Anesthesia depth monitoring system for general anesthesia patient based on medical information processing

Abstract

The present application relates to general anesthesia patient anesthesia depth monitoring system based on medical information processing, the system includes multidimensional data acquisition module, preliminary optimization processing module, multidimensional consensus model construction module and mirror generation contrast module. The present application obtains original data through data acquisition; adopts waveform data processing, text data processing, numerical data processing, data alignment assembly and data set division preliminary optimization processing method; adopt multidimensional consensus model output anesthesia depth grade prediction result, through constructing three kinds of sub-networks combined with debate consensus module, aligning the feature space of each mode and referring to each other information to correct the judgment, improve the anesthesia depth monitoring reliability and explainability of complex, atypical cases; adopt mirror generation contrast method to detect whether the multidimensional consensus model produces biased prediction due to demographic attributes, test whether the model decision is different due to sensitive attributes, provide better safety guarantee for clinical decision.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring of the depth of anesthesia in patients undergoing general anesthesia, specifically to a monitoring system for the depth of anesthesia in patients undergoing general anesthesia based on medical information processing. Background Technology

[0002] Monitoring the depth of anesthesia in patients undergoing general anesthesia involves real-time acquisition and analysis of physiological parameters such as electroencephalogram (EEG), heart rate, and blood pressure during the procedure. This data, combined with advanced algorithms and signal processing techniques, is used to assess and adjust the depth of anesthesia. This monitoring allows for precise control of the depth of anesthesia, avoiding the risks associated with excessively deep or shallow anesthesia, reducing the likelihood of intraoperative consciousness loss or drug overdose, improving patient safety, and optimizing the use of anesthetic drugs.

[0003] However, traditional anesthesia depth monitoring systems for general anesthesia patients suffer from several technical problems. Firstly, the multimodal data fusion strategies employed struggle to balance the contributions of highly heterogeneous EEG waveforms, textual semantics, and structured indicators, which exhibit significant differences in noise distribution. Secondly, when different modal evidence points to contradictory conclusions, there is a lack of effective conflict resolution mechanisms and contingency plans for failed resolutions, leading to insufficient model robustness and decision confidence. Furthermore, traditional anesthesia depth monitoring systems often focus solely on overall prediction accuracy, passively examining the distribution or overall performance differences among different groups in the training data. They lack methods to detect whether model decisions are systematically influenced by demographic attributes, potentially resulting in systematic underestimation or overestimation of anesthesia depth for specific populations, thus weakening the universality of clinical decision-making. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides a system for monitoring the depth of anesthesia in general anesthesia patients based on medical information processing. Traditional systems for monitoring the depth of anesthesia in general anesthesia suffer from several problems. First, the multimodal data fusion strategies employed by these systems struggle to balance the contributions of highly heterogeneous and noise-distributed EEG waveforms, textual semantics, and structured indicators. Second, when different modal evidence points to contradictory conclusions, the systems lack effective negotiation mechanisms and contingency plans for failed negotiations, leading to insufficient model robustness and decision confidence. This invention creatively employs a multidimensional consensus model to output the predicted level of anesthesia depth. By constructing three sub-networks combined with a debate consensus module, the system retains the sensitivity of each sub-network to specific neurophysiological signals and clinical indicators. Finally, through a debate consensus mechanism, the feature spaces of each modality are aligned, and information from each other is referenced to correct the judgment. This approach enhances the reliability and interpretability of anesthesia depth monitoring for complex and atypical cases. Traditional anesthesia depth monitoring systems for general anesthesia patients often focus solely on overall prediction accuracy, passively examining the distribution or overall performance differences among different groups in the training data. They lack methods to detect whether model decisions are systematically influenced by demographic attributes, potentially leading to systematic underestimation or overestimation of anesthesia depth for specific populations and weakening the universality of clinical decisions. This innovative approach employs a mirror-generated comparison method to detect whether multidimensional consensus models produce biased predictions due to demographic attributes. While maintaining the original disease characteristics, only the patient's demographic profile is modified to generate a medically reasonable virtual control sample, examining whether model decisions are distorted by sensitive attributes, thus providing better safety assurance for clinical decision-making.

[0005] The technical solution adopted by the present invention is as follows: The anesthesia depth monitoring system for general anesthesia patients based on medical information processing provided by the present invention includes a multi-dimensional data acquisition module, a preliminary optimization processing module, a multi-dimensional consensus model construction module, and a mirror generation and comparison module;

[0006] The multidimensional data acquisition module obtains the raw dataset of anesthesia depth monitoring through data acquisition and sends the raw dataset of anesthesia depth monitoring to the preliminary optimization processing module;

[0007] The preliminary optimization processing module employs waveform data processing, text data processing, numerical data processing, data alignment and assembly, and dataset partitioning as preliminary optimization processing methods to obtain the dataset to be monitored, the deep monitoring training set, and the deep monitoring test set. The dataset to be monitored is then sent to the mirror generation and comparison module, and the deep monitoring training set and the deep monitoring test set are sent to the multidimensional consensus model construction module.

[0008] The multidimensional consensus model construction module obtains a multidimensional consensus model by constructing three sub-networks combined with a debate consensus module, and sends the multidimensional consensus model to the mirror generation and comparison module.

[0009] The mirror generation and comparison module constructs a variational autoencoder model to perform mirror generation and comparison, thereby obtaining the final reference result for anesthesia depth monitoring.

[0010] Furthermore, in the multidimensional data acquisition module, the original dataset for anesthesia depth monitoring specifically includes a historical original dataset for anesthesia depth monitoring and a current original dataset for anesthesia depth monitoring. Both the historical original dataset for anesthesia depth monitoring and the current original dataset for anesthesia depth monitoring contain waveform data, text data, and numerical data. The historical original dataset for anesthesia depth monitoring also includes anesthesia depth level label data.

