Anesthetic dosage prediction method and system based on machine learning

Abstract

The invention relates to the technical field of knowledge bases, in particular to an anesthetic dosage prediction method and system based on machine learning, and the method comprises the following steps: obtaining preoperative genome data, metabonomics data, intraoperative real-time vital sign data and operating room environment data of a patient, the genome data comprises CYP450 enzyme gene polymorphism data, and the metabonomics data comprises CYP450 enzyme gene polymorphism data; the metabonomics data comprises propofol metabolite concentration data, and the intraoperative real-time vital sign data comprises a BIS value, an MAP value and an HR value; according to the method, a dynamic hypergraph model containing drug nodes, gene nodes, metabolite nodes and environment nodes is constructed, the drug nodes and the gene nodes are connected through hyperedge to represent the regulation and control effect of drug metabolism genes on drug metabolism, the anesthesia safety and effectiveness can be improved, and the urgent clinical requirements for personalized and precise anesthesia are met.

Description

Technical Field

[0001] This invention relates to the field of anesthesia technology, and more specifically to a method and system for predicting anesthetic drug dosage based on machine learning. Background Technology

[0002] In clinical surgery, commonly used anesthetic drugs are classified into intravenous anesthetics (such as propofol and etomidate), inhaled anesthetics (such as sevoflurane and desflurane), sedatives (such as midazolam), analgesics (such as sufentanil and remifentanil), and muscle relaxants (such as succinylcholine, rocuronium bromide, and atracurium).

[0003] These sedatives, analgesics, and anesthetics require strict supervision by relevant hospital departments. For anesthetic drugs and Class I psychotropic drugs (referred to as "narcotic drugs"), a fully closed-loop supervision principle is necessary, including double-locked storage, restrictions on prescription rights, empty ampoule recycling, and electronic traceability systems connected to the national monitoring network. However, current anesthetic drug management has pain points. For example, anesthesiologists mainly rely on basic parameters such as age, weight, and BMI, and experience to estimate the operation time when selecting medications. However, patient history, surgical type and grade, intraoperative stimulation intensity, anesthesia area range, and surgeon's skill level all affect the operation time and dosage of anesthetic drugs. This leads to prediction errors and waste due to incompatibilities. Data shows that the waste rate of narcotic drugs during surgery is as high as 20%-25%, of which about 15% is due to dosage prediction errors. Furthermore, clinical practice has not yet achieved correlation analysis between preoperative examination data (such as complete blood count, comprehensive biochemistry, coagulation tests, and urinalysis) and intraoperative monitoring data with anesthetic dosage. Although these data are readily available and have broad applicability, current technologies have not incorporated them into multimodal data modeling, resulting in a lack of comprehensive clinical guidance for adjusting anesthetic drug dosages. Therefore, there is an urgent need to construct a multidimensional predictive model integrating patient genetic differences, metabolite variations, vital sign monitoring, and environmental variables to reduce unnecessary drug waste. This can be achieved by establishing genotype-phenotype association maps and developing machine learning-based dynamic adjustment algorithms to form an intelligent anesthesia decision support system. This system will overcome the limitations of traditional experience-based anesthesia medicine and achieve the clinical goal of precise anesthesia throughout the entire process. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a machine learning-based method for predicting anesthetic drug dosage. The method includes the following steps: acquiring preoperative genomic data, metabolomics data, intraoperative real-time vital signs data, and operating room environmental data. The genomic data includes gene polymorphism data of liver and kidney function indicators. The metabolomics data includes metabolic concentration data of anesthetic drugs such as remifentanil citrate injection and propofol. The intraoperative real-time vital signs data include bispectral index (BIS), mean arterial pressure (MAP), oxygen saturation (SpO2), non-invasive blood pressure (BP), heart rate (HR), body temperature, and muscle relaxation monitoring values. A dynamic hypergraph model is constructed, comprising drug nodes, gene nodes, metabolite nodes, and environmental nodes. The drug nodes and gene nodes are connected by hyperedges to characterize the drug. The regulatory role of metabolic genes on drug metabolism is explored. Metabolite nodes and environmental nodes are connected by dynamic weighted edges to characterize the influence of environmental factors on metabolite concentration. A dynamic hypergraph convolutional network is used to learn features from the dynamic hypergraph model, extracting high-order interaction features between multimodal data. These high-order interaction features, along with the temporal features of real-time intraoperative vital signs data, are input into a causal Transformer network. The causal Transformer network captures the lag effect of drug infusion through causal convolutional layers and models the long-range dependence of physiological indicators at different times using a multi-head self-attention mechanism, outputting a predicted feature of the current anesthesia depth. A reinforcement learning algorithm is used to generate an anesthetic drug dosage adjustment strategy. The reward function of the reinforcement learning algorithm includes a penalty term for anesthesia depth deviation, a penalty term for hemodynamic fluctuations, and a clinical safety constraint term. Optimizing the reward function enables dynamic adjustment of the anesthetic dosage.

