Artificial intelligence-based anesthesia depth classification method and system

Abstract

The embodiment of the invention relates to the technical field of anesthesia depth classification, and particularly discloses an anesthesia depth classification method and system based on artificial intelligence. According to the embodiment of the invention, the method comprises the steps: selecting a target basic model through obtaining a plurality of standard feature data, carrying out the model training, and generating a target classification model; evaluating and optimizing the target classification model to generate an optimized classification model; acquiring physiological parameter data, and performing data preprocessing on the physiological parameter data to generate standard parameter data; performing multi-feature extraction to obtain parameter feature data; and importing the parameter characteristic data into the optimization classification model, carrying out anesthesia depth classification, and obtaining and displaying a depth classification result. According to the method, the optimized classification model can be constructed, the physiological parameter data can be collected, data preprocessing and multi-feature extraction can be carried out, the data can be imported into the optimized classification model, the depth classification result can be obtained and displayed, automatic and accurate judgment of the anesthesia depth can be realized, judgment errors can be effectively avoided, and smooth operation can be ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of anesthesia depth classification, and in particular relates to an anesthesia depth classification method and system based on artificial intelligence. Background Art

[0002] The depth of anesthesia is usually classified according to the patient's consciousness and reaction level during anesthesia, and is mainly divided into the following categories: light anesthesia, moderate anesthesia, and deep anesthesia. The depth of anesthesia is mainly determined by clinical signs and electrophysiological methods. Clinical signs include blood pressure, heart rate, airway resistance, tearing, body movement, and salivation, while electrophysiology includes bispectral index (BIS) monitoring and anesthesia depth monitor.

[0003] Since the depth of anesthesia varies with the patient's condition, in the prior art, doctors mainly judge the depth of anesthesia based on clinical signs and electrophysiological data. Inexperienced doctors are likely to make misjudgments, affecting the smooth progress of the operation. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide an artificial intelligence-based anesthesia depth classification method and system, aiming to solve the problems raised in the background technology.

[0005] To achieve the above objectives, the embodiments of the present invention provide the following technical solutions: An artificial intelligence-based anesthesia depth classification method, the method specifically comprising the following steps: Based on multiple physiological parameters for anesthesia depth classification, multiple standard feature data are obtained, and a target basic model is selected from multiple preset models to be selected, and model training is performed to generate a target classification model; Evaluating and optimizing the target classification model to generate an optimized classification model; According to the plurality of physiological parameters, data collection is performed to obtain physiological parameter data, and data preprocessing is performed on the physiological parameter data to generate standard parameter data; Performing multiple feature extraction on the standard parameter data to obtain parameter feature data; The parameter characteristic data is imported into the optimized classification model to perform anesthesia depth classification analysis, and the depth classification result is obtained and displayed.

[0006] As a further limitation of the technical solution of the embodiment of the present invention, the multiple physiological parameters based on the anesthesia depth classification, obtaining multiple standard feature data, and selecting a target basic model from multiple preset models to be selected, performing model training, and generating a target classification model specifically include the following steps: Based on multiple physiological parameters of anesthesia depth classification, multiple standard feature data are obtained; According to a preset data division ratio, the plurality of standard feature data are divided into a training set and a test set; Select a target basic model from multiple preset models to be selected; The target basic model is trained and tested by using the training set and the test set to generate a target classification model.

[0007] As a further limitation of the technical solution of the embodiment of the present invention, the multiple models to be selected include: a convolutional neural network model, a long short-term memory model, a random forest model and a support vector machine model.

[0008] As a further limitation of the technical solution of the embodiment of the present invention, the evaluating and optimizing the target classification model to generate the optimized classification model specifically includes the following steps: Evaluate the accuracy, precision, recall rate and F1 score of the target classification model, and record first evaluation data; Based on the ROC curve and the AUC value, the target classification model is evaluated and second evaluation data is recorded; Performing hyperparameter tuning on the target classification model and recording hyperparameter tuning data; The target classification model is optimized based on the first evaluation data, the second evaluation data and the hyperparameter tuning data to generate an optimized classification model.

[0009] As a further limitation of the technical solution of the embodiment of the present invention, the data collection according to the plurality of physiological parameters, the acquisition of physiological parameter data, and the data preprocessing of the physiological parameter data to generate standard parameter data specifically include the following steps: According to the plurality of physiological parameters, data collection is performed to obtain physiological parameter data; Performing denoising processing on the physiological parameter data to generate denoised parameter data; The denoising parameter data is standardized to generate standard parameter data.

[0010] As a further limitation of the technical solution of the embodiment of the present invention, the performing of multiple feature extraction on the standard parameter data to obtain parameter feature data specifically comprises the following steps: Performing time domain feature extraction on the standard parameter data to obtain time domain feature data; Performing frequency domain feature extraction on the standard parameter data to obtain frequency domain feature data; Extracting time-frequency features from the standard parameter data to obtain time-frequency feature data; Performing nonlinear feature extraction on the standard parameter data to obtain nonlinear feature data; The time domain feature data, the frequency domain feature data, the time-frequency feature data and the nonlinear feature data are integrated to generate parameter feature data.

