Artificial intelligence-based anesthetic drug dosage optimization system

Abstract

This invention relates to an artificial intelligence-based anesthetic drug dosage optimization system. The system includes a medical data acquisition module, a preliminary data processing module, a state prediction model construction module, an anesthetic drug dosage optimization module, and a decision execution module. The invention obtains raw data through data acquisition; employs preliminary data processing methods such as spatiotemporal alignment, feature construction, normalization, and dataset segmentation; utilizes an indirect path of predicting the state – optimizing the state – converting the dosage to ensure the physiological rationality of the dosage decision, thus enhancing the interpretability and safety of the entire system; employs a state prediction deep learning model as the state prediction model, simulating the dynamic interactions of complex human physiological systems to achieve accurate prediction of drug effects; and designs a dosage optimization algorithm to optimize anesthetic drug dosage while simultaneously considering multiple clinical objectives such as depth of anesthesia, physiological stability, and drug safety, significantly improving the accuracy of dosage optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent anesthetic drug dosage optimization technology, specifically to an anesthetic drug dosage optimization system based on artificial intelligence. Background Technology

[0002] Optimization of anesthetic drug dosage is a technique that uses artificial intelligence algorithms such as machine learning or deep learning to scientifically estimate and dynamically adjust the dosage of anesthetic drugs by comprehensively considering data such as individual patient characteristics, surgical type, and real-time physiological indicators. Its goal is to minimize and personalize drug use while ensuring the depth and safety of anesthesia, improve the accuracy and safety of surgical anesthesia, reduce the risk of anesthesia-related complications, and optimize the use of drug resources.

[0003] However, traditional anesthetic drug dosage optimization suffers from several technical problems. Firstly, it attempts to establish a simple input-output mapping within a complex physiological response system, neglecting the dynamic transmission process of drug action. This fails to explain the nonlinear relationship between dose and effect, and lacks a deep understanding of the evolution of physiological states, leading to predictive failures when significant individual differences or abnormal physiological responses occur. Secondly, traditional anesthetic drug dosage optimization relies on limited monitoring indicators and fixed dosage formulas, failing to fully reflect the complexity of individual patient physiological characteristics. It ignores the synergistic and antagonistic effects between various physiological systems, making it difficult to accurately predict the pharmacokinetic characteristics of drugs in complex surgical situations, resulting in dosage decisions lacking adaptability to the patient's real-time physiological state. Thirdly, traditional anesthetic drug dosage optimization lacks a systematic optimization mechanism, making it difficult to achieve an optimal balance among multiple conflicting clinical goals, and failing to fully realize the therapeutic effect of drugs while minimizing side effects. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an artificial intelligence-based anesthetic drug dosage optimization system. Traditional anesthetic drug dosage optimization attempts to establish simple input-output mappings within complex physiological response systems, neglecting the dynamic transmission process of drug action, failing to explain the nonlinear relationship between dose and effect, and lacking a deep understanding of the evolution of physiological states. This leads to prediction failures when there are significant individual differences or abnormal physiological responses. This solution creatively adopts an indirect path of predicting the state, optimizing the state, and converting the dose. It bases dosage optimization on a deep understanding of physiological mechanisms, ensuring the physiological rationality of dosage decisions by finding an ideal physiological state as an intermediate bridge, thus giving the entire system better interpretability and safety. Furthermore, traditional anesthetic drug dosage optimization relies on limited monitoring indicators and fixed dosage formulas, failing to fully reflect the complexity of individual patient physiological characteristics, neglecting the synergistic and antagonistic effects between various physiological systems, and making it difficult to accurately predict drug effects. The pharmacokinetic characteristics of complex surgical scenarios lead to a lack of adaptability in dosage decisions to the patient's real-time physiological state. This solution creatively employs a state prediction deep learning model as the state prediction model. By simulating the dynamic interactions of complex human physiological systems, it achieves accurate prediction of drug effects. It can learn the nonlinear relationships between physiological indicators from multi-source heterogeneous data and capture the dynamic response process of drugs in vivo, providing a reliable physiological basis for subsequent dosage optimization. Addressing the technical problem that traditional anesthetic drug dosage optimization lacks a systematic optimization mechanism, making it difficult to achieve the best balance among multiple conflicting clinical goals and fully realize the therapeutic effect of drugs while minimizing side effects, this solution creatively designs a dosage optimization algorithm for anesthetic drug dosage optimization. Through an intelligent search algorithm, it systematically explores the optimal dosage range while ensuring clinical safety, simultaneously considering multiple clinical goals such as anesthesia depth, physiological stability, and drug safety, significantly improving the accuracy of dosage optimization.

[0005] The technical solution adopted by the present invention is as follows: The anesthetic drug dosage optimization system based on artificial intelligence provided by the present invention includes a medical data acquisition module, a preliminary data processing module, a state prediction model construction module, an anesthetic drug dosage optimization module, and a decision execution module;

[0006] The medical data acquisition module obtains a dose-optimized raw dataset through data collection and sends the dose-optimized raw dataset to the data preliminary processing module.

[0007] The data preliminary processing module uses data spatiotemporal alignment, feature construction, normalization processing, and dataset segmentation as preliminary data processing methods to obtain the dataset to be optimized, the state prediction training set, and the state prediction test set. The dataset to be optimized is then sent to the anesthetic drug dosage optimization module, and the state prediction training set and the state prediction test set are sent to the state prediction model construction module.

[0008] The state prediction model building module is used to build the model required to provide a reliable data foundation for subsequent anesthetic drug dosage optimization. It constructs a state prediction deep learning model as the state prediction model and sends the state prediction model to the anesthetic drug dosage optimization module.

[0009] The anesthetic drug dosage optimization module is used to optimize the anesthetic drug dosage based on the predicted comprehensive physiological state, optimize the anesthetic drug dosage by designing a dosage optimization algorithm, obtain the anesthetic drug dosage optimization decision, and send the anesthetic drug dosage optimization decision to the decision execution module.

[0010] The decision execution module specifically sends the anesthetic drug dosage optimization decision to relevant medical staff, and comprehensively considers the anesthetic drug dosage based on this.

[0011] Furthermore, in the medical data acquisition module, the dose optimization raw dataset specifically includes a past dose optimization raw dataset and a real-time dose optimization raw dataset. Both the past dose optimization raw dataset and the real-time dose optimization raw dataset include patient basic information data, physiological monitoring data, drug-related data, surgery-related data, and environmental data. The past dose optimization raw dataset also includes clinical outcome data.