[0005] Further preferably, the step of constructing the dynamic hypergraph model includes: mapping different categories of data to corresponding nodes, wherein drug nodes correspond to target anesthetic drugs, gene nodes correspond to key physiological indicators affecting drug metabolism, metabolite nodes correspond to key metabolites in drug metabolism pathways, and environmental nodes correspond to operating room temperature and humidity parameters; establishing hyperedges connecting drug nodes and related gene nodes, wherein the connection relationship of the hyperedges is determined based on clinically known correlations between liver and kidney function and drug metabolism, characterizing the regulatory effect of specific genes on the activity of drug-metabolizing enzymes; and establishing dynamic weighted edges connecting metabolite nodes and environmental nodes, wherein the initial weights of the dynamic weighted edges are set based on experimental data on the stability of metabolites caused by environmental factors, and are dynamically adjusted during model training using a gradient descent algorithm.

[0006] Further preferably, the feature learning process of the dynamic hypergraph convolutional network includes: hierarchical feature aggregation of the hyperedge set to which each node belongs; transformation of the node features within the hyperedge using a multilayer perceptron (MLP), where the input of the MLP is the sum of the feature vectors of all nodes within the hyperedge, and the output is the feature vector of the current node after hypergraph convolution; and calculation of dynamic weights between nodes within the hyperedge using an attention mechanism to adaptively adjust the contribution of different nodes to the features of the target node. The calculation process of the attention mechanism includes inputting the feature vectors of the target node and its neighboring nodes into the LeakyReLU activation function after linear transformation, and normalizing them using the softmax function to obtain the attention weights between nodes.

[0007] Further preferably, the input data of the causal Transformer network includes multimodal feature vectors of the current time and the past n time steps, the kernel size of the causal convolutional layer is n+1, and the length of the output sequence is ensured to be consistent with the length of the input sequence through padding operation, and the multi-head self-attention mechanism includes at least 8 attention heads, each attention head independently calculates the query vector, key vector and value vector, and the output results of each attention head are integrated through concatenation operation.

[0008] More preferably, the attention weights between nodes in the dynamic hypergraph convolutional network are calculated as follows: ; in, Indicates the superedge Middle node With nodes Attention weights and The weight matrix is ​​a learnable matrix. and They represent the first Nodes in hypergraph convolution and nodes eigenvectors, Indicates the superedge All nodes within, This is the activation function.

[0009] More preferably, the causal convolutional layer output features of the causal Transformer network are calculated as follows: ; in, Indicates the current time Causal convolution output features The kernel size is [size]. For the first Each convolutional kernel weight, Indicates lag Input characteristics at time step The preset drug infusion lag time constant is determined using pharmacokinetic experimental data.

[0010] More preferably, the reward function expression of the reinforcement learning algorithm is: ; in, Indicates the current time The reward value, , , The preset weighting coefficients, This is the current bispectral index measurement value of the EEG. The bispectral index of the electroencephalogram (EEG) corresponding to the target depth of anesthesia. This represents the difference between the mean arterial pressure at the current time and the mean arterial pressure at the previous time. This represents the current dosage of anesthetic drugs. Within the clinically safe dosage range, This is an indicator function that outputs 1 when the dose exceeds the safe range, and 0 otherwise.

[0011] A machine learning-based anesthetic drug dosage prediction system, applied to any of the machine learning-based anesthetic dosage prediction methods described above, is characterized by comprising: a data acquisition module configured to acquire preoperative genomic data, metabolomics data, intraoperative real-time vital signs data, and operating room environmental data of the patient; wherein the genomic data is acquired via a gene sequencer, the metabolomics data is obtained via mass spectrometry, the intraoperative real-time vital signs data is acquired in real time via a monitor, and the operating room environmental data is acquired via a sensor network; and a dynamic hypergraph construction module configured to construct a dynamic hypergraph model including drug nodes, gene nodes, metabolite nodes, and environmental nodes, wherein drug nodes and gene nodes are connected by hyperedges, and metabolite nodes and environmental nodes are connected by hyperedges. The environment nodes are connected via dynamically weighted edges; the feature learning module is configured to perform feature learning on the dynamic hypergraph model through a dynamic hypergraph convolutional network to extract high-order interaction features between multimodal data; the temporal modeling module is configured to input the high-order interaction features and intraoperative real-time vital signs data into a causal Transformer network, which outputs anesthesia depth prediction features through causal convolutional layers and a multi-head self-attention mechanism; the dose adjustment module is configured to generate anesthesia drug dose adjustment strategies based on a reinforcement learning algorithm, which achieves dynamic adjustment of the anesthesia dose by optimizing a reward function that includes anesthesia depth deviation penalty term, hemodynamic fluctuation penalty term, and clinical safety constraint term.

[0012] Further preferably, the dynamic hypergraph convolutional network includes at least three hypergraph convolutional layers, each containing a hyperedge attention calculation unit and a feature aggregation unit. The hyperedge attention calculation unit is used to calculate the dynamic attention weights between nodes within the hyperedge, and the feature aggregation unit aggregates the node features within the hyperedge using a multilayer perceptron. The causal Transformer network includes at least two encoder layers, each containing a causal convolutional sublayer and a multi-head self-attention sublayer. The output of the causal convolutional sublayer serves as the input of the multi-head self-attention sublayer.