[0011] Furthermore, in the preliminary optimization processing module, the waveform data processing is used to remove interference in the waveform and unify the time resolution. Specifically, it filters out baseline drift and power frequency noise through a bandpass filter, then resamples all waveform data to 200Hz and truncates a fixed-length window, and performs Z-score normalization to obtain the processed waveform data.

[0012] The text data processing is used to convert unstructured medical record text into a form that the model can process. Specifically, it involves word segmentation, stop word removal, entity standardization, and length truncation to obtain text sequence data with controllable length.

[0013] The numerical data processing is used to fill in missing values ​​and eliminate dimensional differences. Specifically, it involves median filling, outlier truncation, min-max scaling, and one-hot encoding of categorical variables to obtain the processed numerical features.

[0014] The data alignment assembly is used to align different modal data of the same patient in time. Specifically, it involves concatenating processed waveform, text and numerical data into the same sample through patient association and timestamp matching to obtain a feature format that the model can process.

[0015] The dataset partitioning is used to obtain training data and test data. Specifically, the original historical anesthesia depth monitoring dataset is split to obtain a depth monitoring training set and a depth monitoring test set.

[0016] The current anesthesia depth monitoring raw dataset is preliminarily optimized through waveform data processing, text data processing, numerical data processing, and data alignment assembly to obtain the dataset to be monitored. The historical anesthesia depth monitoring raw dataset is preliminarily optimized through waveform data processing, text data processing, numerical data processing, data alignment assembly, and dataset partitioning to obtain the depth monitoring training set and the depth monitoring test set.

[0017] Furthermore, in the multidimensional consensus model construction module, different modal data are processed from three independent subnets respectively, and a consensus depth level is reached through a debate mechanism. Specifically, a multidimensional consensus model is obtained by constructing a waveform processing subnet, a text processing subnet, and a numerical processing subnet, combined with the debate consensus module.

[0018] The multidimensional consensus model construction module specifically includes a waveform processing subnet, a text processing subnet, a numerical processing subnet, a consensus quality calculation and debate consensus module;

[0019] The waveform processing subnetwork is used to extract features from high-frequency physiological waveforms and provide a preliminary assessment of anesthesia depth. Its content includes:

[0020] The residual dilated convolutional block design is used to gradually extract local to global temporal patterns from multi-channel waveforms. Specifically, it involves stacking multiple residual dilated convolutional blocks with progressively increasing dilation rates after 1×1 convolution to increase the dimensionality. Each residual dilated convolutional block contains two dilation convolutions and residual connections to obtain waveform temporal features.

[0021] Waveform-dimensional anesthesia depth prediction is used to compress temporal features into a fixed-length vector and map it to anesthesia depth judgment. Specifically, it obtains waveform hidden features by performing global average pooling on the waveform temporal features in the time dimension, and then outputs waveform-dimensional confidence scores and waveform-dimensional preliminary anesthesia depth level labels through fully connected layers and softmax classifiers respectively.

[0022] The text processing subnetwork is used to extract semantic features from structured medical record text and provide a preliminary assessment of anesthesia depth. Its content includes:

[0023] Lightweight BERT encoding is used to transform text sequences into context-rich semantic vectors. Specifically, it processes the input text sequence data through a pre-trained BioBERT-mini model, extracts the output at the [CLS] position, and obtains the text encoding features.

[0024] The text-dimensional anesthesia depth prediction is used to compress high-dimensional semantic vectors to a uniform dimension and output an anesthesia depth judgment. Specifically, the text encoding features are reduced in dimensionality through a fully connected layer to obtain text hidden layer features. Then, the text dimension confidence score and the text dimension preliminary anesthesia depth level label are output through a fully connected layer and a softmax classifier, respectively.

[0025] The numerical processing subnet is used to extract numerical features from laboratory values ​​and vital sign statistics and to provide a preliminary judgment of anesthesia depth. Specifically, it reduces the dimensionality of the numerical data layer by layer through three fully connected layers. Each fully connected layer is followed by LeakyReLU activation and random inactivation to output the numerical hidden layer features. Then, it outputs the numerical dimension confidence score and the numerical dimension preliminary anesthesia depth level label through fully connected layers and softmax classifiers respectively.

[0026] The consensus quality calculation, used to summarize the initial judgment results of the three subnets and determine whether a debate is needed, includes:

[0027] Majority voting is used to determine the level of anesthesia depth that is agreed upon by the majority in the three subnets. Specifically, it involves counting the number of initial anesthesia depth level labels in the three subnets and taking the one with the most of the same number as the majority anesthesia depth level label.

[0028] The consensus quality score is used to quantify the degree of consistency among the three subnetworks. Specifically, it calculates the proportion of subnetworks that are consistent with the majority anesthesia depth level label to obtain the consensus quality score.

[0029] When the consensus quality score is 1, the initial anesthesia depth level labels output by the three subnets are consistent. The initial anesthesia depth level label and the consensus quality score are directly output as the output of the multidimensional consensus model. Otherwise, the content of the debate consensus module continues to be executed.

[0030] The debate consensus module is used to reach a higher consensus through feature-level interaction when there are disagreements among the three subnets. Its content includes:

[0031] Interactive space feature projection is used to map the hidden features of different subnets to the same dimensional space for interaction. Specifically, it maps waveform hidden features, text hidden features, and numerical hidden features to the same feature space through a shared fully connected layer to obtain waveform interactive features, text interactive features, and numerical interactive features.

[0032] Adversarial modality alignment is used to force the features of the three subnets to be modally indistinguishable in the feature space, thereby eliminating modality-specific bias. Specifically, it is achieved by adversarial training of the gradient inversion layer and the modality discriminator to obtain waveform alignment features, text alignment features, and numerical alignment features.