[0011] An anesthesia depth classification system based on artificial intelligence, the system comprises a model selection training unit, a model evaluation optimization unit, a data collection and processing unit, a multiple feature extraction unit and an anesthesia depth classification unit, wherein: A model selection training unit is used to obtain multiple standard feature data based on multiple physiological parameters of anesthesia depth classification, and select a target basic model from multiple preset models to be selected, perform model training, and generate a target classification model; A model evaluation and optimization unit, used to evaluate and optimize the target classification model to generate an optimized classification model; A data collection and processing unit, used to collect data according to the plurality of physiological parameters, obtain physiological parameter data, and perform data preprocessing on the physiological parameter data to generate standard parameter data; A multiple feature extraction unit, used for performing multiple feature extraction on the standard parameter data to obtain parameter feature data; The anesthesia depth classification unit is used to import the parameter characteristic data into the optimized classification model, perform anesthesia depth classification analysis, and obtain and display depth classification results.

[0012] As a further limitation of the technical solution of the embodiment of the present invention, the model selection training unit specifically includes: A data acquisition module, used for acquiring a plurality of standard characteristic data based on a plurality of physiological parameters classified according to the depth of anesthesia; A data partitioning module, used to divide the plurality of standard feature data into a training set and a test set according to a preset data partitioning ratio; A model selection module, used to select a target basic model from a plurality of preset models to be selected; The training and testing module is used to train and test the target basic model through the training set and the test set to generate a target classification model.

[0013] As a further limitation of the technical solution of the embodiment of the present invention, the model evaluation optimization unit specifically includes: A first evaluation module is used to evaluate the accuracy, precision, recall rate and F1 score of the target classification model and record first evaluation data; A second evaluation module is used to evaluate the target classification model based on the ROC curve and the AUC value, and record second evaluation data; A hyperparameter tuning module, used to perform hyperparameter tuning on the target classification model and record hyperparameter tuning data; A model optimization module is used to optimize the target classification model by integrating the first evaluation data, the second evaluation data and the hyperparameter tuning data to generate an optimized classification model.

[0014] As a further limitation of the technical solution of the embodiment of the present invention, the multiple feature extraction unit specifically includes: A time domain feature extraction module, used to extract time domain features from the standard parameter data to obtain time domain feature data; A frequency domain feature extraction module, used to extract frequency domain features from the standard parameter data to obtain frequency domain feature data; A time-frequency feature extraction module, used to extract time-frequency features from the standard parameter data to obtain time-frequency feature data; A nonlinear feature extraction module, used for performing nonlinear feature extraction on the standard parameter data to obtain nonlinear feature data; The feature data generating module is used to generate parameter feature data by integrating the time domain feature data, the frequency domain feature data, the time-frequency feature data and the nonlinear feature data.

[0015] Compared with the prior art, the present invention has the following beneficial effects: The embodiment of the present invention obtains multiple standard feature data, selects a target basic model, performs model training, and generates a target classification model; evaluates and optimizes the target classification model to generate an optimized classification model; obtains physiological parameter data, performs data preprocessing on the physiological parameter data, and generates standard parameter data; performs multiple feature extraction to obtain parameter feature data; imports the parameter feature data into the optimized classification model, performs anesthesia depth classification, and obtains and displays the depth classification results. It is possible to construct an optimized classification model, collect physiological parameter data, perform data preprocessing and multiple feature extraction, import it into the optimized classification model, obtain and display the depth classification results, and realize automatic and accurate judgment of anesthesia depth, effectively avoid misjudgment, and ensure the smooth progress of the operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention.

[0017] Figure 1 A flow chart of a method provided by an embodiment of the present invention is shown.

[0018] Figure 2 A flow chart of generating a target classification model in the method provided in an embodiment of the present invention is shown.

[0019] Figure 3A flow chart of generating an optimized classification model in the method provided in an embodiment of the present invention is shown.

[0020] Figure 4 A flow chart of generating standard parameter data in the method provided in an embodiment of the present invention is shown.

[0021] Figure 5 A flow chart of obtaining parameter characteristic data in the method provided in an embodiment of the present invention is shown.

[0022] Figure 6 The application architecture diagram of the system provided by the embodiment of the present invention is shown.

[0023] Figure 7 A structural block diagram of a model selection training unit in a system provided by an embodiment of the present invention is shown.

[0024] Figure 8 A structural block diagram of a model evaluation and optimization unit in a system provided by an embodiment of the present invention is shown.

[0025] Fig. 9 A structural block diagram of a multiple feature extraction unit in a system provided by an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0027] It is understandable that since the depth of anesthesia varies with the patient's condition, and in the prior art, doctors mainly judge the depth of anesthesia based on clinical signs and electrophysiological data, inexperienced doctors are likely to make misjudgments and affect the smooth progress of the operation.

[0028] To solve the above problems, the embodiment of the present invention obtains multiple standard feature data based on multiple physiological parameters of anesthesia depth classification, and selects a target basic model from multiple preset models to be selected, performs model training, and generates a target classification model; evaluates and optimizes the target classification model to generate an optimized classification model; collects data according to multiple physiological parameters, obtains physiological parameter data, and performs data preprocessing on the physiological parameter data to generate standard parameter data; performs multiple feature extraction on the standard parameter data to obtain parameter feature data; imports the parameter feature data into the optimized classification model, performs anesthesia depth classification analysis, and obtains and displays the depth classification results. It is possible to construct an optimized classification model, collect physiological parameter data, perform data preprocessing and multiple feature extraction, import it into the optimized classification model, obtain and display the depth classification results, and realize automatic and accurate judgment of anesthesia depth, effectively avoid misjudgment, and ensure the smooth progress of the operation.

[0029] Figure 1 A flow chart of a method provided by an embodiment of the present invention is shown.