[0013] Further preferably, the system also includes an interpretability analysis module configured to calculate the contribution of each input feature to the anesthetic dose prediction result using the SHAP value analysis method, and generate a visualized feature importance ranking chart. The input features include gene polymorphism data of liver and kidney function indicators; metabolic concentration data of anesthetic drugs such as remifentanil citrate injection and propofol; real-time intraoperative vital signs data including BIS value, MAP value, SpO2 value, BP value, HR value, body temperature value, muscle relaxation monitoring value, and inhaled anesthetic drug concentration (MAC); operating room temperature and humidity parameters; the system's hardware architecture includes a server cluster, which contains a graphics processing unit (GPU) and a central processing unit (CPU), and the data acquisition module is connected to the server cluster via a wired network and / or a wireless network.

[0014] Technical effects: This invention effectively addresses the problem of insufficient accuracy in anesthetic dose prediction in previous techniques by integrating multi-dimensional data, constructing a dynamic hypergraph model, and combining it with reinforcement learning. It creatively incorporates genomic and metabolomics data, utilizes a dynamic hypergraph convolutional network to mine high-order relationships between data, models drug lag effects through a causal Transformer network, and achieves dynamic dose adjustment using a reinforcement learning reward function with multiple penalty terms. Compared to traditional techniques, this approach comprehensively considers individual differences and environmental factors, accurately captures drug metabolism patterns, significantly improves the accuracy of anesthetic dose prediction, reduces the risk of excessively deep or shallow anesthesia, and provides safe, precise, and personalized anesthesia plans for clinical practice. Attached Figure Description

[0015] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the machine learning-based anesthetic drug dosage prediction method of this application; Figure 2 This is a block diagram of the machine learning-based anesthetic drug dosage prediction system of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are part of this invention.

[0018] Anesthetic drugs include potent analgesics and sedatives such as fentanyl, remifentanil, and propofol. Due to their narrow therapeutic window, significant individual variability, and complex synergistic effects, these drugs require close monitoring by healthcare professionals in clinical practice. However, in current clinical practice, the determination of anesthetic drug dosage still heavily relies on the physician's personal experience, typically estimated based only on basic parameters such as surgical duration, patient height, and weight. This experience-driven approach to medication has significant limitations: on the one hand, insufficient dosage accuracy can lead to anesthesia that is too deep or too shallow; on the other hand, insufficient consideration of individual differences can result in drug waste or suboptimal efficacy. Existing prediction methods often rely solely on the patient's basic physiological indicators or single-modality data, failing to comprehensively consider individual genetic differences, dynamic changes in metabolites, and the impact of environmental factors on drug metabolism, leading to insufficient prediction accuracy. Furthermore, existing methods lack effective modeling of drug infusion lag effects and long-term dependence on physiological indicators, making it impossible to achieve dynamic and precise adjustment of anesthetic dosage, and also making it difficult to ensure both anesthetic efficacy and patient hemodynamic stability and clinical safety.

[0019] Based on this, please refer to Figure 1This embodiment provides a machine learning-based method for predicting anesthetic drug dosage, including: acquiring preoperative genomic data, metabolomics data, intraoperative real-time vital signs data, and operating room environmental data of the patient; simultaneously acquiring basic patient information, including age, weight, body type classification (standard weight / obese / underweight), and historical duration data of similar surgeries, such as the average duration of similar surgeries for adults / children of the same surgeon; genomic data covering gene polymorphism data of liver and kidney function indicators; and metabolomics data including the metabolic concentrations of anesthetic drugs such as remifentanil citrate injection and propofol. The intraoperative real-time vital signs data include BIS value, MAP value, SpO2 value, BP value, HR value, body temperature value, muscle relaxation monitoring value, and inhaled anesthetic drug concentration (MAC). A dynamic hypergraph model is constructed, in which drug nodes and gene nodes are connected by hyperedges, and surgical feature nodes correspond to the estimated operation time. Based on historical data and patient age and weight classification, hyperedges are connected to drug nodes to characterize the correlation between operation time and total anesthetic drug requirements. For example, for every 30-minute increase in operation time, the basic requirement for propofol maintenance dose increases by X mg.

[0020] This study demonstrates the regulation of drug metabolism by genes related to drug metabolism. Metabolite nodes and environmental nodes are connected via dynamic weighted edges, reflecting the influence of environmental factors on metabolite concentration. A dynamic hypergraph convolutional network is used for feature learning to extract high-order interaction features between multimodal data. These high-order interaction features, along with temporal features from real-time intraoperative vital signs data, are input into a causal Transformer network. This network captures the lag effect of drug infusion through causal convolutional layers and models the long-range dependencies of physiological indicators at different times using a multi-head self-attention mechanism, outputting a predicted feature of the current anesthesia depth. A dose adjustment strategy is generated based on a reinforcement learning algorithm. The reward function of the reinforcement learning algorithm includes a penalty term for anesthesia depth deviation, a penalty term for hemodynamic fluctuations, and a clinical safety constraint term. By optimizing the reward function, the anesthetic dose is dynamically adjusted to ensure that the predicted anesthesia depth approaches the target depth while maintaining dose adjustment within safe boundaries.