[0033] Cross-modal attention interaction is used to allow each subnet to adjust its own alignment features based on the features of other subnets. Specifically, each alignment feature is used as a query, and the other two alignment features are used as keys and values ​​in turn. Through multi-head cross-attention processing, updated waveform hidden features, updated text hidden features, and updated numerical hidden features are obtained.

[0034] The re-voting and final output are used to make a new judgment on the depth of anesthesia based on the features after the interaction and to determine the final output. Specifically, the updated waveform hidden features, updated text hidden features, and updated numerical hidden features are mapped to the same dimension as the hidden features of their corresponding subnets through linear transformation, and the re-judgment confidence score and re-judgment anesthesia depth level label of each subnet are output instead of the original hidden features.

[0035] The consensus quality score is recalculated based on the re-decision confidence score of each subnet. When the re-decision consensus quality score is 1, that is, the re-decision anesthesia depth level label of each subnet is consistent, the re-decision anesthesia depth level label and the re-decision consensus quality score are output as the output of the multidimensional consensus model. Otherwise, the re-decision anesthesia depth level label and the re-decision consensus quality score of the subnet with the highest re-decision confidence score among the three subnets are output as the output of the multidimensional consensus model.

[0036] Furthermore, in the mirror generation and comparison module, the method used to detect whether the multidimensional consensus model produces biased predictions due to the patient's demographic attributes is to generate a mirror sample that differs from the original sample only in demographic data by constructing a variational autoencoder. The original sample and the mirror sample are then input into the frozen multidimensional consensus model, and the output differences are compared to obtain the final reference result for anesthesia depth monitoring.

[0037] The mirror generation and comparison module specifically includes variational autoencoder design, medical rationality constraints, mirror sample generation, and frozen model inference.

[0038] The variational autoencoder design is used to encode multidimensional patient data into separable latent variables of pathophysiological and demographic attributes, including:

[0039] Multimodal coding is used to encode data from three modalities into a unified joint latent space. Specifically, waveform, text, and numerical features are extracted and concatenated by three independent small encoders. The resulting joint features are processed by a fully connected layer to obtain the sampling mean and sampling variance. The joint latent variables are obtained by random sampling based on a normal distribution with a dimension of 128.

[0040] Latent space decoupling is used to separate latent variables into two parts: pathophysiological attributes and demographic attributes. Specifically, the first 96 dimensions of the joint latent variables are used as the pathophysiological subspace and the last 32 dimensions are used as the demographic subspace. The mutual information between the two is minimized by the CLUB estimator to obtain two statistically independent subspaces.

[0041] Causal intervention design is used to ensure that demographic attributes affect pathophysiological attributes only through known causal paths. Specifically, it involves adding the demographic subspace to the pathophysiological subspace using a sparse causal matrix and applying L1 regularization to the sparse causal matrix to obtain enhanced joint latent variables.

[0042] Decoder construction is used to ensure that the decoupled latent variables can still completely reconstruct the generated samples. Specifically, the decoder reconstructs waveform, text and numerical data based on the enhanced joint latent variables to obtain the generated samples.

[0043] The medical rationality constraint is used to ensure that the generated sample is equivalent to the original sample in terms of pathophysiology and conforms to medical common sense. Specifically, the generated sample is processed by a Gaussian mixture model pre-trained based on the original dataset of historical anesthesia depth monitoring, and a rationality score in the range of [0,1] is output. When the variational autoencoder is trained, the Gaussian mixture model is fixed, and the variational autoencoder forces the rationality score to be close to 1 through adversarial loss.

[0044] The mirror sample generation is used to create a virtual patient sample with different demographic attributes but the same pathophysiological attributes. Specifically, for any original sample, its pathophysiological subspace after intervention is first obtained, then another original sample is randomly selected and its demographic subspace is obtained. The two are combined as mirror joint latent variables, and finally the mirror sample corresponding to the original sample is obtained through the decoder.

[0045] The frozen model inference is used to compare whether the anesthesia depth judgment of the original sample and the mirror sample are consistent. Specifically, the original sample and its corresponding mirror sample are respectively input into the frozen multidimensional consensus model to obtain their respective consensus quality scores and anesthesia depth level labels. If the anesthesia depth level labels obtained by the two are consistent, the consensus quality score and anesthesia depth level label corresponding to the original sample are directly output as the final anesthesia depth monitoring reference result. Otherwise, the consensus quality score and anesthesia depth level label of the two and the mandatory manual re-evaluation alarm are output as the final anesthesia depth monitoring reference result.

[0046] The beneficial effects achieved by the present invention using the above solution are as follows:

[0047] (1) In view of the technical problems of traditional general anesthesia patient anesthesia depth monitoring system, the multimodal data fusion strategy used is difficult to balance the contributions of EEG waveforms, text semantics and structured indicators with high heterogeneity and significant differences in noise distribution. Furthermore, when different modal evidence points to contradictory conclusions, there is a lack of effective negotiation and decision-making plan after negotiation failure, resulting in insufficient model robustness and decision confidence. This solution creatively adopts a multidimensional consensus model to output the anesthesia depth level prediction results. By constructing three sub-networks combined with a debate consensus module, the sensitivity of each sub-network to specific neurophysiological signals and clinical indicators is preserved. Then, through the debate consensus mechanism, the feature spaces of each modality are aligned and the judgment is corrected by referring to each other's information, which improves the reliability and interpretability of anesthesia depth monitoring for complex and atypical cases.