[0030] Specifically, a method for classifying anesthesia depth based on artificial intelligence comprises the following steps: Step S101, based on multiple physiological parameters for anesthesia depth classification, multiple standard feature data are obtained, and a target basic model is selected from multiple preset models to be selected, and model training is performed to generate a target classification model.

[0031] In an embodiment of the present invention, multiple physiological parameters based on anesthesia depth classification are used to obtain multiple standard feature data, and the multiple standard feature data are divided and processed according to a preset data division ratio into a training set and a test set. Then, a target basic model is selected from multiple preset models to be selected, and then the target basic model is trained through the training set, and then the model is tested through the test set to construct a target classification model.

[0032] It is understandable that multiple physiological parameters, including electrocardiogram (ECG), blood oxygen saturation (SpO2), blood pressure, respiratory rate, etc.

[0033] It is understandable that there are multiple preset models to be selected, including convolutional neural network models, long short-term memory models, random forest models, and support vector machine models.

[0034] Specifically, Figure 2 A flow chart of generating a target classification model in the method provided in an embodiment of the present invention is shown.

[0035] Among them, in the preferred embodiment provided by the present invention, the multiple physiological parameters based on the anesthesia depth classification, obtaining multiple standard feature data, and selecting a target basic model from multiple preset models to be selected, performing model training, and generating a target classification model specifically include the following steps: Step S1011, acquiring a plurality of standard feature data based on a plurality of physiological parameters classified by anesthesia depth; Step S1012, dividing the plurality of standard feature data into a training set and a test set according to a preset data division ratio; Step S1013, selecting a target basic model from a plurality of preset models to be selected; Step S1014: training and testing the target basic model using the training set and the test set to generate a target classification model.

[0036] Among them, in the preferred embodiment provided by the present invention, the prediction process corresponding to the target basic model specifically includes the following steps: Obtain the patient's multi-dimensional physiological parameter time series data and anesthesia depth label data; The PC algorithm was used to analyze the causal effect strength of each physiological parameter and the anesthesia depth label in the multidimensional physiological parameter time series data to obtain the potential causal parameter candidate set; The chi-square test and G-test were used to perform statistical tests on the independence of each physiological parameter and the anesthesia depth label in the potential causal parameter candidate set. If the test results showed that the physiological parameter and the anesthesia depth label were not independent, it was determined that there was a potential causal relationship, and a list of significant parameters was obtained. By calculating the intervention likelihood ratio between each physiological parameter and the probability of anesthesia depth classification in the significant parameter column, the change range of the probability of anesthesia depth classification when a unit change occurs in the physiological parameter in the causal correlation parameter set is quantified, and the causal contribution of each physiological parameter is obtained; Obtain real-time physiological parameter features and surgical stage knowledge graph, and convert the surgical stage knowledge graph into a key-value pair structure; Encode the current physiological parameter features into a query vector, and calculate the similarity between the query vector and the key in the knowledge graph to determine the current surgical stage and obtain the matching result; Based on the monitoring focus of different physiological parameters in different surgical stages, weights are assigned to key parameters in the current surgical stage according to the matching results to obtain the dynamic weight of each physiological parameter; The dynamic weight of each physiological parameter is input into the gating network, and the causal contribution and the dynamic weight are weighted fused in the gating network to obtain the fused feature importance; Input the fused feature importance into the Sigmoid function to generate the gate value; Scaling the real-time physiological parameter feature vector according to the gate value to obtain a feature vector after gate adjustment; The gated feature vector is input into the Transformer classification model to obtain the classification result.

[0037] In the above scheme, the present invention designs causal contribution to ensure that the model focuses on the physiological parameters that really affect the depth of anesthesia, and effectively excludes interference parameters that are related to the depth of anesthesia but have no causal relationship. In addition, clinical prior knowledge is introduced, and through the attention mechanism, the model automatically switches the monitoring focus at different surgical stages, and can dynamically adjust the importance of features according to the surgical stage. Finally, the results of the first two steps are fused to perform comprehensive calculations of the gated function, and ultimately achieve intelligent screening of input features.

[0038] Among them, in the preferred implementation manner provided by the present invention, the process of training the target basic model specifically includes the following steps: Obtain the gated feature vector and the corresponding basic patient information, including ASA grade, age and medical history; Extract parameters strongly correlated with clinical risk from the gated and adjusted feature vectors to obtain a list of key risk parameters; The key risk parameter list is mapped according to the ASA classification to classify the patient into the preset risk level and obtain the patient's risk level; Set corresponding dynamic disturbances for patient risk levels; Input the real-time physiological parameter feature vector into the target basic model to obtain the current prediction result; Calculate the loss value based on the difference between the predicted result and the true label; The gradient of each input physiological parameter to the loss value in the real-time physiological parameter feature vector is calculated by the chain rule to obtain the gradient direction of each physiological parameter; Apply dynamic perturbations consistent with the gradient direction to all physiological parameters to obtain adversarial samples that link risk and causality; Patients are divided into meta-learning task groups according to their risk levels. The adversarial samples corresponding to each meta-learning task group are used to perform short-step gradient descent training on the gating network to obtain the optimized gating network parameters of each meta-learning task group and the updated gradient direction of each meta-learning task group. A dual-path training method is adopted to train the gated network based on the optimized gated network parameters, adversarial samples and feature vectors after gate adjustment to obtain the final fusion model.