[0021] The intraoperative real-time vital signs data also include end-tidal carbon dioxide partial pressure (PetCO2). If necessary, arterial analysis of various ion levels can also be performed intraoperatively. Real-time data acquisition, along with BIS, MAP, SpO2, BP, HR, body temperature, muscle relaxation monitoring, and inhaled anesthetic drug concentration (MAC), constitutes a multidimensional physiological state assessment index. End-tidal carbon dioxide partial pressure indirectly reflects the degree of respiratory function inhibition by anesthetic drugs, further optimizing the calculation accuracy of anesthetic depth deviation in the reward function. The target anesthetic drugs include commonly used clinical anesthetics such as propofol, remifentanil, and desflurane. Propofol metabolite concentration data, combined with its infusion rate (e.g., a 50 mg / ml infusion dose), allows for more accurate modeling.

[0022] The causal Transformer network models the lag effect of drug infusion and the long-term dependence of physiological indicators, enabling the system to more accurately predict the trend of anesthesia depth over time. The reinforcement learning algorithm, combined with a reward function containing multiple penalty and safety constraints, ensures anesthetic efficacy while also considering patient hemodynamic stability and clinical safety, avoiding the risks of excessively deep or shallow anesthesia. This provides a more reliable and personalized dose prediction and adjustment scheme for clinical anesthesia.

[0023] When constructing multimodal data association models, traditional methods cannot clearly and accurately express the complex relationships between different types of data, especially the intrinsic connections between drugs, genes, metabolites and environmental factors.

[0024] Traditional methods lack effective modeling tools for understanding the regulatory role of drug metabolism genes in drug metabolism and the influence of environmental factors on metabolite concentrations. This results in the inability to fully utilize this crucial information, affecting the accuracy and reliability of anesthetic dose prediction.

[0025] Based on this, the steps of constructing the dynamic hypergraph model include: mapping different categories of data to corresponding nodes, with drug nodes corresponding to target anesthetic drugs, gene nodes corresponding to key physiological indicators affecting drug metabolism, metabolite nodes corresponding to key metabolites in drug metabolism pathways, and environmental nodes corresponding to operating room temperature and humidity parameters.

[0026] Hyperedges are established to connect drug nodes and related gene nodes. The connections of these hyperedges are determined based on a drug metabolism pathway database to characterize the regulatory role of specific genes on the activity of drug-metabolizing enzymes. Dynamically weighted edges are established to connect metabolite nodes and environmental nodes. The initial weights of these dynamic weighted edges are set based on experimental data on the impact of environmental factors on metabolite stability and are dynamically adjusted during model training using a gradient descent algorithm.

[0027] The introduction of dynamic weighted edges takes into account the dynamic influence of environmental factors on metabolite concentrations and can adaptively adjust the weights during training, enabling the model to better adapt to changes in metabolites under different environmental conditions.

[0028] Meanwhile, traditional methods lack a dynamic adjustment mechanism for the relationships between nodes, making it difficult to adapt to dynamic changes in data and affecting the effectiveness of feature learning and the accuracy of anesthetic dosage prediction. Based on this, the feature learning process of the proposed dynamic hypergraph convolutional network includes: hierarchical feature aggregation of the hyperedge set to which each node belongs; transformation of the node features within the hyperedge using a multilayer perceptron (MLP); the input of the MLP is the sum of the feature vectors of all nodes within the hyperedge, and the output is the feature vector of the current node after hypergraph convolution. A dynamic weight between nodes within the hyperedge is calculated through an attention mechanism to adaptively adjust the contribution of different nodes to the target node's features. The calculation process of the attention mechanism includes linearly transforming the feature vectors of the target node and its neighboring nodes, inputting them into a LeakyReLU activation function, and then normalizing them using a softmax function to obtain the attention weights between nodes.

[0029] When processing the temporal characteristics of real-time intraoperative vital signs data, traditional methods often ignore the lag effect of drug infusion, failing to accurately capture the delayed time that drugs affect physiological indicators after administration, leading to a lag in the prediction of changes in anesthesia depth. Simultaneously, traditional methods struggle to effectively model the long-range dependencies between physiological indicators at different times, failing to fully consider the evolution of physiological indicators over time. This results in anesthetic dosage predictions not adapting well to the dynamic changes in the patient's physiological state, affecting the accuracy and timeliness of predictions.

[0030] Based on this, the input data of the causal Transformer network includes multimodal feature vectors from the current time step and the past n time steps. The kernel size of the causal convolutional layer is n+1, and padding operations are used to ensure that the length of the output sequence is consistent with the length of the input sequence. The multi-head self-attention mechanism includes at least 8 attention heads, each of which independently calculates the query vector, key vector, and value vector, and the output results of each attention head are integrated through a concatenation operation. By using the multimodal feature vectors from the current time step and multiple past time steps as input to the causal Transformer network and setting the kernel size of the causal convolutional layer to a specific size, this scheme can effectively capture the lag effect of drug infusion, accurately simulate the process that the drug takes a certain amount of time to affect physiological indicators after infusion, avoid the lag in prediction, and make the prediction of anesthesia depth more consistent with reality.