[0012] Furthermore, in the preliminary data processing module, the spatiotemporal alignment of the data is used to ensure the consistency of physiological data from different sources and sampling frequencies in the time dimension. Specifically, it is achieved by using resampling, time delay compensation, and timestamp synchronization techniques to obtain a data sequence with a unified time reference.

[0013] The feature construction is used to extract meaningful feature representations. Specifically, feature engineering is performed on patient basic information data, physiological monitoring data, drug-related data, surgery-related data, and environmental data to obtain feature representations for each data.

[0014] The normalization process is used to normalize the data scale to eliminate the influence of dimensions. Specifically, it processes the data using the Z-score standardization method to obtain data with uniform scale.

[0015] The dataset segmentation is used to obtain training data and test data, specifically by segmenting the original dataset of past dose optimization.

[0016] By performing the data spatiotemporal alignment, feature construction, and normalization, the real-time dose optimization raw dataset is preliminarily processed to obtain the dataset to be optimized. By performing the data spatiotemporal alignment, feature construction, normalization, and dataset segmentation, the past dose optimization raw dataset is preliminarily processed to obtain the state prediction training set and the state prediction test set.

[0017] Furthermore, in the state prediction model construction module, a model is used to construct the model required to provide a reliable data foundation for subsequent anesthetic drug dosage optimization. Specifically, a state prediction deep learning model is constructed as the state prediction model. The state prediction deep learning model outputs a predicted comprehensive physiological state by combining a graph neural network architecture and an adversarial training mechanism.

[0018] The state prediction model construction module specifically includes physiological system graph construction, multi-scale spatiotemporal graph convolution, adversarial physiological state encoding, drug response prediction, and model construction and training.

[0019] The physiological system diagram construction is used to model the human physiological system as a dynamic graph structure to capture the interaction relationships between physiological systems, and includes:

[0020] Physiological system node definition is used to identify key physiological system nodes in the graph structure. Specifically, the cardiovascular system, central nervous system, respiratory system, autonomic nervous system, and metabolic system are selected as key physiological systems to construct nodes.

[0021] The node edge weight calculation is used to quantify the interaction strength between different physiological systems. Specifically, it calculates the connection edge weights between nodes through time-varying mutual information to obtain the adjacency matrix of the physiological systems.

[0022] Multi-scale graph structure construction is used to construct graph structures at different granularity levels to capture multi-scale physiological features. Specifically, multi-scale graph structures are obtained by defining local graph structures and global system graph structures. The local graph structure is specifically a graph structure with the key physiological system as nodes and retaining edges with a weight greater than 0.7. The global system graph structure is specifically a fully connected graph with the key physiological system as nodes.

[0023] Anatomical constraints are applied to ensure that the graph structure conforms to prior knowledge of human anatomy. Specifically, the weights of the connecting edges are constrained by an anatomical connectivity mask obtained from prior anatomical knowledge, resulting in constrained local adjacency matrices and constrained global system adjacency matrices.

[0024] The multi-scale spatiotemporal graph convolution is used to extract spatiotemporal features from the graph structure, and includes:

[0025] Local graph convolution computation is used to extract local interaction features within a physiological system. Specifically, it uses a graph attention mechanism to aggregate information on the local graph structure based on a constrained local adjacency matrix to obtain local physiological features.

[0026] Global graph convolution computation is used to capture the global synergistic effects between physiological systems. Specifically, it involves using graph convolution operations to aggregate information on the global system graph structure based on the constrained global system adjacency matrix to obtain global physiological features.

[0027] Temporal enhancement is used to integrate spatial graph structure and temporal series information. Specifically, it fuses physiological local features at consecutive time steps through a temporal convolutional network to obtain temporal enhancement features.

[0028] Multi-scale feature aggregation is used to adaptively fuse feature information at different scales. Specifically, it combines physiological local features, physiological global features, and temporal enhancement features through a weighted summation mechanism to obtain physiological multi-scale features.

[0029] The adversarial physiological state encoding, used to learn decoupled and stable physiological state representations, includes:

[0030] Dual-channel feature encoding is used to encode physiological signal features and drug effect information separately. Specifically, it processes physiological multi-scale features and drug effect features obtained from drug-related data through two independent fully connected networks to obtain physiological encoding features and drug encoding features.

[0031] State reconstruction generation is used to reconstruct physiological state representation through a generator network. Specifically, a generator network is used to map encoded features back to the physiological state space to obtain the reconstructed physiological state. The generator network is specifically an independent multilayer perceptron.

[0032] Adversarial discriminative training is used to improve the realism of physiological state representation. Specifically, it involves distinguishing between real physiological states and reconstructed physiological states through a discriminator network and training it adversarially with the generator network to obtain a reinforced generator network, which serves as a physiological state encoder. The discriminator network is specifically an independent multilayer perceptron.

[0033] State uncertainty estimation, used to quantify the confidence of state estimation, specifically involves processing encoded features through an independent multilayer perceptron to obtain uncertainty values;

[0034] The drug response prediction is used to establish a model of the relationship between drug dosage and physiological response, and includes:

[0035] Confounding factor identification is used to identify non-drug factors that affect drug efficacy. Specifically, it involves using a convolutional neural network to extract features based on patient basic information data, surgery-related data, and environmental data to obtain confounding factor features.

[0036] Intervention effect estimation is used to estimate the direct intervention effect of a drug. Specifically, it involves designing a model to separate the direct intervention effect from confounding effects. The formula used is as follows:

[0037] ;

[0038] In the formula, This indicates the intervention activation function. This represents the slope of the intervention activation function. This represents the center point parameter of the intervention activation function. This represents the direct intervention effect function of the drug. Let represent the independent multilayer perceptron operating function used to obtain the direct intervention effect of the drug, Se represent the drug sensitivity matrix based on medical prior knowledge, and bs represent the baseline offset vector. Represents the confounding effect function. Let x and represent the independent multilayer perceptron operating functions used to capture the promiscuous effects. This represents distinct input arguments. This indicates the direct intervention effect of the drug. Indicates the confounding effect. This represents the enhanced reconstructed physiological state obtained by the physiological state encoder; De represents the drug effect features obtained based on drug-related data; and conf represents the confounding factor features.

[0039] Comprehensive state prediction is used to obtain a predicted comprehensive physiological state. Specifically, it combines the enhanced reconstructed physiological state obtained by the physiological state encoder with the direct intervention effect of the drug and confounding effects to obtain the predicted comprehensive physiological state.