[0039] Among them, in the preferred embodiment provided by the present invention, a dual-path training method is adopted, and the gated network is trained based on the optimized gated network parameters, adversarial samples and feature vectors after gate adjustment, and the final fusion model is obtained, which specifically includes the following steps: According to the risk level of the meta-learning task group, the updated gradient direction of each meta-learning task group is weighted averaged to obtain the global gradient vector and the consensus parameter label list; Analyze the correlation between the gradient direction of the consensus parameters in the consensus parameter marker list and the causal contribution, and obtain the common pattern; The specificity patterns were obtained by comparing the parameter differences of the optimized gating network parameters in the high-risk and low-risk meta-learning task groups; The common pattern encoding is set as the general rule, and the specific pattern encoding is set as the conditional rule, so as to obtain a cross-task feature importance map containing the general rule and the conditional rule, and the cross-task feature importance map is used to mark the common key parameters; The dual pathways include the causal pathway and the adversarial pathway; In the causal path, the gated network is trained using the gated conditioned feature vector; In the adversarial path, the optimized gating network parameters are loaded into the gating network, and the adversarial samples are used to train the gating network. The processing strategies for the meta-learning task group for high-risk patients learned in the meta-learning process are selectively distilled into the gating network according to the importance of the common key parameters to learn the characteristic processing patterns of high-risk patients. After the training is completed, the final fusion gating model is obtained, and the first gating network output is output in the causal path and the second gating network output is output in the adversarial path at the same time; The feature distribution of the first gating network output and the second gating network output are aligned to obtain the feature similarity matrix, and the feature similarity matrix is ​​used as the dynamic guidance weight of the attention mechanism in the Transformer classification model. The corresponding process has the following relationship: ; in, represents the attention calculation operation, Represent the query vector, key vector and value vector respectively; represents the feature similarity matrix, is a learnable parameter, , Represents the scaling factor.

[0040] In the above scheme, the present invention uses risk-graded adversarial training and ASA grading to dynamically adjust the strength of adversarial samples, so that the model remains stable in extreme physiological fluctuation scenarios of high-risk patients (ASA ≥ 3). In addition, it is designed with causal-adversarial dual-path protection, the causal gating mechanism blocks interference from irrelevant features, and adversarial training enhances the robustness of key parameters. The two work together to reduce false awakening events in intraoperative awakening tests. Through meta-learning to quickly adjust parameters, the model only needs 5-10 samples to complete parameter fine-tuning for different patient groups (such as slow metabolism in elderly patients and large fluctuations in physiological parameters in children), greatly improving the speed of application. In addition, risk feature distillation encapsulates the treatment strategies for high-risk patients (such as ASA grade 4 combined with heart failure) into an interpretable rule base, which automatically matches patient features to activate corresponding strategies during clinical deployment without manual configuration.

[0041] Furthermore, the artificial intelligence-based anesthesia depth classification method further includes the following steps: Step S102, evaluating and optimizing the target classification model to generate an optimized classification model.

[0042] In an embodiment of the present invention, the accuracy, precision, recall rate and F1 score of the target classification model are evaluated to reflect the accuracy of the positive class prediction, and the first evaluation data is recorded. Then, based on the ROC curve and the AUC value, the target classification model is evaluated to evaluate the performance of the model under different classification thresholds, and the second evaluation data is recorded. The target classification model is hyperparameter tuned by grid search, random search or Bayesian optimization to find the optimal learning rate, number of layers, number of neurons and other hyperparameters, and the hyperparameter tuning data is recorded. By comprehensively combining the first evaluation data, the second evaluation data and the hyperparameter tuning data, the target classification model is optimized to generate an optimized classification model.

[0043] Specifically, Figure 3 A flow chart of generating an optimized classification model in the method provided in an embodiment of the present invention is shown.

[0044] Among them, in the preferred embodiment provided by the present invention, the evaluation and optimization of the target classification model to generate the optimized classification model specifically includes the following steps: Step S1021, evaluating the accuracy, precision, recall rate and F1 score of the target classification model, and recording first evaluation data; Step S1022, evaluating the target classification model based on the ROC curve and the AUC value, and recording second evaluation data; Step S1023, performing hyperparameter tuning on the target classification model and recording hyperparameter tuning data; Step S1024, comprehensively analyzing the first evaluation data, the second evaluation data, and the hyperparameter tuning data, optimizing the target classification model, and generating an optimized classification model.

[0045] Among them, in the preferred implementation manner provided by the present invention, the target classification model is optimized by combining the first evaluation data, the second evaluation data and the hyperparameter tuning data, and generating the optimized classification model specifically comprises the following steps: Step S10241, obtaining an F1 score and an AUC value from the first evaluation data and the second evaluation data, and recording the F1 score and the AUC value in the form of decimals to obtain an F1 value and an AUC value; Step S10242, obtaining the original model parameter quantity of the target basic model and the tuned parameter quantity after hyperparameter tuning, and calculating the change ratio of the absolute difference between the original model parameter quantity and the tuned parameter quantity to obtain the parameter quantity change ratio; Step S10243, combining the F1 value and the AUC value in a harmonic mean form to obtain a core indicator balance item; Step S10244, calculating the complexity penalty term according to the parameter change ratio; Step S10245: merge the core indicator balance item and the complexity penalty item to obtain the final optimization score Step S10246, taking the parameter amounts of different hyperparameter combinations in the hyperparameter tuning process as input, repeating steps S10241 to S10245 to obtain final optimization scores of different hyperparameter combinations; Step S10247, by comparing the final optimization scores of different hyperparameter combinations, the target classification model corresponding to the hyperparameter combination with the highest final optimization score is selected as the optimized classification model.