[0048] (2) Traditional general anesthesia depth monitoring systems often only focus on the overall accuracy of predictions and can only passively check the distribution of different groups or the overall performance differences in the training data. They lack the means to detect whether the model decision is systematically affected by demographic attributes, which may lead to the model underestimating or overestimating the depth of anesthesia for specific populations and weakening the universality of clinical decision-making. This solution creatively adopts the mirror generation comparison method to detect whether the multidimensional consensus model produces biased predictions due to demographic attributes. Under the premise of maintaining the original disease characteristics, only the patient's demographic profile is changed to generate a medically reasonable virtual control sample to test whether the model decision is distorted due to sensitive attributes, thus providing better safety assurance for clinical decision-making. Attached Figure Description

[0049] Figure 1 A schematic diagram of the module of the anesthesia depth monitoring system for general anesthesia patients based on medical information processing provided by the present invention;

[0050] Figure 2 A flowchart illustrating the initial optimization of the processing module;

[0051] Figure 3 A flowchart illustrating the process of building modules for a multidimensional consensus model;

[0052] Figure 4 A flowchart illustrating the process of generating a comparison module for mirror images.

[0053] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0054] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0055] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0056] Example 1, see Figure 1The present invention provides a system for monitoring the depth of anesthesia in general anesthesia patients based on medical information processing, which includes a multi-dimensional data acquisition module, a preliminary optimization processing module, a multi-dimensional consensus model construction module, and a mirror generation and comparison module.

[0057] The multidimensional data acquisition module obtains the raw dataset of anesthesia depth monitoring through data acquisition and sends the raw dataset of anesthesia depth monitoring to the preliminary optimization processing module;

[0058] The preliminary optimization processing module employs waveform data processing, text data processing, numerical data processing, data alignment and assembly, and dataset partitioning as preliminary optimization processing methods to obtain the dataset to be monitored, the deep monitoring training set, and the deep monitoring test set. The dataset to be monitored is then sent to the mirror generation and comparison module, and the deep monitoring training set and the deep monitoring test set are sent to the multidimensional consensus model construction module.

[0059] The multidimensional consensus model construction module obtains a multidimensional consensus model by constructing three sub-networks combined with a debate consensus module, and sends the multidimensional consensus model to the mirror generation and comparison module.

[0060] The mirror generation and comparison module constructs a variational autoencoder model to perform mirror generation and comparison, thereby obtaining the final reference result for anesthesia depth monitoring.

[0061] Example 2, see Figure 1 Based on the above embodiments, in the multidimensional data acquisition module, the original dataset for anesthesia depth monitoring specifically includes a historical original dataset for anesthesia depth monitoring and a current original dataset for anesthesia depth monitoring. Both the historical original dataset for anesthesia depth monitoring and the current original dataset for anesthesia depth monitoring contain waveform data, text data, and numerical data. The historical original dataset for anesthesia depth monitoring also includes anesthesia depth level label data.

[0062] The waveform data specifically includes electroencephalogram (EEG) waveform data, electrocardiogram (ECG) data, blood pressure waveform data, pulse oxygen saturation waveform data, and respiratory waveform data. The text data specifically includes medical history records, surgical records, anesthesia records, and disease diagnosis documents. The medical history records specifically include past medical history data, allergy history data, surgical history data, and medication history data. The surgical records specifically include surgical type data and surgical duration data. The anesthesia records specifically include anesthesia method data, anesthetic drug type data, and anesthetic drug dosage data. The numerical data specifically includes vital sign data and demographic data. The vital sign data specifically includes blood test data, resting heart rate data, respiratory rate data, body temperature data, and BMI data. The demographic data specifically includes age data, gender data, ethnicity data, weight data, and height data.

[0063] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In the preliminary optimization processing module, the waveform data processing is used to remove interference in the waveform and unify the time resolution. Specifically, it filters out baseline drift and power frequency noise through a bandpass filter, then resamples all waveform data to 200Hz and truncates a fixed-length window, and performs Z-score normalization to obtain the processed waveform data.

[0064] The text data processing is used to convert unstructured medical record text into a form that the model can process. Specifically, it involves word segmentation, stop word removal, entity standardization, and length truncation to obtain text sequence data with controllable length.

[0065] The numerical data processing is used to fill in missing values ​​and eliminate dimensional differences. Specifically, it involves median filling, outlier truncation, min-max scaling, and one-hot encoding of categorical variables to obtain the processed numerical features.

[0066] The data alignment assembly is used to align different modal data of the same patient in time. Specifically, it involves concatenating processed waveform, text and numerical data into the same sample through patient association and timestamp matching to obtain a feature format that the model can process.

[0067] The dataset partitioning is used to obtain training data and test data. Specifically, the original historical anesthesia depth monitoring dataset is split to obtain a depth monitoring training set and a depth monitoring test set.

[0068] The current anesthesia depth monitoring raw dataset is preliminarily optimized through waveform data processing, text data processing, numerical data processing, and data alignment assembly to obtain the dataset to be monitored. The historical anesthesia depth monitoring raw dataset is preliminarily optimized through waveform data processing, text data processing, numerical data processing, data alignment assembly, and dataset partitioning to obtain the depth monitoring training set and the depth monitoring test set.

[0069] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In the multidimensional consensus model construction module, it is used to process different modal data from three independent subnets and reach a consensus depth level through a debate mechanism. Specifically, by constructing a waveform processing subnet, a text processing subnet, and a numerical processing subnet, and combining them with the debate consensus module, a multidimensional consensus model is obtained.

[0070] The multidimensional consensus model construction module specifically includes a waveform processing subnet, a text processing subnet, a numerical processing subnet, a consensus quality calculation and debate consensus module;

[0071] The waveform processing subnetwork is used to extract features from high-frequency physiological waveforms and provide a preliminary assessment of anesthesia depth. Its content includes:

[0072] The residual dilated convolutional block design is used to progressively extract local to global temporal patterns from multi-channel waveforms. Specifically, it involves stacking multiple residual dilated convolutional blocks with progressively increasing dilation rates after 1×1 convolution to increase the dimensionality. Each residual dilated convolutional block contains two dilation convolutions and residual connections to obtain the waveform temporal features. The formula used is as follows:

[0073] ;

[0074] In the formula, This represents the waveform features after 1×1 convolution for dimensionality enhancement. This represents a 1×1 convolution function. This represents the input waveform data. This represents the output feature of the k-th residual dilated convolution block. Represents the ReLU activation function. This represents the output feature of the (k-1)th residual dilated convolution block. This represents the batch normalization function. The dilated convolution function represents the dilation rate di.