[0040] The model construction and training specifically involves building a state prediction deep learning model by integrating the physiological system graph construction, the multi-scale spatiotemporal graph convolution, the adversarial physiological state encoding, and the drug response prediction. The model is then trained and its performance is verified based on the state prediction training set and the state prediction test set, resulting in the trained state prediction deep learning model as the state prediction model.

[0041] Furthermore, in the anesthetic drug dosage optimization module, the anesthetic drug dosage is optimized based on the predicted comprehensive physiological state. Specifically, the anesthetic drug dosage is optimized by designing a dosage optimization algorithm to obtain the anesthetic drug dosage optimization decision.

[0042] The anesthetic drug dosage optimization module specifically includes particle initialization, fitness assessment, optimal physiological state search, and dosage conversion.

[0043] The particle initialization is used to initialize and optimize the search process for the current patient, and includes:

[0044] Patient similarity matching is used to find historical cases similar to the current patient. Specifically, the dataset to be optimized is used as the input of the state prediction model to obtain the predicted comprehensive physiological state of the current patient, and the similarity is calculated with the set of historical predicted comprehensive physiological states obtained based on the training set. The top K historical predicted comprehensive physiological states with the lowest similarity are selected as the initial positions of the particles to obtain the preliminary particle swarm.

[0045] The safe search space is defined to ensure that the optimization process is carried out within the clinically safe range. Specifically, it is obtained by combining the uncertainty value obtained by using the dataset to be optimized as the input of the state prediction model with the predicted comprehensive physiological state of the current patient.

[0046] Diverse particle initialization is used to ensure that the optimization algorithm has sufficient exploration capabilities. Specifically, particles outside the safe search space in the initial particle swarm are removed, and the number of particles is replenished to I by random sampling within the safe search space to obtain the initial particle swarm.

[0047] The fitness assessment, used to comprehensively evaluate the fitness of different particles, includes:

[0048] Multi-objective fitness calculation is used to evaluate the fitness of particles from multiple dimensions, specifically from three aspects: anesthesia depth, stability, and safety.

[0049] The overall fitness is obtained by adjusting the relative importance of each objective according to the surgical stage. Specifically, the fitness of multiple objectives is calculated by weighting the fitness based on the medical prior knowledge of the surgical stage to obtain the overall fitness value.

[0050] The optimal physiological state search is used to search for the optimal physiological state that the current patient can achieve, and includes:

[0051] A hybrid update strategy is designed to balance global exploration and local development capabilities. Specifically, it combines standard particle swarm update and gradient-guided search to obtain the updated particle positions.

[0052] Constraint handling and repair are used to ensure that the search process meets clinical safety constraints. Specifically, feasible solution space exploration is obtained through projection operators and penalty function mechanisms.

[0053] To obtain the optimal physiological state, specifically, iterative updates are continuously performed until the iteration termination condition is met, then the iteration updates are stopped and the global best position is obtained as the current patient's optimal physiological state. The iteration termination condition specifically includes the particle's comprehensive fitness value being continuously greater than a preset threshold and the maximum number of iterations being reached.

[0054] The dose conversion is used to convert the current optimal physiological state of the patient into a specific anesthetic drug dose. Specifically, it uses a pre-trained multilayer perceptron to convert the current optimal physiological state of the patient into an optimal anesthetic drug dose, and uses the optimal anesthetic drug dose as the output of the anesthetic drug dose optimization decision.

[0055] Furthermore, in the decision execution module, the optimized anesthetic drug dosage decision is sent to relevant medical personnel, and the dosage of anesthetic drugs is comprehensively considered based on the optimized anesthetic drug dosage decision and combined with medical experience.

[0056] The beneficial effects achieved by the present invention using the above solution are as follows:

[0057] (1) Traditional anesthetic drug dosage optimization attempts to establish a simple input-output mapping in a complex physiological response system, ignoring the dynamic transmission process of drug action, failing to explain the nonlinear relationship between dose and effect, and lacking a deep understanding of the evolution of physiological state, resulting in prediction failure when there are significant individual differences or abnormal physiological reactions. This solution creatively adopts an indirect path of predicting state - optimizing state - changing dose, establishing dosage optimization on a deep understanding of physiological mechanisms. By finding the ideal physiological state as an intermediate bridge, the physiological rationality of dosage decision is ensured, making the whole system more interpretable and safe.

[0058] (2) Traditional anesthetic drug dosage optimization relies on limited monitoring indicators and fixed dosage formulas, which cannot fully reflect the complexity of individual patient physiological characteristics, ignore the synergistic and antagonistic effects between various physiological systems, and make it difficult to accurately predict the pharmacokinetic characteristics of drugs in complex surgical situations. This results in a lack of adaptability of dosage decisions to the real-time physiological state of patients. This solution creatively adopts a state prediction deep learning model as the state prediction model. By simulating the dynamic interaction of complex human physiological systems, it can accurately predict drug effects. It can learn the nonlinear relationship between physiological indicators from multi-source heterogeneous data, capture the dynamic response process of drugs in the body, and provide a reliable physiological basis for subsequent dosage optimization.

[0059] (3) In view of the technical problems of traditional anesthetic drug dosage optimization, which lacks a systematic optimization mechanism, makes it difficult to achieve the best balance among multiple conflicting clinical goals, and fails to give full play to the therapeutic effect of the drug while minimizing side effects, this solution creatively designs a dosage optimization algorithm to optimize the dosage of anesthetic drugs. Through intelligent search algorithm, the optimal dosage range is systematically explored under the premise of ensuring clinical safety. It can simultaneously take into account multiple clinical goals such as anesthesia depth, physiological stability and drug safety, and significantly improve the accuracy of dosage optimization. Attached Figure Description

[0060] Figure 1 A schematic diagram of the modules of the artificial intelligence-based anesthetic drug dosage optimization system provided by the present invention;

[0061] Figure 2 This is a flowchart illustrating the preliminary data processing module.

[0062] Figure 3 A flowchart illustrating the process of building modules for the state prediction model;

[0063] Figure 4 A flowchart illustrating the anesthetic drug dosage optimization module.

[0064] 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

[0065] 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.

[0066] 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.

[0067] Example 1, see Figure 1 The artificial intelligence-based anesthetic drug dosage optimization system provided by the present invention includes a medical data acquisition module, a preliminary data processing module, a state prediction model construction module, an anesthetic drug dosage optimization module, and a decision execution module.