[0046] In the above scheme, the present invention overcomes the shortcoming of the traditional arithmetic mean being insensitive to extreme values ​​by combining the F1 score (focusing on category balance) with the AUC value (global ranking ability) in the form of a harmonic mean. In anesthesia depth monitoring, it is necessary to avoid both intraoperative awareness (requiring high recall) and excessive sedation (requiring high accuracy). This balancing mechanism makes the model optimization direction meet the core clinical needs. In addition, the use of ln(1+parameter quantity change ratio) to perform nonlinear penalty on parameter quantity growth can suppress invalid parameter quantity expansion earlier than linear penalty (such as directly using parameter quantity change ratio). For example, when the parameter quantity change ratio = 0.5, the penalty value is 0.405 / 10≈0.04, while the linear penalty is 0.5 / 10=0.05-the logarithmic function is more sensitive to initial growth. In computationally constrained scenarios such as embedded anesthesia monitors (such as GE Datex-Ohmeda S / 5), this method can reduce the number of model parameters.

[0047] Furthermore, the artificial intelligence-based anesthesia depth classification method further includes the following steps: Step S103, collecting data according to the plurality of physiological parameters, acquiring physiological parameter data, and preprocessing the physiological parameter data to generate standard parameter data.

[0048] In an embodiment of the present invention, data collection is performed in real time during surgical anesthesia according to multiple physiological parameters to obtain physiological parameter data, and then the physiological parameter data is denoised by bandpass filtering, wavelet denoising and / or adaptive filtering to generate denoised parameter data, and then the denoised parameter data is zero-meaned or standardized to generate standard parameter data to ensure data consistency.

[0049] Specifically, Figure 4 A flow chart of generating standard parameter data in the method provided in an embodiment of the present invention is shown.

[0050] Among them, in the preferred embodiment provided by the present invention, the data collection is performed according to the multiple physiological parameters, the physiological parameter data is obtained, and the physiological parameter data is preprocessed to generate standard parameter data, which specifically includes the following steps: Step S1031, collecting data according to the plurality of physiological parameters to obtain physiological parameter data; Step S1032, performing denoising processing on the physiological parameter data to generate denoised parameter data; Step S1033, standardizing the denoising parameter data to generate standard parameter data.

[0051] Furthermore, the artificial intelligence-based anesthesia depth classification method further includes the following steps: Step S104, performing multiple feature extraction on the standard parameter data to obtain parameter feature data.

[0052] In an embodiment of the present invention, time domain feature data is obtained by performing time domain feature extraction (mean, variance, kurtosis and skewness) on the standard parameter data, and frequency domain feature extraction (power spectral density, frequency band energy) is performed on the standard parameter data to obtain frequency domain feature data, time-frequency feature extraction (wavelet transform, short-time Fourier transform) is performed on the standard parameter data to obtain time-frequency feature data, nonlinear feature extraction (entropy, Lyapunov exponent) is performed on the standard parameter data to obtain nonlinear feature data, and then the time domain feature data, frequency domain feature data, time-frequency feature data and nonlinear feature data are comprehensively sorted to generate parameter feature data.

[0053] Specifically, Figure 5A flow chart of obtaining parameter characteristic data in the method provided in an embodiment of the present invention is shown.

[0054] Among them, in the preferred embodiment provided by the present invention, the multiple feature extraction of the standard parameter data to obtain parameter feature data specifically includes the following steps: Step S1041, extracting time domain features from the standard parameter data to obtain time domain feature data; Step S1042, extracting frequency domain features from the standard parameter data to obtain frequency domain feature data; Step S1043, extracting time-frequency features from the standard parameter data to obtain time-frequency feature data; Step S1044, performing nonlinear feature extraction on the standard parameter data to obtain nonlinear feature data; Step S1045, synthesizing the time domain feature data, the frequency domain feature data, the time-frequency feature data and the nonlinear feature data to generate parameter feature data.

[0055] Furthermore, the artificial intelligence-based anesthesia depth classification method further includes the following steps: Step S105, importing the parameter characteristic data into the optimized classification model, performing anesthesia depth classification analysis, and acquiring and displaying depth classification results.

[0056] In an embodiment of the present invention, parameter feature data is imported into an optimized classification model, and the parameter feature data is automatically classified and analyzed for anesthesia depth through the optimized classification model to obtain depth classification results, and the depth classification results are displayed in real time.

[0057] Furthermore, Figure 6 The application architecture diagram of the system provided by the embodiment of the present invention is shown.

[0058] Among them, in another preferred embodiment provided by the present invention, an anesthesia depth classification system based on artificial intelligence includes: The model selection training unit 101 is used to obtain multiple standard feature data based on multiple physiological parameters for anesthesia depth classification, and select a target basic model from multiple preset models to be selected, perform model training, and generate a target classification model.

[0059] In an embodiment of the present invention, the model selection training unit 101 obtains multiple standard feature data based on multiple physiological parameters of anesthesia depth classification, divides and processes the multiple standard feature data according to a preset data division ratio, divides them into a training set and a test set, and then selects a target basic model from multiple preset models to be selected, and then trains the target basic model through the training set, and then tests the model through the test set to construct a target classification model.

[0060] Specifically, Figure 7 A structural block diagram of the model selection training unit 101 in the system provided by an embodiment of the present invention is shown.