[0075] Waveform-dimensional anesthesia depth prediction is used to compress temporal features into a fixed-length vector and map it to anesthesia depth judgment. Specifically, it obtains waveform hidden features by performing global average pooling on the waveform temporal features in the time dimension, and then outputs waveform-dimensional confidence scores and waveform-dimensional preliminary anesthesia depth level labels through fully connected layers and softmax classifiers respectively.

[0076] The text processing subnetwork is used to extract semantic features from structured medical record text and provide a preliminary assessment of anesthesia depth. Its content includes:

[0077] Lightweight BERT encoding is used to transform text sequences into context-rich semantic vectors. Specifically, it processes the input text sequence data through a pre-trained BioBERT-mini model, extracts the output at the [CLS] position, and obtains the text encoding features.

[0078] The text-dimensional anesthesia depth prediction is used to compress high-dimensional semantic vectors to a uniform dimension and output an anesthesia depth judgment. Specifically, the text encoding features are reduced in dimensionality through a fully connected layer to obtain text hidden layer features. Then, the text dimension confidence score and the text dimension preliminary anesthesia depth level label are output through a fully connected layer and a softmax classifier, respectively.

[0079] The numerical processing subnet is used to extract numerical features from laboratory values ​​and vital sign statistics and to provide a preliminary judgment of anesthesia depth. Specifically, it reduces the dimensionality of the numerical data layer by layer through three fully connected layers. Each fully connected layer is followed by LeakyReLU activation and random inactivation to output the numerical hidden layer features. Then, it outputs the numerical dimension confidence score and the numerical dimension preliminary anesthesia depth level label through fully connected layers and softmax classifiers respectively.

[0080] The consensus quality calculation, used to summarize the initial judgment results of the three subnets and determine whether a debate is needed, includes:

[0081] Majority voting is used to determine the level of anesthesia depth that is agreed upon by the majority in the three subnets. Specifically, it involves counting the number of initial anesthesia depth level labels in the three subnets and taking the one with the most of the same number as the majority anesthesia depth level label.

[0082] The consensus quality score is used to quantify the degree of consistency among the three subnetworks. Specifically, it calculates the proportion of subnetworks that are consistent with the majority anesthesia depth grade label to obtain the consensus quality score. The formula used is as follows:

[0083] ;

[0084] In the formula, CQ represents the consensus quality score, a represents the subnet index, wa represents the waveform processing subnet, tx represents the text processing subnet, and num represents the numerical processing subnet. This represents the initial anesthesia depth level label output by subnet a. This indicates the majority of anesthesia depth rating labels. This represents an indicator function whose value is 1 when the initial anesthesia depth level label output by subnet a is consistent with the majority anesthesia depth level label, and 0 otherwise.

[0085] When the CQ value is 1, the initial anesthesia depth level labels output by the three subnets are consistent. This initial anesthesia depth level label and the consensus quality score are directly output as the output of the multidimensional consensus model. Otherwise, the content of the debate consensus module continues to be executed.

[0086] The debate consensus module is used to reach a higher consensus through feature-level interaction when there are disagreements among the three subnets. Its content includes:

[0087] Interactive space feature projection is used to map the hidden features of different subnets to the same dimensional space for interaction. Specifically, it maps waveform hidden features, text hidden features, and numerical hidden features to the same feature space through a shared fully connected layer to obtain waveform interactive features, text interactive features, and numerical interactive features.

[0088] Adversarial modality alignment is used to force the features of the three subnets to be modally indistinguishable in the feature space, eliminating modality-specific bias. Specifically, it obtains waveform alignment features, text alignment features, and numerical alignment features through adversarial training of a gradient inversion layer and a modality discriminator. The modality discriminator is specifically a two-layer multilayer perceptron, whose output is a subnet label used to identify the source subnet of the input interactive features. The loss function formula is as follows:

[0089] ;

[0090] In the formula, This represents the adversarial modal alignment loss value. This represents the function for calculating cross-entropy loss. This represents the function that runs the modal discriminator. This represents the gradient reversal layer function. This indicates the interaction characteristics originating from subnet a. Indicates the tag of the interactive feature subnet;

[0091] Cross-modal attention interaction is used to allow each subnet to adjust its own alignment features based on the features of other subnets. Specifically, each alignment feature is used as a query, and the other two alignment features are used as keys and values ​​in turn. Through multi-head cross-attention processing, updated waveform hidden features, updated text hidden features, and updated numerical hidden features are obtained.

[0092] The re-voting and final output are used to make a new judgment on the depth of anesthesia based on the features after the interaction and to determine the final output. Specifically, the updated waveform hidden features, updated text hidden features, and updated numerical hidden features are mapped to the same dimension as the hidden features of their corresponding subnets through linear transformation, and the re-judgment confidence score and re-judgment anesthesia depth level label of each subnet are output instead of the original hidden features.

[0093] The consensus quality score is recalculated based on the re-decision confidence score of each subnet. When the re-decision consensus quality score is 1, that is, the re-decision anesthesia depth level label of each subnet is consistent, the re-decision anesthesia depth level label and the re-decision consensus quality score are output as the output of the multidimensional consensus model. Otherwise, the re-decision anesthesia depth level label and the re-decision consensus quality score of the subnet with the highest re-decision confidence score among the three subnets are output as the output of the multidimensional consensus model.