[0068] The medical data acquisition module obtains a dose-optimized raw dataset through data collection and sends the dose-optimized raw dataset to the data preliminary processing module.

[0069] The data preliminary processing module uses data spatiotemporal alignment, feature construction, normalization processing, and dataset segmentation as preliminary data processing methods to obtain the dataset to be optimized, the state prediction training set, and the state prediction test set. The dataset to be optimized is then sent to the anesthetic drug dosage optimization module, and the state prediction training set and the state prediction test set are sent to the state prediction model construction module.

[0070] The state prediction model building module is used to build the model required to provide a reliable data foundation for subsequent anesthetic drug dosage optimization. It constructs a state prediction deep learning model as the state prediction model and sends the state prediction model to the anesthetic drug dosage optimization module.

[0071] The anesthetic drug dosage optimization module is used to optimize the anesthetic drug dosage based on the predicted comprehensive physiological state, optimize the anesthetic drug dosage by designing a dosage optimization algorithm, obtain the anesthetic drug dosage optimization decision, and send the anesthetic drug dosage optimization decision to the decision execution module.

[0072] The decision execution module specifically sends the anesthetic drug dosage optimization decision to relevant medical staff, and comprehensively considers the anesthetic drug dosage based on this.

[0073] By performing the above operations, this solution addresses the technical problems of traditional anesthetic drug dosage optimization, which attempts to establish a simple input-output mapping in a complex physiological response system, ignores the dynamic transmission process of drug action, cannot explain the nonlinear relationship between dose and effect, and lacks a deep understanding of the evolution of physiological states, leading to prediction failure when there are significant individual differences or abnormal physiological reactions. This solution creatively adopts an indirect path of predicting the state - optimizing the state - converting the dose, establishing dosage optimization on a deep understanding of physiological mechanisms. By finding the ideal physiological state as an intermediate bridge, it ensures the physiological rationality of dosage decisions, making the entire system more interpretable and safer.

[0074] Example 2, see Figure 1Based on the above embodiments, in the medical data acquisition module, the dose optimization raw dataset specifically includes a past dose optimization raw dataset and a real-time dose optimization raw dataset. Both the past dose optimization raw dataset and the real-time dose optimization raw dataset include patient basic information data, physiological monitoring data, drug-related data, surgery-related data, and environmental data. The past dose optimization raw dataset also includes clinical outcome data.

[0075] The patient's basic information data specifically includes patient age data, patient gender data, patient height and weight data, patient medical history data, patient long-term medication history data, patient drug allergy history data, and patient laboratory test data;

[0076] The physiological monitoring data specifically includes cardiovascular system monitoring data, central nervous system monitoring data, respiratory system monitoring data, autonomic nervous system monitoring data, and metabolic system monitoring data. The cardiovascular system monitoring data specifically includes electrocardiogram (ECG) data, blood pressure monitoring data, and cardiac output monitoring data. The central nervous system monitoring data specifically includes bispectral index (BSE) data, EEG data, and evoked potential monitoring data. The respiratory system monitoring data specifically includes blood oxygen saturation data, end-tidal carbon dioxide data, respiratory rate data, and airway pressure data. The autonomic nervous system monitoring data specifically includes skin conductance level data, pupil diameter change rate data, baroreflex sensitivity index data, and thermoregulatory response index data. The metabolic system monitoring data specifically includes core body temperature change data, surface-core temperature gradient data, metabolic equivalent estimate data, and drug clearance rate index data.

[0077] The drug-related data specifically refers to the time, dosage, and rate of anesthetic drug infusion for the patient.

[0078] The surgery-related data specifically includes surgery type data, surgery stage data, and surgical stimulation intensity score data;

[0079] The environmental data specifically includes indoor temperature and humidity data, noise level data, and light intensity data;

[0080] The clinical outcome data specifically refers to anesthesia quality label data.

[0081] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In the data preliminary processing module, the data spatiotemporal alignment is used to ensure the consistency of physiological data from different sources and sampling frequencies in the time dimension. Specifically, it obtains a data sequence with a unified time reference through resampling, time delay compensation and timestamp synchronization technology.

[0082] The feature construction is used to extract meaningful feature representations. Specifically, feature engineering is performed on patient basic information data, physiological monitoring data, drug-related data, surgery-related data, and environmental data to obtain feature representations for each data.

[0083] The normalization process is used to normalize the data scale to eliminate the influence of dimensions. Specifically, it processes the data using the Z-score standardization method to obtain data with uniform scale.

[0084] The dataset segmentation is used to obtain training data and test data, specifically by segmenting the original dataset of past dose optimization.

[0085] By performing the data spatiotemporal alignment, feature construction, and normalization, the real-time dose optimization raw dataset is preliminarily processed to obtain the dataset to be optimized. By performing the data spatiotemporal alignment, feature construction, normalization, and dataset segmentation, the past dose optimization raw dataset is preliminarily processed to obtain the state prediction training set and the state prediction test set.

[0086] Example 4, see Figure 1 and Figure 3 This embodiment is based on the above embodiment. In the state prediction model construction module, a model is used to construct the model required to provide a reliable data foundation for subsequent anesthetic drug dosage optimization. Specifically, a state prediction deep learning model is constructed as the state prediction model. The state prediction deep learning model outputs a predicted comprehensive physiological state by combining a graph neural network architecture and an adversarial training mechanism.

[0087] The state prediction model construction module specifically includes physiological system graph construction, multi-scale spatiotemporal graph convolution, adversarial physiological state encoding, drug response prediction, and model construction and training.

[0088] The physiological system diagram construction is used to model the human physiological system as a dynamic graph structure to capture the interaction relationships between physiological systems, and includes:

[0089] Physiological system node definition is used to identify key physiological system nodes in the graph structure. Specifically, the cardiovascular system, central nervous system, respiratory system, autonomic nervous system, and metabolic system are selected as key physiological systems to construct nodes.

[0090] The node edge weight calculation is used to quantify the interaction strength between different physiological systems. Specifically, it calculates the edge weights between nodes using time-varying mutual information to obtain the adjacency matrix of the physiological systems. The formula used is as follows:

[0091] ;

[0092] In the formula, This represents the weight of the edge connecting node a and node b. This represents the mutual information calculation function, where t represents the time step. Indicates the size of the time window. Indicates time window The node characteristics of internal node a, Indicates time window The node characteristics of internal node b. Represents the mean of mutual information. Indicates the standard deviation of mutual information;

[0093] Multi-scale graph structure construction is used to construct graph structures at different granularity levels to capture multi-scale physiological features. Specifically, multi-scale graph structures are obtained by defining local graph structures and global system graph structures. The local graph structure is specifically a graph structure with the key physiological system as nodes and retaining edges with a weight greater than 0.7. The global system graph structure is specifically a fully connected graph with the key physiological system as nodes.