[0061] Among them, in the preferred embodiment provided by the present invention, the model selection training unit 101 specifically includes: The data acquisition module 1011 is used to acquire a plurality of standard characteristic data based on a plurality of physiological parameters classified by anesthesia depth; A data partitioning module 1012 is used to partition the plurality of standard feature data into a training set and a test set according to a preset data partitioning ratio; A model selection module 1013 is used to select a target basic model from a plurality of preset models to be selected; The training and testing module 1014 is used to train and test the target basic model through the training set and the test set to generate a target classification model.

[0062] Furthermore, the artificial intelligence-based anesthesia depth classification system also includes: The model evaluation and optimization unit 102 is used to evaluate and optimize the target classification model to generate an optimized classification model.

[0063] In an embodiment of the present invention, the model evaluation and optimization unit 102 evaluates the accuracy, precision, recall rate and F1 score of the target classification model to reflect the accuracy of the positive class prediction, records the first evaluation data, and then evaluates the target classification model based on the ROC curve and the AUC value to evaluate the performance of the model under different classification thresholds, records the second evaluation data, and performs hyperparameter tuning on the target classification model through grid search, random search or Bayesian optimization and other methods to find the optimal learning rate, number of layers, number of neurons and other hyperparameters, records the hyperparameter tuning data, optimizes the target classification model by combining the first evaluation data, the second evaluation data and the hyperparameter tuning data to generate an optimized classification model.

[0064] Specifically, Figure 8 It shows a structural block diagram of the model evaluation and optimization unit 102 in the system provided by an embodiment of the present invention.

[0065] Among them, in the preferred embodiment provided by the present invention, the model evaluation optimization unit 102 specifically includes: A first evaluation module 1021 is used to evaluate the accuracy, precision, recall rate and F1 score of the target classification model and record first evaluation data; A second evaluation module 1022 is used to evaluate the target classification model based on the ROC curve and the AUC value, and record second evaluation data; A hyperparameter tuning module 1023 is used to perform hyperparameter tuning on the target classification model and record hyperparameter tuning data; The model optimization module 1024 is used to optimize the target classification model by integrating the first evaluation data, the second evaluation data and the hyperparameter tuning data to generate an optimized classification model.

[0066] Furthermore, the artificial intelligence-based anesthesia depth classification system also includes: The data collection and processing unit 103 is used to collect data according to the multiple physiological parameters, obtain physiological parameter data, and perform data preprocessing on the physiological parameter data to generate standard parameter data.

[0067] In an embodiment of the present invention, the data collection and processing unit 103 collects data in real time according to multiple physiological parameters during surgical anesthesia to obtain physiological parameter data, and then performs denoising processing such as bandpass filtering, wavelet denoising and / or adaptive filtering on the physiological parameter data to generate denoised parameter data, and then performs zero mean or standardization processing on the denoised parameter data to generate standard parameter data to ensure data consistency.

[0068] The multiple feature extraction unit 104 is used to perform multiple feature extraction on the standard parameter data to obtain parameter feature data.

[0069] In an embodiment of the present invention, the multiple feature extraction unit 104 performs time domain feature extraction (mean, variance, kurtosis and skewness) on the standard parameter data to obtain time domain feature data, performs frequency domain feature extraction (power spectral density, frequency band energy) on the standard parameter data to obtain frequency domain feature data, performs time-frequency feature extraction (wavelet transform, short-time Fourier transform) on the standard parameter data to obtain time-frequency feature data, performs nonlinear feature extraction (entropy, Lyapunov exponent) on the standard parameter data to obtain nonlinear feature data, and then comprehensively organizes the time domain feature data, frequency domain feature data, time-frequency feature data and nonlinear feature data to generate parameter feature data.

[0070] Specifically, Fig. 9 The structure block diagram of the multiple feature extraction unit 104 in the system provided by the embodiment of the present invention is shown.

[0071] Among them, in the preferred embodiment provided by the present invention, the multiple feature extraction unit 104 specifically includes: The time domain feature extraction module 1041 is used to extract the time domain features of the standard parameter data to obtain the time domain feature data; The frequency domain feature extraction module 1042 is used to extract the frequency domain features of the standard parameter data to obtain frequency domain feature data; The time-frequency feature extraction module 1043 is used to extract the time-frequency features of the standard parameter data to obtain the time-frequency feature data; A nonlinear feature extraction module 1044 is used to extract nonlinear features from the standard parameter data to obtain nonlinear feature data; The feature data generating module 1045 is used to generate parameter feature data by integrating the time domain feature data, the frequency domain feature data, the time-frequency feature data and the nonlinear feature data.

[0072] Furthermore, the artificial intelligence-based anesthesia depth classification system also includes: The anesthesia depth classification unit 105 is used to import the parameter characteristic data into the optimized classification model, perform anesthesia depth classification analysis, and obtain and display depth classification results.

[0073] In an embodiment of the present invention, the anesthesia depth classification unit 105 imports the parameter feature data into the optimized classification model, performs automatic classification analysis of the anesthesia depth on the parameter feature data through the optimized classification model, obtains the depth classification result, and displays the depth classification result in real time.

[0074] It should be understood that, although each step in the flow chart of each embodiment of the present invention is shown in sequence according to the indication of the arrow, these steps are not necessarily performed in sequence according to the order indicated by the arrow. Unless there is a clear explanation in this article, the execution of these steps does not have a strict order restriction, and these steps can be performed in other orders. Moreover, at least a portion of the steps in each embodiment may include a plurality of sub-steps or a plurality of stages, and these sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the execution order of these sub-steps or stages is not necessarily performed in sequence, but can be performed in turn or alternately with at least a portion of other steps or sub-steps or stages of other steps.