[0094] By performing the above operations, this solution addresses the technical problems of traditional general anesthesia patient anesthesia depth monitoring systems. These problems include the difficulty in balancing the contributions of EEG waveforms, text semantics, and structured indicators, which exhibit high heterogeneity and significant differences in noise distribution, and the lack of effective disagreement negotiation and contingency plans for failed negotiation when different modal evidence points to contradictory conclusions. This results in insufficient model robustness and decision confidence. The solution creatively employs a multidimensional consensus model to output anesthesia depth level prediction results. By constructing three sub-networks combined with a debate consensus module, the sensitivity of each sub-network to specific neurophysiological signals and clinical indicators is preserved. Furthermore, through a debate consensus mechanism, the feature spaces of each modality are aligned, and mutual information is referenced to correct judgments. This improves the reliability and interpretability of anesthesia depth monitoring for complex and atypical cases.

[0095] Example 5, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In the mirror generation and comparison module, it is used to detect whether the multidimensional consensus model produces biased predictions due to the patient's demographic attributes. Specifically, it generates a mirror sample that differs from the original sample only in demographic data by constructing a variational autoencoder. Then, the original sample and the mirror sample are respectively input into the frozen multidimensional consensus model, and the output differences are compared to obtain the final anesthesia depth monitoring reference result.

[0096] The mirror generation and comparison module specifically includes variational autoencoder design, medical rationality constraints, mirror sample generation, and frozen model inference.

[0097] The variational autoencoder design is used to encode multidimensional patient data into separable latent variables of pathophysiological and demographic attributes, including:

[0098] Multimodal coding is used to encode data from three modalities into a joint latent space. Specifically, waveform, text, and numerical features are extracted and concatenated by three independent mini encoders. The resulting joint features are processed by a fully connected layer to obtain the sampling mean and sampling variance. The joint latent variables are obtained by random sampling based on a normal distribution with a dimension of 128. The three independent mini encoders have the same structure as the three subnets but different parameters.

[0099] Latent space decoupling is used to separate latent variables into two parts: pathophysiological attributes and demographic attributes. Specifically, the first 96 dimensions of the joint latent variables are used as the pathophysiological subspace and the last 32 dimensions are used as the demographic subspace. The mutual information between the two is minimized by the CLUB estimator to obtain two statistically independent subspaces.

[0100] Causal intervention designs are used to ensure that demographic attributes influence pathophysiological attributes only through known causal paths. Specifically, a sparse causal matrix is ​​used to add the demographic subspace to the pathophysiological subspace, and L1 regularization is applied to the sparse causal matrix to obtain enhanced joint latent variables. The formula used is as follows:

[0101] ;

[0102] In the formula, This represents the pathophysiological subspace after intervention. Let M denote the pathophysiological subspace, and M denote the sparse causal matrix. The term "demographic subspace" is used to represent the enhanced joint latent variable, which is composed of the post-intervention pathophysiological subspace and the demographic subspace. The term "sparse causal matrix" is a pre-trained causal matrix, where each row corresponds to a dimension of the pathophysiological subspace, and only rows where the known demographic attributes have a causal relationship with the pathophysiological attributes are non-zero.

[0103] The decoder is constructed to ensure that the decoupled latent variables can still completely reconstruct the generated samples. Specifically, the decoder reconstructs waveform, text, and numerical data based on the enhanced joint latent variables to obtain the generated samples. The decoder specifically includes a waveform decoder, a text decoder, and a numerical decoder. The waveform decoder is a transposed convolutional network with mean squared error loss. The text decoder consists of a linear layer and a softmax classifier with cross-entropy loss. The numerical decoder is a multilayer perceptron with mean squared error loss.

[0104] The medical rationality constraint is used to ensure that the generated sample is equivalent to the original sample in terms of pathophysiology and conforms to medical common sense. Specifically, the generated sample is processed by a Gaussian mixture model pre-trained based on the original dataset of historical anesthesia depth monitoring, and a rationality score in the range of [0,1] is output. When the variational autoencoder is trained, the Gaussian mixture model is fixed, and the variational autoencoder forces the rationality score to be close to 1 through adversarial loss.

[0105] The mirror sample generation is used to create a virtual patient sample with different demographic attributes but the same pathophysiological attributes. Specifically, for any original sample, its pathophysiological subspace after intervention is first obtained, then another original sample is randomly selected and its demographic subspace is obtained. The two are combined as mirror joint latent variables, and finally the mirror sample corresponding to the original sample is obtained through the decoder.

[0106] The frozen model inference is used to compare whether the anesthesia depth judgment of the original sample and the mirror sample are consistent. Specifically, the original sample and its corresponding mirror sample are respectively input into the frozen multidimensional consensus model to obtain their respective consensus quality scores and anesthesia depth level labels. If the anesthesia depth level labels obtained by the two are consistent, the consensus quality score and anesthesia depth level label corresponding to the original sample are directly output as the final anesthesia depth monitoring reference result. Otherwise, the consensus quality score and anesthesia depth level label of the two and the mandatory manual re-evaluation alarm are output as the final anesthesia depth monitoring reference result.

[0107] By performing the above operations, this solution addresses the technical problem of traditional general anesthesia patient depth monitoring systems, which often only focus on the overall accuracy of predictions and can only passively examine the distribution or overall performance differences of different groups in the training data. It lacks a means to detect whether the model's decisions are systematically affected by demographic attributes, potentially leading to systematic underestimation or overestimation of anesthesia depth for specific populations and weakening the universality of clinical decisions. This solution creatively employs a mirror-generated comparison method to detect whether multidimensional consensus models produce biased predictions due to demographic attributes. While maintaining the original disease characteristics, it only changes the patient's demographic profile to generate a medically reasonable virtual control sample, examining whether the model's decisions are distorted by sensitive attributes, thus providing better safety assurance for clinical decision-making.