[0094] Anatomical constraints are applied to ensure that the graph structure conforms to prior knowledge of human anatomy. Specifically, the weights of connecting edges are constrained using an anatomical connectivity mask obtained from prior anatomical knowledge, resulting in constrained local adjacency matrices and constrained global system adjacency matrices. The formulas used are as follows:

[0095] ;

[0096] In the formula, Represents the constrained local adjacency matrix. CM represents the adjacency matrix of the constrained global system, and CM represents the dissecting connectivity mask. Represents the local adjacency matrix. Represents the global system adjacency matrix. This represents element-wise multiplication.

[0097] The multi-scale spatiotemporal graph convolution is used to extract spatiotemporal features from the graph structure, and includes:

[0098] Local graph convolution computation is used to extract local interaction features within a physiological system. Specifically, it uses a graph attention mechanism to aggregate information on the local graph structure based on a constrained local adjacency matrix to obtain local physiological features. The formula used is as follows:

[0099] ;

[0100] In the formula, Indicates local physiological characteristics, This represents the graph attention convolution function, where nf represents the node features;

[0101] Global graph convolution computation is used to capture global synergistic effects between physiological systems. Specifically, it involves using graph convolution operations to aggregate information on the global system graph structure based on the constrained global system adjacency matrix to obtain global physiological features. The formula used is as follows:

[0102] ;

[0103] In the formula, Indicates global physiological characteristics. Represents the graph convolution function;

[0104] Temporal enhancement is used to integrate spatial graph structure and temporal series information. Specifically, it fuses physiological local features at consecutive time steps using a temporal convolutional network to obtain temporally enhanced features. The formula used is as follows:

[0105] ;

[0106] In the formula, Indicates temporal enhancement features, This represents the function that runs a temporal convolutional network. Indicates in Physiological local characteristics of time step This represents the local physiological characteristics at time step t. Indicates a splicing operation;

[0107] Multi-scale feature aggregation is used to adaptively fuse feature information at different scales. Specifically, it combines physiological local features, physiological global features, and temporal enhancement features through a weighted summation mechanism to obtain physiological multi-scale features. The formula used is as follows:

[0108] ;

[0109] In the formula, Indicates multi-scale weights, This represents the softmax activation function. This indicates that multi-scale weights generate learnable weights. This represents the pooling operation function. Represents physiological multiscale characteristics, Represents the weight of physiological local scales. Represents the physiological global scale weight. Indicates the temporal augmentation scale weight;

[0110] The adversarial physiological state encoding, used to learn decoupled and stable physiological state representations, includes:

[0111] Dual-channel feature encoding is used to encode physiological signal features and drug effect information separately. Specifically, it processes physiological multi-scale features and drug effect features obtained from drug-related data through two independent fully connected networks to obtain physiological encoding features and drug encoding features.

[0112] State reconstruction generation is used to reconstruct physiological state representations through a generator network. Specifically, it employs a generator network to map encoded features back to the physiological state space to obtain the reconstructed physiological state. The generator network is specifically an independent multilayer perceptron. The formula used for state reconstruction generation is as follows:

[0113] ;

[0114] In the formula, This indicates a reconstructed physiological state. This represents the generator network execution function. Represents physiological coding features, Indicates drug coding characteristics;

[0115] Adversarial discriminative training is used to improve the realism of physiological state representation. Specifically, it involves distinguishing between real physiological states and reconstructed physiological states through a discriminator network and training it adversarially with the generator network to obtain a reinforced generator network, which serves as a physiological state encoder. The discriminator network is specifically an independent multilayer perceptron.

[0116] State uncertainty estimation, used to quantify the confidence level of state estimation, specifically involves processing the encoded features through an independent multilayer perceptron to obtain the uncertainty value, using the following formula:

[0117] ;

[0118] In the formula, Indicates the value of uncertainty. This represents the multilayer perceptron operating function used for state uncertainty estimation;

[0119] The drug response prediction is used to establish a model of the relationship between drug dosage and physiological response, and includes:

[0120] Confounding factor identification is used to identify non-drug factors that affect drug efficacy. Specifically, it involves using a convolutional neural network to extract features based on patient basic information data, surgery-related data, and environmental data to obtain confounding factor features.

[0121] Intervention effect estimation is used to estimate the direct intervention effect of a drug. Specifically, it involves designing a model to separate the direct intervention effect from confounding effects. The formula used is as follows:

[0122] ;

[0123] In the formula, This indicates the intervention activation function. This represents the slope of the intervention activation function. This represents the center point parameter of the intervention activation function. This represents the direct intervention effect function of the drug. Let represent the independent multilayer perceptron operating function used to obtain the direct intervention effect of the drug, Se represent the drug sensitivity matrix based on medical prior knowledge, and bs represent the baseline offset vector. Represents the confounding effect function. Let x and represent the independent multilayer perceptron operating functions used to capture the promiscuous effects. This represents distinct input arguments. This indicates the direct intervention effect of the drug. Indicates the confounding effect. This represents the enhanced reconstructed physiological state obtained by the physiological state encoder; De represents the drug effect features obtained based on drug-related data; and conf represents the confounding factor features.

[0124] Comprehensive state prediction is used to obtain a predicted comprehensive physiological state. Specifically, it combines the enhanced reconstructed physiological state obtained by the physiological state encoder with the direct intervention effect of the drug and confounding effects to obtain the predicted comprehensive physiological state. The formula used is as follows:

[0125] ;

[0126] In the formula, This indicates a prediction of the overall physiological state at the next moment. This indicates that the physiological state is being reinforced and reconstructed at the current moment. This indicates the direct intervention effect of the drug at the current moment. This indicates the confounding effect at the current moment;

[0127] The model construction and training specifically involves building a state prediction deep learning model by integrating the physiological system graph construction, the multi-scale spatiotemporal graph convolution, the adversarial physiological state encoding, and the drug response prediction. The model is then trained and its performance is verified based on the state prediction training set and the state prediction test set, resulting in the trained state prediction deep learning model as the state prediction model.

