level.

10. The method for monitoring sedation levels in children under esketamine anesthesia according to claim 4, characterized in that: During the offline training phase of the neural network model for assessing sedation levels in children, the overall loss L of the neural network model for assessing sedation levels in children is... total Classification loss L CE Supervised local contrast loss L local Compared with supervised global comparison loss L global The proportion of these three types of losses in the total loss is controlled by the designed balance coefficient β.

Images