[0075] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0076] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0077] The above-mentioned embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

[0078] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An artificial intelligence-based anesthesia depth classification method, characterized in that: The method specifically comprises the following steps: Based on multiple physiological parameters for anesthesia depth classification, multiple standard feature data are obtained, and a target basic model is selected from multiple preset models to be selected, and model training is performed to generate a target classification model; Evaluating and optimizing the target classification model to generate an optimized classification model; According to the plurality of physiological parameters, data collection is performed to obtain physiological parameter data, and data preprocessing is performed on the physiological parameter data to generate standard parameter data; Performing multiple feature extraction on the standard parameter data to obtain parameter feature data; The parameter characteristic data is imported into the optimized classification model to perform anesthesia depth classification analysis, and the depth classification result is obtained and displayed.

2. The method for classifying anesthesia depth based on artificial intelligence according to claim 1, characterized in that: The method of obtaining a plurality of standard feature data based on a plurality of physiological parameters for anesthesia depth classification, selecting a target basic model from a plurality of preset models to be selected, and performing model training to generate a target classification model specifically comprises the following steps: Based on multiple physiological parameters of anesthesia depth classification, multiple standard feature data are obtained; According to a preset data division ratio, the plurality of standard feature data are divided into a training set and a test set; Select a target basic model from multiple preset models to be selected; The target basic model is trained and tested by using the training set and the test set to generate a target classification model.

3. The method for classifying anesthesia depth based on artificial intelligence according to claim 2, characterized in that: The prediction process corresponding to the target basic model specifically includes the following steps: Obtain the patient's multi-dimensional physiological parameter time series data and anesthesia depth label data; The PC algorithm was used to analyze the causal effect strength of each physiological parameter and the anesthesia depth label in the multidimensional physiological parameter time series data to obtain the potential causal parameter candidate set; For each physiological parameter and anesthesia depth label in the potential causal parameter candidate set, the chi-square test and G-test independence are used for statistical tests. If the test results show that the physiological parameter and the anesthesia depth label are not independent, it is determined that there is a potential causal relationship, and a list of significant parameters is obtained; By calculating the intervention likelihood ratio between each physiological parameter and the probability of anesthesia depth classification in the significant parameter column, the change amplitude of the probability of anesthesia depth classification when the physiological parameter changes by one unit is quantified, and the causal contribution of each physiological parameter is obtained; Obtain real-time physiological parameter features and surgical stage knowledge graph, and convert the surgical stage knowledge graph into a key-value pair structure; Encode the current physiological parameter features into a query vector, and calculate the similarity between the query vector and the key in the knowledge graph to determine the current surgical stage and obtain the matching result; Based on the monitoring focus of different physiological parameters in different surgical stages, weights are assigned to key parameters in the current surgical stage according to the matching results to obtain the dynamic weight of each physiological parameter; The dynamic weight of each physiological parameter is input into the gating network, and the causal contribution and the dynamic weight are weighted fused in the gating network to obtain the fused feature importance; Input the fused feature importance into the Sigmoid function to generate the gate value; Scaling the real-time physiological parameter feature vector according to the gate value to obtain a feature vector after gate adjustment; The gated feature vector is input into the Transformer classification model to obtain the classification result.

4. The method for classifying anesthesia depth based on artificial intelligence according to claim 3, characterized in that: The process of training the target basic model specifically includes the following steps: Obtain the gated feature vector and the corresponding basic patient information, including ASA grade, age and medical history; Extract parameters strongly correlated with clinical risk from the gated and adjusted feature vectors to obtain a list of key risk parameters; The key risk parameter list is mapped according to the ASA classification to classify the patient into the preset risk level and obtain the patient's risk level; Set corresponding dynamic disturbances for patient risk levels; Input the real-time physiological parameter feature vector into the target basic model to obtain the current prediction result; Calculate the loss value based on the difference between the predicted result and the true label; The gradient of each input physiological parameter to the loss value in the real-time physiological parameter feature vector is calculated by the chain rule to obtain the gradient direction of each physiological parameter; Apply dynamic perturbations consistent with the gradient direction to all physiological parameters to obtain adversarial samples that link risk and causality; Patients are divided into meta-learning task groups according to their risk levels. The adversarial samples corresponding to each meta-learning task group are used to perform short-step gradient descent training on the gating network to obtain the optimized gating network parameters of each meta-learning task group and the updated gradient direction of each meta-learning task group. A dual-path training method is adopted to train the gated network based on the optimized gated network parameters, adversarial samples and feature vectors after gate adjustment to obtain the final fusion model.