[0128] By performing the above operations, we can address the technical problems of traditional anesthetic drug dosage optimization, which relies on limited monitoring indicators and fixed dosage formulas, fails to fully reflect the complexity of individual patient physiological characteristics, ignores the synergistic and antagonistic effects between various physiological systems, and makes it difficult to accurately predict the pharmacokinetic characteristics of drugs in complex surgical situations. This results in a lack of adaptability of dosage decisions to the real-time physiological state of patients. This solution creatively adopts a state prediction deep learning model as the state prediction model. By simulating the dynamic interaction of complex human physiological systems, it achieves accurate prediction of drug effects. It can learn the nonlinear relationships between physiological indicators from multi-source heterogeneous data, capture the dynamic response process of drugs in vivo, and provide a reliable physiological basis for subsequent dosage optimization.

[0129] Example 5, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In the anesthetic drug dosage optimization module, it is used to optimize the anesthetic drug dosage based on the predicted comprehensive physiological state. Specifically, it optimizes the anesthetic drug dosage by designing a dosage optimization algorithm to obtain the anesthetic drug dosage optimization decision.

[0130] The anesthetic drug dosage optimization module specifically includes particle initialization, fitness assessment, optimal physiological state search, and dosage conversion.

[0131] The particle initialization is used to initialize and optimize the search process for the current patient, and includes:

[0132] Patient similarity matching is used to find historical cases similar to the current patient. Specifically, the dataset to be optimized is used as input to the state prediction model to obtain the predicted comprehensive physiological state of the current patient, and the similarity is calculated with the set of historical predicted comprehensive physiological states obtained based on the training set. The top K historical predicted comprehensive physiological states in descending order of similarity are selected as the initial positions of the particles to obtain a preliminary particle swarm. The formula used for similarity calculation is as follows:

[0133] ;

[0134] In the formula, Sim represents the similarity between the current predicted comprehensive physiological state and the historical predicted comprehensive physiological state of the patient. This represents the function for calculating cosine similarity. This indicates the patient's current predicted overall physiological state. This indicates the overall physiological state predicted from historical data.

[0135] The safe search space is defined to ensure that the optimization process is conducted within clinically safe limits. Specifically, it is obtained by combining the uncertainty value obtained by using the dataset to be optimized as input to the state prediction model with the predicted comprehensive physiological state of the current patient. The formula used is as follows:

[0136] ;

[0137] In the formula, SP represents the safe search space. This represents the uncertainty value obtained by using the dataset to be optimized as input to the state prediction model;

[0138] Diverse particle initialization is used to ensure that the optimization algorithm has sufficient exploration capabilities. Specifically, particles outside the safe search space in the initial particle swarm are removed, and the number of particles is replenished to I by random sampling within the safe search space to obtain the initial particle swarm.

[0139] The fitness assessment, used to comprehensively evaluate the fitness of different particles, includes:

[0140] Multi-objective fitness calculation is used to evaluate the fitness of particles from multiple dimensions, specifically from three aspects: anesthesia depth, stability, and safety. The formula used is as follows:

[0141] ;

[0142] In the formula, This indicates the anesthesia depth fitness value. Indicates the stability fitness value. Indicates the safety fitness value. This represents the sigmoid activation function. This represents the pre-trained fully connected layer function. This represents the ideal depth of anesthesia. This represents the function for calculating standard deviation. This represents the stability reference threshold, and D represents the number of dimensions for predicting the overall physiological state. This represents the lower bound of the safe range for predicting the comprehensive physiological state dimension d. This represents the upper bound of the safe range for predicting the comprehensive physiological state dimension d. The d-th dimension represents the prediction of overall physiological state. This indicates an indicator function, which has a value of 1 when the predicted d-th dimension of the overall physiological state is within a safe range, and a value of 0 otherwise.

[0143] The overall fitness is obtained by adjusting the relative importance of each objective based on the surgical stage. Specifically, it is calculated by weighting the fitness of multiple objectives based on fitness weights derived from prior medical knowledge of the surgical stage, and obtaining the overall fitness value using the following formula:

[0144] ;

[0145] In the formula, This represents the overall fitness value. The fitness weight of objective c is represented by the prior medical knowledge of the surgical phase. This represents the fitness value of target c;

[0146] The optimal physiological state search is used to search for the optimal physiological state that the current patient can achieve, and includes:

[0147] A hybrid update strategy is designed to balance global exploration and local exploitation capabilities. Specifically, it combines standard particle swarm optimization and gradient-guided search to obtain the updated particle positions, using the following formula:

[0148] ;

[0149] In the formula, This represents the velocity of the i-th particle in the d-th dimension at the (dt+1)-th iteration. Indicates inertia weight, This represents the velocity of the i-th particle in the d-th dimension at the d-th iteration. Represents individual learning factors. Represents the global learning factor. and This represents distinct random values ​​within the range [0,1]. This represents the historical best position of the i-th particle in dimension d. This represents the position of the i-th particle in the d-th dimension at the d-th iteration. This represents the global optimal position of the particle in dimension d. This represents the position of the i-th particle in the d-th dimension at the (dt+1)-th iteration. This represents the gradient-guided search learning rate. This represents the gradient of fitness with respect to particle position;

[0150] Constraint handling and repair are used to ensure that the search process meets clinical safety constraints. Specifically, feasible solution space exploration is obtained through projection operators and penalty function mechanisms, and the formulas used are as follows:

[0151] ;

[0152] In the formula, This represents the position of the i-th out-of-bounds particle after the d-th dimension has been repaired. This represents the projection operator that projects the particle position onto the safe search space. This represents the position of the i-th particle in the d-th dimension. This represents the fitness value after adjusting for constraint violation penalties. The penalty coefficient is represented by vio, and the constraint violation amount is represented by vio.

[0153] To obtain the optimal physiological state, specifically, iterative updates are continuously performed until the iteration termination condition is met, then the iteration updates are stopped and the global best position is obtained as the current patient's optimal physiological state. The iteration termination condition specifically includes the particle's comprehensive fitness value being continuously greater than a preset threshold and the maximum number of iterations being reached.

[0154] The dose conversion is used to convert the optimized physiological state of the current patient into a specific anesthetic drug dose. Specifically, it involves using a pre-trained multilayer perceptron to convert the current patient's optimal physiological state into an optimal anesthetic drug dose, and then using the optimal anesthetic drug dose as the output of the anesthetic drug dose optimization decision. The formula used for the dose conversion is as follows:

[0155] ;

[0156] In the formula, This indicates the optimal dosage of anesthetic drugs. This represents the pre-trained multilayer perceptron running function used for dose conversion. Sf represents the optimal physiological state, and Sf represents the current physiological state of the patient.