[0108] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0109] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0110] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A system for monitoring the depth of anesthesia in patients undergoing general anesthesia based on medical information processing, characterized in that: The system includes a multi-dimensional data acquisition module, a preliminary optimization processing module, a multi-dimensional consensus model construction module, and a mirror generation and comparison module; The multidimensional data acquisition module obtains the original dataset for anesthesia depth monitoring through data acquisition. The original dataset for anesthesia depth monitoring specifically includes the historical original dataset for anesthesia depth monitoring and the current original dataset for anesthesia depth monitoring. The preliminary optimization processing module employs waveform data processing, text data processing, numerical data processing, data alignment and assembly, and dataset partitioning as preliminary optimization processing methods to obtain the dataset to be monitored, the deep monitoring training set, and the deep monitoring test set. The multidimensional consensus model construction module is used to process different modal data from three independent subnets and reach a consensus depth level through a debate mechanism. Specifically, it is used to construct a waveform processing subnet, a text processing subnet, and a numerical processing subnet, and combine them with the debate consensus module to obtain a multidimensional consensus model. The mirror generation and comparison module is used to detect whether the multidimensional consensus model produces biased predictions due to the patient's demographic attributes. Specifically, it generates a mirror sample that differs from the original sample only in demographic data by constructing a variational autoencoder. The original sample and the mirror sample are then input into the frozen multidimensional consensus model, and the output differences are compared to obtain the final reference result for anesthesia depth monitoring.

2. The anesthesia depth monitoring system for general anesthesia patients based on medical information processing according to claim 1, characterized in that: The multidimensional consensus model construction module specifically includes a waveform processing subnet, a text processing subnet, a numerical processing subnet, a consensus quality calculation module, and a debate consensus module.

3. The anesthesia depth monitoring system for general anesthesia patients based on medical information processing according to claim 2, characterized in that: The waveform processing subnetwork is used to extract features from high-frequency physiological waveforms and provide a preliminary assessment of anesthesia depth. Its content includes: The residual dilated convolutional block design is used to gradually extract local to global temporal patterns from multi-channel waveforms. Specifically, it involves stacking multiple residual dilated convolutional blocks with progressively increasing dilation rates after 1×1 convolution to increase the dimensionality. Each residual dilated convolutional block contains two dilation convolutions and residual connections to obtain waveform temporal features. Waveform-dimensional anesthesia depth prediction is used to compress temporal features into a fixed-length vector and map it to anesthesia depth judgment. Specifically, it obtains waveform hidden features by performing global average pooling on the waveform temporal features in the time dimension, and then outputs waveform-dimensional confidence scores and waveform-dimensional preliminary anesthesia depth level labels through fully connected layers and softmax classifiers respectively. The text processing subnetwork is used to extract semantic features from structured medical record text and provide a preliminary assessment of anesthesia depth. Its content includes: Lightweight BERT encoding is used to transform text sequences into context-rich semantic vectors. Specifically, it processes the input text sequence data through a pre-trained BioBERT-mini model, extracts the output at the [CLS] position, and obtains the text encoding features. The text-dimensional anesthesia depth prediction is used to compress high-dimensional semantic vectors to a uniform dimension and output an anesthesia depth judgment. Specifically, the text encoding features are reduced in dimensionality through a fully connected layer to obtain text hidden layer features. Then, the text dimension confidence score and the text dimension preliminary anesthesia depth level label are output through a fully connected layer and a softmax classifier, respectively. The numerical processing subnet is used to extract numerical features from laboratory values ​​and vital sign statistics and to provide a preliminary judgment of anesthesia depth. Specifically, it reduces the dimensionality of the numerical data layer by layer through three fully connected layers. Each fully connected layer is followed by LeakyReLU activation and random inactivation to output the numerical hidden layer features. Then, it outputs the numerical dimension confidence score and the numerical dimension preliminary anesthesia depth level label through fully connected layers and softmax classifiers respectively. The consensus quality calculation, used to summarize the initial judgment results of the three subnets and determine whether a debate is needed, includes: Majority voting is used to determine the level of anesthesia depth that is agreed upon by the majority in the three subnets. Specifically, it involves counting the number of initial anesthesia depth level labels in the three subnets and taking the one with the most of the same number as the majority anesthesia depth level label. The consensus quality score is used to quantify the degree of consistency among the three subnetworks. Specifically, it calculates the proportion of subnetworks that are consistent with the majority anesthesia depth level label to obtain the consensus quality score. When the consensus quality score is 1, the initial anesthesia depth level labels output by the three subnets are consistent. The initial anesthesia depth level label and the consensus quality score are directly output as the output of the multidimensional consensus model. Otherwise, the content of the debate consensus module continues to be executed. The debate consensus module is used to reach a higher consensus through feature-level interaction when there are disagreements among the three subnets. Its content includes: Interactive space feature projection is used to map the hidden features of different subnets to the same dimensional space for interaction. Specifically, it maps waveform hidden features, text hidden features, and numerical hidden features to the same feature space through a shared fully connected layer to obtain waveform interactive features, text interactive features, and numerical interactive features. Adversarial modality alignment is used to force the features of the three subnets to be modally indistinguishable in the feature space, thereby eliminating modality-specific bias. Specifically, it is achieved by adversarial training of the gradient inversion layer and the modality discriminator to obtain waveform alignment features, text alignment features, and numerical alignment features. Cross-modal attention interaction is used to allow each subnet to adjust its own alignment features based on the features of other subnets. Specifically, each alignment feature is used as a query, and the other two alignment features are used as keys and values ​​in turn. Through multi-head cross-attention processing, updated waveform hidden features, updated text hidden features, and updated numerical hidden features are obtained. The re-voting and final output are used to make a new judgment on the depth of anesthesia based on the features after the interaction and to determine the final output. Specifically, the updated waveform hidden features, updated text hidden features, and updated numerical hidden features are mapped to the same dimension as the hidden features of their corresponding subnets through linear transformation, and the re-judgment confidence score and re-judgment anesthesia depth level label of each subnet are output instead of the original hidden features. The consensus quality score is recalculated based on the re-decision confidence score of each subnet. When the re-decision consensus quality score is 1, that is, the re-decision anesthesia depth level label of each subnet is consistent, the re-decision anesthesia depth level label and the re-decision consensus quality score are output as the output of the multidimensional consensus model. Otherwise, the re-decision anesthesia depth level label and the re-decision consensus quality score of the subnet with the highest re-decision confidence score among the three subnets are output as the output of the multidimensional consensus model.