[0031] In dynamic hypergraph convolutional networks, traditional node feature aggregation methods typically use fixed weights, which cannot adaptively adjust the weights between nodes according to the differences in node features. This results in the inability to fully highlight the contribution of key nodes to the features of the target node, making it difficult to effectively explore the complex relationships between nodes within the hyperedge, thus affecting the effectiveness of feature learning and the accuracy of anesthetic dose prediction.

[0032] Based on this, the attention weights between nodes in a dynamic hypergraph convolutional network are calculated as follows: ; in, Indicates the superedge Middle node With nodes Attention weights and The weight matrix is ​​a learnable matrix. and They represent the first Nodes in hypergraph convolution and nodes eigenvectors, Indicates the superedge All nodes within, To activate the function, a nonlinear transformation is introduced to solve the problems of traditional methods. Addressing the neuron death problem in functions and enhancing the network's ability to express complex features.

[0033] In dynamic hypergraph convolutional networks, this formula is used to calculate hyperedges. Middle node With nodes Attention weights Its core purpose is to adaptively adjust the contribution of different nodes to the features of the target node in order to more accurately explore the complex relationships between nodes within the hyperedge.

[0034] In the formula, and They represent the first Nodes in hypergraph convolution and nodes The feature vectors of the nodes contain the data information represented by the nodes. For example, in the scenario of anesthetic dosage prediction, the nodes may represent drugs, genes, metabolites or environmental factors, and their feature vectors encode the attributes and states of the corresponding data.

[0035] and These are learnable weight matrices that are continuously optimized during model training using the backpropagation algorithm. The goal is to adjust the mapping relationship of node features so that the network can better capture the intrinsic connections between data.

[0036] It is an activation function that solves the problems of traditional The problem of neurons "dying" when the gradient of a function is zero on the negative half-axis. In the formula, For nodes and nodes The sum of the eigenvectors after weight matrix transformation undergoes a nonlinear transformation, introducing a nonlinear factor to enhance the network's expressive power. Through this activation function, the network can learn more complex feature relationships. (Molecular part) By performing an exponential operation on the output of the activation function, the differences in features are amplified, making the important feature relationships more prominent in the weight calculation.

[0037] denominator Then for the superedge All nodes and nodes The summation of the characteristic combinations is used as a normalization factor. Through operations on the numerator and denominator, the final result is... It is a normalized weight value, ranging from 0 to 1, which reflects the node's... For nodes The relative importance of feature aggregation. Throughout the hypergraph convolution process, by continuously calculating and adjusting these attention weights, the network can adaptively focus on key nodes and suppress secondary information, thereby more effectively extracting high-order interaction features between multimodal data and providing more valuable feature representations for subsequent anesthetic dose prediction.

[0038] Traditional time-series data processing models often overlook the time delay in the effects of drugs on the body after infusion when analyzing intraoperative physiological indicators. This leads to an inability to accurately simulate the drug onset process, resulting in a disconnect between predicted anesthetic depth and actual drug efficacy. Furthermore, conventional convolution operations struggle to accurately capture the dynamic characteristics of drug lag effects and cannot effectively extract key temporal features containing lag information. This affects the real-time performance and accuracy of anesthetic dose prediction, failing to meet the clinical need for precise control of anesthetic status.

[0039] Based on this, the output features of the causal convolutional layer in the causal Transformer network are calculated as follows: ; in, Indicates the current time Causal convolution output features The kernel size is [size]. For the first Each convolutional kernel weight, Indicates lag Input characteristics at time step The preset drug infusion lag time constant was determined using pharmacokinetic experimental data. This scheme utilizes a causal Transformer network with a specific kernel size. The multimodal feature vectors of the current and historical moments are calculated and combined with the drug infusion lag time constant precisely measured experimentally. , for lag The input features at each time point are weighted and summed to incorporate the lag effect of drug infusion into the feature calculation process, thereby enabling the extraction of time-series features containing lag information.

[0040] This formula is used to calculate the causal convolutional layer in a causal Transformer network at the current time step. Output characteristics Its core function is to capture the lag effect of drug infusion, enabling the network to accurately simulate the delayed process from drug infusion to the production of physiological effects.

[0041] In the formula, The kernel size determines the range of historical data used in the feature calculation at the current time step. For example, if... Then, the data from the current moment and the past four moments will be considered in the feature calculation. For the first Each convolutional kernel has weights, which are learnable parameters that are continuously adjusted during model training using optimization algorithms to adapt to different drug metabolism and physiological response characteristics. This represents the contribution weight of the corresponding historical data to the characteristics of the current moment.

[0042] Indicates lag Input features at time t, where This is a preset drug infusion lag time constant, determined through pharmacokinetic experimental data. During anesthesia, there is a certain time delay between drug infusion and its effect on human physiological indicators. It's about quantifying this delay time. For example, the delay time of a certain anesthetic drug. The timeframe is 3 minutes, which means that when calculating the characteristics of the current moment, the impact of the drug infusion that started 3 minutes ago on the current physiological state needs to be considered.