5. The method for classifying anesthesia depth based on artificial intelligence according to claim 4, characterized in that: Using a dual-path training method, the gated network is trained based on the optimized gated network parameters, adversarial samples, and feature vectors after gate adjustment. The final fusion model includes the following steps: According to the risk level of the meta-learning task group, the updated gradient direction of each meta-learning task group is weighted averaged to obtain the global gradient vector and the consensus parameter label list; Analyze the correlation between the gradient direction of the consensus parameters in the consensus parameter marker list and the causal contribution, and obtain the common pattern; The specificity patterns were obtained by comparing the parameter differences of the optimized gating network parameters in the high-risk and low-risk meta-learning task groups; The common pattern encoding is set as the general rule, and the specific pattern encoding is set as the conditional rule, so as to obtain a cross-task feature importance map containing the general rule and the conditional rule, and the cross-task feature importance map is used to mark the common key parameters; The dual pathways include the causal pathway and the adversarial pathway; In the causal path, the gated network is trained using the gated conditioned feature vector; In the adversarial path, the optimized gating network parameters are loaded into the gating network, and the adversarial samples are used to train the gating network. The processing strategies for the meta-learning task group for high-risk patients learned in the meta-learning process are selectively distilled into the gating network according to the importance of the common key parameters to learn the characteristic processing patterns of high-risk patients. After the training is completed, the final fusion gating model is obtained, and the first gating network output is output in the causal path and the second gating network output is output in the adversarial path at the same time; The feature distribution of the first gating network output and the second gating network output are aligned to obtain a feature similarity matrix, which is used as the dynamic guidance weight of the attention mechanism in the Transformer classification model.

6. The method for classifying anesthesia depth based on artificial intelligence according to claim 5, characterized in that: The process of using the feature similarity matrix as the dynamic guidance weight of the attention mechanism in the Transformer classification model corresponds to the following relationship: ； in, represents the attention calculation operation, Represent the query vector, key vector and value vector respectively; represents the feature similarity matrix, is a learnable parameter, , Represents the scaling factor.

7. The method for classifying anesthesia depth based on artificial intelligence according to claim 6, characterized in that: The evaluating and optimizing the target classification model to generate an optimized classification model specifically comprises the following steps: Evaluate the accuracy, precision, recall rate and F1 score of the target classification model, and record first evaluation data; Based on the ROC curve and the AUC value, the target classification model is evaluated and second evaluation data is recorded; Performing hyperparameter tuning on the target classification model and recording hyperparameter tuning data; The target classification model is optimized based on the first evaluation data, the second evaluation data and the hyperparameter tuning data to generate an optimized classification model.

8. The method for classifying anesthesia depth based on artificial intelligence according to claim 7, characterized in that: Combining the first evaluation data, the second evaluation data and the hyperparameter tuning data to optimize the target classification model, generating an optimized classification model specifically comprises the following steps: Step S10241, obtaining an F1 score and an AUC value from the first evaluation data and the second evaluation data, and recording the F1 score and the AUC value in the form of decimals to obtain an F1 value and an AUC value; Step S10242, obtaining the original model parameter quantity of the target basic model and the tuned parameter quantity after hyperparameter tuning, and calculating the change ratio of the absolute difference between the original model parameter quantity and the tuned parameter quantity to obtain the parameter quantity change ratio; Step S10243, combining the F1 value and the AUC value in a harmonic mean form to obtain a core indicator balance item; Step S10244, calculating the complexity penalty term according to the parameter change ratio; Step S10245, combining the core indicator balance item and the complexity penalty item to obtain the final optimization score; Step S10246, taking the parameter amounts of different hyperparameter combinations in the hyperparameter tuning process as input, repeating steps S10241 to S10245 to obtain final optimization scores of different hyperparameter combinations; Step S10247, by comparing the final optimization scores of different hyperparameter combinations, the target classification model corresponding to the hyperparameter combination with the highest final optimization score is selected as the optimized classification model.

9. The method for classifying anesthesia depth based on artificial intelligence according to claim 8, characterized in that: The step of collecting data according to the plurality of physiological parameters, acquiring physiological parameter data, and preprocessing the physiological parameter data to generate standard parameter data specifically includes the following steps: According to the plurality of physiological parameters, data collection is performed to obtain physiological parameter data; Performing denoising processing on the physiological parameter data to generate denoised parameter data; The denoising parameter data is standardized to generate standard parameter data.

10. The artificial intelligence-based anesthesia depth classification method according to claim 9, characterized in that: The performing of multiple feature extraction on the standard parameter data to obtain parameter feature data specifically comprises the following steps: Performing time domain feature extraction on the standard parameter data to obtain time domain feature data; Performing frequency domain feature extraction on the standard parameter data to obtain frequency domain feature data; Extracting time-frequency features from the standard parameter data to obtain time-frequency feature data; Performing nonlinear feature extraction on the standard parameter data to obtain nonlinear feature data; The time domain feature data, the frequency domain feature data, the time-frequency feature data and the nonlinear feature data are integrated to generate parameter feature data.

11. An artificial intelligence-based anesthesia depth classification system, the system being applied to the artificial intelligence-based anesthesia depth classification method according to any one of claims 1 to 10, characterized in that: The system comprises a model selection training unit, a model evaluation optimization unit, a data collection and processing unit, a multiple feature extraction unit and an anesthesia depth classification unit, wherein: A model selection training unit is used to obtain multiple standard feature data based on multiple physiological parameters of anesthesia depth classification, and select a target basic model from multiple preset models to be selected, perform model training, and generate a target classification model; A model evaluation and optimization unit, used to evaluate and optimize the target classification model to generate an optimized classification model; A data collection and processing unit, used to collect data according to the plurality of physiological parameters, obtain physiological parameter data, and perform data preprocessing on the physiological parameter data to generate standard parameter data; A multiple feature extraction unit, used for performing multiple feature extraction on the standard parameter data to obtain parameter feature data; The anesthesia depth classification unit is used to import the parameter characteristic data into the optimized classification model, perform anesthesia depth classification analysis, and obtain and display depth classification results.

Images