4. The anesthesia depth monitoring system for general anesthesia patients based on medical information processing according to claim 1, characterized in that: The mirror generation and comparison module specifically includes variational autoencoder design, medical rationality constraints, mirror sample generation, and frozen model inference.

5. The anesthesia depth monitoring system for general anesthesia patients based on medical information processing according to claim 4, characterized in that: The variational autoencoder design is used to encode multidimensional patient data into separable latent variables of pathophysiological and demographic attributes, including: Multimodal coding is used to encode data from three modalities into a unified joint latent space. Specifically, waveform, text, and numerical features are extracted and concatenated by three independent small encoders. The resulting joint features are processed by a fully connected layer to obtain the sampling mean and sampling variance. The joint latent variables are obtained by random sampling based on a normal distribution with a dimension of 128. Latent space decoupling is used to separate latent variables into two parts: pathophysiological attributes and demographic attributes. Specifically, the first 96 dimensions of the joint latent variables are used as the pathophysiological subspace and the last 32 dimensions are used as the demographic subspace. The mutual information between the two is minimized by the CLUB estimator to obtain two statistically independent subspaces. Causal intervention design is used to ensure that demographic attributes affect pathophysiological attributes only through known causal paths. Specifically, it involves adding the demographic subspace to the pathophysiological subspace using a sparse causal matrix and applying L1 regularization to the sparse causal matrix to obtain enhanced joint latent variables. Decoder construction is used to ensure that the decoupled latent variables can still completely reconstruct the generated samples. Specifically, the decoder reconstructs waveform, text and numerical data based on the enhanced joint latent variables to obtain the generated samples. The medical rationality constraint is used to ensure that the generated sample is equivalent to the original sample in terms of pathophysiology and conforms to medical common sense. Specifically, the generated sample is processed by a Gaussian mixture model pre-trained based on the original dataset of historical anesthesia depth monitoring, and a rationality score in the range of [0,1] is output. When the variational autoencoder is trained, the Gaussian mixture model is fixed, and the variational autoencoder forces the rationality score to be close to 1 through adversarial loss. The mirror sample generation is used to create a virtual patient sample with different demographic attributes but the same pathophysiological attributes. Specifically, for any original sample, its pathophysiological subspace after intervention is first obtained, then another original sample is randomly selected and its demographic subspace is obtained. The two are combined as mirror joint latent variables, and finally the mirror sample corresponding to the original sample is obtained through the decoder. The frozen model inference is used to compare whether the anesthesia depth judgment of the original sample and the mirror sample are consistent. Specifically, the original sample and its corresponding mirror sample are respectively input into the frozen multidimensional consensus model to obtain their respective consensus quality scores and anesthesia depth level labels. If the anesthesia depth level labels obtained by the two are consistent, the consensus quality score and anesthesia depth level label corresponding to the original sample are directly output as the final anesthesia depth monitoring reference result. Otherwise, the consensus quality score and anesthesia depth level label of the two and the mandatory manual re-evaluation alarm are output as the final anesthesia depth monitoring reference result.

6. The anesthesia depth monitoring system for general anesthesia patients based on medical information processing according to claim 1, characterized in that: Both the historical anesthesia depth monitoring raw dataset and the current anesthesia depth monitoring raw dataset contain waveform data, text data, and numerical data. The historical anesthesia depth monitoring raw dataset also includes anesthesia depth level label data.

7. The anesthesia depth monitoring system for general anesthesia patients based on medical information processing according to claim 1, characterized in that: In the preliminary optimization processing module, the waveform data processing is used to remove interference in the waveform and unify the time resolution. Specifically, it filters out baseline drift and power frequency noise through a bandpass filter, then resamples all waveform data to 200Hz and truncates a fixed-length window, and performs Z-score normalization to obtain the processed waveform data. The text data processing is used to convert unstructured medical record text into a form that the model can process. Specifically, it involves word segmentation, stop word removal, entity standardization, and length truncation to obtain text sequence data with controllable length. The numerical data processing is used to fill in missing values ​​and eliminate dimensional differences. Specifically, it involves median filling, outlier truncation, min-max scaling, and one-hot encoding of categorical variables to obtain the processed numerical features. The data alignment assembly is used to align different modal data of the same patient in time. Specifically, it involves concatenating processed waveform, text and numerical data into the same sample through patient association and timestamp matching to obtain a feature format that the model can process. The dataset partitioning is used to obtain training data and test data. Specifically, the original historical anesthesia depth monitoring dataset is split to obtain a depth monitoring training set and a depth monitoring test set. The current anesthesia depth monitoring raw dataset is preliminarily optimized through waveform data processing, text data processing, numerical data processing, and data alignment assembly to obtain the dataset to be monitored. The historical anesthesia depth monitoring raw dataset is preliminarily optimized through waveform data processing, text data processing, numerical data processing, data alignment assembly, and dataset partitioning to obtain the depth monitoring training set and the depth monitoring test set.

Images