[0043] Traditional methods of adjusting anesthetic dosage lack a comprehensive consideration of anesthetic efficacy, hemodynamic stability, and clinical safety, often resulting in compromises. For example, simply pursuing the target depth of anesthesia may lead to hemodynamic disturbances such as a sudden drop in blood pressure, or cause the risk of drug overdose due to the lack of strict control over the dosage range. These methods fail to balance multiple needs in complex clinical scenarios and make it difficult to ensure the safety and comfort of patients during anesthesia.

[0044] Based on this, the reward function expression for the reinforcement learning algorithm is: ; in, Indicates the current time The reward value, , , The preset weighting coefficients, This is the current bispectral index measurement value of the EEG. The bispectral index of the electroencephalogram (EEG) corresponding to the target depth of anesthesia. This represents the difference between the mean arterial pressure at the current time and the mean arterial pressure at the previous time. This represents the current dosage of anesthetic drugs. Within the clinically safe dosage range, The indicator function outputs 1 when the dose exceeds the safe range and 0 otherwise. The reward function constructed in this scheme takes into account anesthesia depth deviation, hemodynamic fluctuations, and dose safety boundaries. It balances the relationship between these three factors through preset weight coefficients and uses a reinforcement learning algorithm to optimize the reward value, thereby generating a dynamic adjustment strategy for the anesthetic dose.

[0045] The reward function can be dynamically adjusted based on the operation time. When the estimated remaining operation time is greater than 30 minutes, the weight of the hemodynamic fluctuation penalty term is increased. Prioritize maintaining the patient's physiological stability; when the remaining time is less than 10 minutes, increase the weight of the penalty for deviation in anesthesia depth. To avoid delayed awakening.

[0046] This formula defines the reward function in reinforcement learning algorithms. It is used to comprehensively consider multiple factors such as the depth of anesthesia, hemodynamic stability, and clinical safety during the adjustment of anesthetic dosage, so as to guide the algorithm to generate a reasonable anesthetic drug dosage adjustment strategy.

[0047] In the formula, For the current moment The bispectral index of the brain is an important physiological indicator reflecting the depth of anesthesia in patients. The bispectral index of the electroencephalogram (EEG) corresponding to the target depth of anesthesia.

[0048] The squared deviation between the current depth of anesthesia and the target depth of anesthesia is calculated. The method of calculating the squared deviation amplifies the penalty for larger deviations, prompting the algorithm to prioritize ensuring that the anesthesia effect reaches the expected target.

[0049] These are preset weighting coefficients used to adjust the importance of anesthesia depth deviation in the reward function. The degree of emphasis on the accuracy of anesthesia depth can be flexibly adjusted according to different surgical needs and patient conditions.

[0050] This represents the difference between the mean arterial pressure at the current time and the mean arterial pressure at the previous time. The absolute value of the difference is used to measure the degree of fluctuation in hemodynamics.

[0051] Maintaining hemodynamic stability during anesthesia is crucial. This part punishes hemodynamic fluctuations and forces the algorithm to take into account the stability of physiological indicators such as blood pressure when adjusting the dosage, so as to avoid hemodynamic disturbances such as sudden rise or fall in blood pressure caused by dosage adjustment. The corresponding weighting coefficient is used to control the influence of hemodynamic stability on the reward function. This represents the current dosage of anesthetic drugs. This is within the clinically safe dosage range.

[0052] This is an indicator function that outputs 1 when the dose exceeds the safe range, and 0 otherwise.

[0053] As a clinical safety constraint, by penalizing situations that exceed the safe dosage range, the dosage is strictly limited to the safe range, eliminating the risk of drug overdose or underdose. The weighting coefficients for this constraint term are used to ensure the importance of clinical safety in dosage adjustment decisions. The reward function is obtained by adding these three parts. It can comprehensively balance anesthetic efficacy, physiological stability, and medication safety in complex clinical settings. Reinforcement learning algorithms maximize... By learning towards a specific goal and through continuous trial and error, a strategy for adjusting anesthetic drug dosage that takes into account multiple factors is gradually developed. This effectively reduces the risk of complications caused by inappropriate dosage during anesthesia and provides patients with safer and more precise anesthesia plans.

[0054] Existing anesthesia dose prediction systems suffer from incomplete data collection and poor interoperability of functional modules. On the one hand, they rely only on partial physiological indicator data, failing to cover key influencing factors such as genes, metabolites, and the environment. On the other hand, each functional module is relatively independent, lacking deep integration of data and algorithms, which prevents the full exploitation of the value of multi-source data, making it difficult to achieve high-precision and intelligent anesthesia dose prediction and meet the needs of personalized anesthesia in clinical practice.