[0157] By performing the above operations, this solution addresses the technical problems of traditional anesthetic drug dosage optimization, which lacks a systematic optimization mechanism, makes it difficult to achieve the best balance among multiple conflicting clinical goals, and fails to fully exert the therapeutic effect of the drug while minimizing side effects. This solution creatively designs a dosage optimization algorithm for anesthetic drug dosage optimization. Through an intelligent search algorithm, it systematically explores the optimal dosage range while ensuring clinical safety. It can simultaneously take into account multiple clinical goals such as anesthesia depth, physiological stability, and drug safety, significantly improving the accuracy of dosage optimization.

[0158] Example 6, see Figure 1 This embodiment is based on the above embodiment. In the decision execution module, the anesthetic drug dosage optimization decision is sent to relevant medical staff. Based on the anesthetic drug dosage optimization decision, the dosage of anesthetic drugs is comprehensively considered in combination with medical experience.

[0159] 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.

[0160] 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.

[0161] 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. An artificial intelligence-based anesthetic drug dosage optimization system, characterized in that: The system includes a medical data acquisition module, a preliminary data processing module, a state prediction model construction module, an anesthetic drug dosage optimization module, and a decision execution module; The medical data acquisition module obtains a dose optimization raw dataset through data collection. The dose optimization raw dataset specifically includes a past dose optimization raw dataset and a real-time dose optimization raw dataset. The preliminary data processing module employs data spatiotemporal alignment, feature construction, normalization, and dataset segmentation to obtain the dataset to be optimized, the state prediction training set, and the state prediction test set. The state prediction model construction module is used to build the model required to provide a reliable data foundation for subsequent anesthetic drug dosage optimization. Specifically, it constructs a state prediction deep learning model as the state prediction model. The state prediction deep learning model outputs a prediction of the comprehensive physiological state by combining a graph neural network architecture and an adversarial training mechanism. The specific content includes physiological system graph construction, multi-scale spatiotemporal graph convolution, adversarial physiological state encoding, drug response prediction, and model construction and training. The anesthetic drug dosage optimization module is used to optimize the anesthetic drug dosage based on the predicted comprehensive physiological state. Specifically, it optimizes the anesthetic drug dosage by designing a dosage optimization algorithm to obtain the anesthetic drug dosage optimization decision. The decision execution module specifically sends the anesthetic drug dosage optimization decision to relevant medical staff, and based on the anesthetic drug dosage optimization decision, comprehensively considers the dosage of anesthetic drugs in combination with medical experience. The physiological system graph construction is used to model the human physiological system as a dynamic graph structure to capture the interaction relationships between physiological systems. It includes defining physiological system nodes, calculating node edge weights, constructing a multi-scale graph structure, and applying anatomical constraints to obtain a multi-scale graph structure, a constrained local adjacency matrix, and a constrained global system adjacency matrix, which serve as inputs for multi-scale spatiotemporal graph convolution. The multi-scale spatiotemporal graph convolution is used to extract spatiotemporal features from the graph structure. It includes local graph convolution calculation, global graph convolution calculation, temporal enhancement and multi-scale feature aggregation to obtain physiological multi-scale features, which serve as input for adversarial physiological state encoding. The adversarial physiological state encoding is used to learn a decoupled and stable physiological state representation. It includes dual-channel feature encoding, state reconstruction generation, adversarial discrimination training, and state uncertainty estimation to obtain the reconstructed physiological state and uncertainty value, which serve as inputs for drug response prediction. The drug response prediction is used to establish a model of the relationship between drug dosage and physiological response. It includes the identification of confounding factors, estimation of intervention effects, and prediction of the overall state, resulting in a predicted overall physiological state, which serves as the input to the anesthetic drug dosage optimization module.

2. The artificial intelligence-based anesthetic drug dosage optimization system according to claim 1, characterized in that: The physiological system node definition is used to determine the key physiological system nodes in the graph structure. Specifically, the cardiovascular system, central nervous system, respiratory system, autonomic nervous system, and metabolic system are selected as key physiological systems to construct nodes. The node edge weight calculation is used to quantify the interaction strength between different physiological systems. Specifically, it calculates the connection edge weights between nodes through time-varying mutual information to obtain the adjacency matrix of the physiological systems. The multi-scale graph structure construction is used to construct graph structures at different granularity levels to capture multi-scale physiological features. Specifically, it obtains the multi-scale graph structure by defining local graph structures and global system graph structures. The local graph structure is specifically a graph structure with the key physiological system as nodes and retaining edges with a weight greater than 0.