[0055] Based on this, please refer to Figure 2This embodiment provides a machine learning-based anesthetic drug dosage prediction system, including a data acquisition module configured to acquire preoperative genomic data, metabolomics data, intraoperative real-time vital signs data, and operating room environmental data. Genomic data is acquired using a gene sequencer, metabolomics data is obtained using a mass spectrometer, intraoperative real-time vital signs data is acquired in real time using a monitor, and operating room environmental data is acquired through a sensor network. A dynamic hypergraph construction module is configured to construct a dynamic hypergraph model containing drug nodes, gene nodes, metabolite nodes, and environmental nodes. A feature learning module is configured to perform feature learning on the dynamic hypergraph model using a dynamic hypergraph convolutional network. A temporal modeling module is configured to input high-order interaction features and intraoperative real-time vital signs data into a causal Transformer network. A dosage adjustment module is configured to generate an anesthetic drug dosage adjustment strategy based on a reinforcement learning algorithm. Each module has a clear division of labor and collaborates closely, forming a complete data processing and prediction workflow.

[0056] The data acquisition module can also acquire routine preoperative clinical examination data of patients, including but not limited to: Comprehensive biochemical panel: including liver and kidney function, blood glucose, electrolytes, etc., used to assess the functional status of organs metabolizing drugs and to help correct the interpretation of metabolomics data; The four coagulation parameters, including prothrombin time and activated partial thromboplastin time, are used to assess the patient's baseline hemodynamic status and optimize the weighting coefficients of the hemodynamic fluctuation penalty term. Complete blood count (CBC): Hematocrit, hemoglobin, etc., are used to help assess the patient's oxygen supply capacity and are linked with intraoperative arterial blood oxygen saturation data to improve the accuracy of anesthesia depth prediction. The system includes four immune markers: hepatitis B surface antigen, hepatitis C antibody, etc., used to rule out underlying immune diseases that may affect drug metabolism. Blood gas analysis assesses acid-base balance and oxygenation status, optimizing preoperative preparation and intraoperative management. Through multi-type data acquisition modules, the system comprehensively integrates genetic, metabolic, physiological, and environmental data, overcoming the data limitations of traditional systems and providing a rich information foundation for accurate prediction. The dynamic hypergraph construction and feature learning module can deeply mine the complex relationships and high-order features between multi-source data; the temporal modeling module accurately processes the temporal characteristics of data; and the dosage adjustment module achieves intelligent decision-making based on reinforcement learning.

[0057] Traditional neural network architectures suffer from insufficient feature extraction layers and low module collaboration efficiency when processing anesthesia-related multimodal data. Shallow networks struggle to capture the deep semantic features of the data, and the lack of collaborative optimization mechanisms between different functional modules limits the model's ability to understand complex data relationships. This makes it difficult to effectively cope with the dynamic changes in data during anesthesia, thus affecting the accuracy and stability of anesthetic dose prediction.

[0058] Based on this, the Dynamic Hypergraph Convolutional Network (DHCNN) comprises at least three hypergraph convolutional layers, each containing a hyperedge attention computation unit and a feature aggregation unit; the Causal Transformer Network comprises at least two encoder layers, each containing a causal convolutional sub-layer and a multi-head self-attention sub-layer. The DHCNN, through its multi-layer structure, progressively extracts and abstracts features from the hypergraph data. The hyperedge attention computation unit and the feature aggregation unit work collaboratively, adaptively adjusting node weights and aggregating features. The encoder layer of the Causal Transformer Network captures drug hysteresis effects through causal convolutional sub-layers, and the multi-head self-attention sub-layer models long-range dependencies, achieving deep information interaction and feature optimization between layers.

[0059] Meanwhile, the system's hardware architecture is poorly designed, resulting in low data transmission and processing efficiency, failing to meet the real-time needs of clinicians for rapid and accurate prediction of anesthetic doses. Therefore, the system also includes an interpretability analysis module, configured to calculate the contribution of each input feature to the anesthetic dose prediction result using the SHAP value analysis method, generating a visual feature importance ranking chart. The system's hardware architecture includes a server cluster containing a GPU and a CPU, with the data acquisition module connected to the server cluster via a wired or wireless network. The interpretability analysis module uses the SHAP value method to quantify the impact of each feature on the prediction result; the hardware architecture combines a server cluster with GPUs and CPUs, utilizing the parallel computing capabilities of GPUs to accelerate data processing, and ensuring efficient data transmission through network connectivity. By generating a visual feature importance ranking chart, the interpretability analysis module transforms the model's complex decision-making process into intuitive and easy-to-understand information, helping clinicians clearly understand the role of each data feature in anesthetic dose prediction, enhancing the system's transparency and credibility, and promoting clinicians' trust in and application of the prediction results.

[0060] In specific surgical scenarios, the application process of this system is as follows: Preoperative data collection: Gene polymorphism data of liver function indicators and kidney function indicators are obtained through gene sequencer, and the concentration of propofol metabolites is detected by mass spectrometer; at the same time, routine examination data such as complete blood count, red blood cell count, platelet count, biochemical items, albumin, creatinine, and coagulation four items are collected from the patient. Among them, liver and kidney function indicators are used to adjust the initial weight of metabolite nodes in the dynamic hypergraph model.