7. The global system graph structure is specifically a fully connected graph with the key physiological system as nodes. The anatomical constraints are applied to ensure that the graph structure conforms to prior knowledge of human anatomy. Specifically, the weights of the connecting edges are constrained by an anatomical connectivity mask obtained from prior anatomical knowledge, resulting in a constrained local adjacency matrix and a constrained global system adjacency matrix. The local graph convolution computation is used to extract local interaction features within the physiological system. Specifically, it uses a graph attention mechanism to aggregate information on the local graph structure based on a constrained local adjacency matrix to obtain physiological local features. The global graph convolution computation is used to capture the global synergistic effect between physiological systems. Specifically, it involves performing information aggregation on the global system graph structure based on the constrained global system adjacency matrix through graph convolution operations to obtain global physiological features. The temporal enhancement is used to integrate spatial graph structure and temporal sequence information. Specifically, it fuses the physiological local features of consecutive time steps through a temporal convolutional network to obtain temporal enhancement features. The multi-scale feature aggregation is used to adaptively fuse feature information at different scales. Specifically, it combines physiological local features, physiological global features, and temporal enhancement features through a weighted summation mechanism to obtain physiological multi-scale features. The dual-channel feature encoding is used to encode physiological signal features and drug effect information respectively. Specifically, it processes physiological multi-scale features and drug effect features obtained from drug-related data through two independent fully connected networks to obtain physiological encoding features and drug encoding features. The state reconstruction generation is used to reconstruct the physiological state representation through a generator network. Specifically, the generator network is used to map the encoded features back to the physiological state space to obtain the reconstructed physiological state. The generator network is specifically an independent multilayer perceptron. The adversarial discriminant training is used to improve the realism of physiological state representation. Specifically, it distinguishes between real physiological states and reconstructed physiological states through a discriminator network and performs adversarial training with the generator network to obtain a reinforced generator network, which serves as a physiological state encoder. The discriminator network is specifically an independent multilayer perceptron. The state uncertainty estimation is used to quantify the confidence level of the state estimation. Specifically, it is obtained by processing the encoded features through an independent multilayer perceptron to obtain the uncertainty value. The confounding factor identification is used to identify non-drug factors that affect drug effects. Specifically, it involves extracting features based on patient basic information data, surgery-related data, and environmental data using a convolutional neural network to obtain confounding factor features. The intervention effect estimation is used to estimate the direct intervention effect of the drug. Specifically, it involves designing a model to separate the direct intervention effect and confounding effects of the drug. The formula used is as follows: ； In the formula, This indicates the intervention activation function. This represents the slope of the intervention activation function. This represents the center point parameter of the intervention activation function. This represents the direct intervention effect function of the drug. Let represent the independent multilayer perceptron operating function used to obtain the direct intervention effect of the drug, Se represent the drug sensitivity matrix based on medical prior knowledge, and bs represent the baseline offset vector. Represents the confounding effect function. Let x and represent the independent multilayer perceptron operating functions used to capture the promiscuous effects. This represents distinct input arguments. This indicates the direct intervention effect of the drug. Indicates the confounding effect. This represents the enhanced reconstructed physiological state obtained by the physiological state encoder; De represents the drug effect features obtained based on drug-related data; and conf represents the confounding factor features. The comprehensive state prediction is used to obtain a predicted comprehensive physiological state, specifically by combining the reconstructed physiological state obtained by the physiological state encoder with the direct intervention effect of the drug and confounding effects to obtain the predicted comprehensive physiological state. The model construction and training specifically involves building a state prediction deep learning model by integrating the physiological system graph construction, the multi-scale spatiotemporal graph convolution, the adversarial physiological state encoding, and the drug response prediction. The model is then trained and its performance is verified based on the state prediction training set and the state prediction test set, resulting in the trained state prediction deep learning model as the state prediction model.

3. The artificial intelligence-based anesthetic drug dosage optimization system according to claim 1, characterized in that: The anesthetic drug dosage optimization module specifically includes particle initialization, fitness assessment, optimal physiological state search, and dosage conversion.

4. The artificial intelligence-based anesthetic drug dosage optimization system according to claim 3, characterized in that: The particle initialization is used to initialize and optimize the search process for the current patient, and includes: Patient similarity matching is used to find historical cases similar to the current patient. Specifically, the dataset to be optimized is used as the input of the state prediction model to obtain the predicted comprehensive physiological state of the current patient, and the similarity is calculated with the set of historical predicted comprehensive physiological states obtained based on the training set. The top K historical predicted comprehensive physiological states with the lowest similarity are selected as the initial positions of the particles to obtain the preliminary particle swarm. The safe search space is defined to ensure that the optimization process is carried out within the clinically safe range. Specifically, it is obtained by combining the uncertainty value obtained by using the dataset to be optimized as the input of the state prediction model with the predicted comprehensive physiological state of the current patient. Diverse particle initialization is used to ensure that the optimization algorithm has sufficient exploration capabilities. Specifically, particles outside the safe search space in the initial particle swarm are removed, and the number of particles is replenished to I by random sampling within the safe search space to obtain the initial particle swarm. The fitness assessment, used to comprehensively evaluate the fitness of different particles, includes: Multi-objective fitness calculation is used to evaluate the fitness of particles from multiple dimensions, specifically from three aspects: anesthesia depth, stability, and safety. The overall fitness is obtained by adjusting the relative importance of each objective according to the surgical stage. Specifically, the fitness of multiple objectives is calculated by weighting the fitness based on the medical prior knowledge of the surgical stage to obtain the overall fitness value. The optimal physiological state search is used to search for the optimal physiological state that the current patient can achieve, and includes: A hybrid update strategy is designed to balance global exploration and local development capabilities. Specifically, it combines standard particle swarm update and gradient-guided search to obtain the updated particle positions. Constraint handling and repair are used to ensure that the search process meets clinical safety constraints. Specifically, feasible solution space exploration is obtained through projection operators and penalty function mechanisms. To obtain the optimal physiological state, specifically, iterative updates are continuously performed until the iteration termination condition is met, then the iteration updates are stopped and the global best position is obtained as the current patient's optimal physiological state. The iteration termination condition specifically includes the particle's comprehensive fitness value being continuously greater than a preset threshold and the maximum number of iterations being reached. The dose conversion is used to convert the current optimal physiological state of the patient into a specific anesthetic drug dose. Specifically, it uses a pre-trained multilayer perceptron to convert the current optimal physiological state of the patient into an optimal anesthetic drug dose, and uses the optimal anesthetic drug dose as the output of the anesthetic drug dose optimization decision.

5. The artificial intelligence-based anesthetic drug dosage optimization system according to claim 1, characterized in that: Both the historical dose optimization raw dataset and the real-time dose optimization raw dataset include patient basic information data, physiological monitoring data, drug-related data, surgery-related data, and environmental data. The historical dose optimization raw dataset also includes clinical outcome data.

6. The artificial intelligence-based anesthetic drug dosage optimization system according to claim 1, characterized in that: In the preliminary data processing module, the spatiotemporal alignment of the data is used to ensure the consistency of physiological data from different sources and sampling frequencies in the time dimension. Specifically, it is achieved by using resampling, time delay compensation, and timestamp synchronization techniques to obtain a data sequence with a unified time reference. The feature construction is used to extract meaningful feature representations. Specifically, feature engineering is performed on patient basic information data, physiological monitoring data, drug-related data, surgery-related data, and environmental data to obtain feature representations for each data. The normalization process is used to normalize the data scale to eliminate the influence of dimensions. Specifically, it processes the data using the Z-score standardization method to obtain data with uniform scale. The dataset segmentation is used to obtain training data and test data, specifically by segmenting the original dataset of past dose optimization. By performing the data spatiotemporal alignment, feature construction, and normalization, the real-time dose optimization raw dataset is preliminarily processed to obtain the dataset to be optimized. By performing the data spatiotemporal alignment, feature construction, normalization, and dataset segmentation, the past dose optimization raw dataset is preliminarily processed to obtain the state prediction training set and the state prediction test set.

Images