[0061] Intraoperative data monitoring: Real-time data collection via monitors includes BIS, MAP, SpO2, BP, HR, body temperature, muscle relaxation monitoring, end-tidal carbon dioxide partial pressure, and inhaled anesthetic drug concentration (MAC). A sensor network records operating room temperature (22-25℃) and humidity (50-60%). Intraoperative infusion times and dosage changes for propofol and remifentanil are also recorded. Model adaptation and adjustment: For general anesthesia, the drug infusion lag time constant of the causal Transformer network is set based on propofol and remifentanil pharmacokinetic experimental data (e.g., propofol d = 2-3 minutes). The time window parameters of the time sequence features are optimized by combining the time difference between the surgery start time (e.g., 09:05) and the anesthesia start time (e.g., 08:47).

[0062] Safety constraints are reinforced: The clinically safe dose range [Dosemin, Dosemax] in the reward function references the clinical use guidelines for anesthetic drugs. For example, the induction dose of propofol is set at 1.5-2.5 mg / kg, and the maintenance dose is set at 4-12 mg / kg / h. The weighting coefficient (γ2) of the hemodynamic fluctuation penalty term is dynamically adjusted based on the patient's intraoperative blood loss (e.g., <50 ml). When blood loss increases, γ2 is increased to reduce blood pressure fluctuations. This invention integrates readily available clinical examination data, such as complete blood count, comprehensive biochemical tests, coagulation tests, and intraoperative monitoring data, making the model input closer to actual clinical scenarios, reducing reliance on complex data such as gene sequencing, and improving the feasibility of technology implementation.

[0063] After optimizing model parameters based on the characteristics of anesthetic drugs such as propofol and remifentanil, the risks of excessively deep anesthesia (e.g., respiratory depression) or insufficient anesthesia (e.g., intraoperative awareness) can be significantly reduced. This is particularly suitable for anesthesia management in special cases such as ASA classification II-III and patients with a history of cerebral infarction. It effectively synergizes with medication guidelines in clinical anesthesia records (e.g., preoperative fasting, intraoperative fluid control), providing complete technical support for precise dosage control of anesthetic drugs. The key input data involved in this invention are all clinically readily available data: liver function (ALT, AST), kidney function (creatinine, blood urea nitrogen, etc.): obtained from preoperative comprehensive biochemical examinations and hospital laboratory tests; concentrations of metabolites such as remifentanil citrate and propofol: obtained through intraoperative micro-blood sample testing and routine clinical metabolite detection methods; operation duration and historical data of similar surgeries: obtained from the hospital's electronic surgical record system; patient basic information (age, weight, etc.): obtained from the patient's electronic medical record system.

[0064] The above data requires no special testing procedures and is feasible and readily available. The above is merely a preferred embodiment of the present invention and is not intended to limit the invention in any other way. Any person skilled in the art may use the disclosed technical content to make changes or modifications to create equivalent embodiments applicable to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A machine learning-based method for predicting anesthetic drug dosage, characterized in that, The procedure includes the following steps: acquiring the patient's preoperative genomic data, metabolomics data, intraoperative real-time vital signs data, and operating room environmental data. The genomic data includes gene polymorphism data for liver and kidney function indicators. The metabolomics data includes metabolic concentration data of anesthetic drugs such as remifentanil citrate injection and propofol. The intraoperative real-time vital signs data include BIS value, MAP value, SpO2 value, BP value, HR value, body temperature value, and muscle relaxation monitoring value. A dynamic hypergraph model is constructed, comprising drug nodes, gene nodes, metabolite nodes, and environmental nodes. The drug nodes and gene nodes are connected by hyperedges to characterize the drug metabolism gene pair. The regulation of drug metabolism is demonstrated by connecting metabolite nodes and environmental nodes through dynamic weighted edges to characterize the influence of environmental factors on metabolite concentration. A dynamic hypergraph convolutional network is used to learn features of the dynamic hypergraph model, extracting high-order interaction features between multimodal data. These high-order interaction features, along with temporal features of real-time intraoperative vital signs data, are input into a causal Transformer network. This causal Transformer network captures the lag effect of drug infusion through causal convolutional layers, models the long-range dependence of physiological indicators at different times using a multi-head self-attention mechanism, and outputs a predictive feature of the current anesthesia depth. A strategy for adjusting anesthetic drug dosage is generated based on a reinforcement learning algorithm. The reward function of the reinforcement learning algorithm includes a penalty term for deviation in anesthesia depth, a penalty term for hemodynamic fluctuations, and a clinical safety constraint term. The dynamic adjustment of the anesthetic dosage is achieved by optimizing the reward function.

2. The method for predicting anesthetic drug dosage based on machine learning according to claim 1, characterized in that, The steps for constructing the dynamic hypergraph model include: mapping different categories of data to corresponding nodes, where drug nodes correspond to target anesthetic drugs, gene nodes correspond to key physiological indicators affecting drug metabolism, metabolite nodes correspond to key metabolites in drug metabolism pathways, and environmental nodes correspond to operating room temperature and humidity parameters; establishing hyperedges connecting drug nodes and related gene nodes, the connection relationships of which are determined based on clinically known correlations between liver and kidney function and drug metabolism, characterizing the regulatory role of specific genes on the activity of drug-metabolizing enzymes; and establishing dynamic weighted edges connecting metabolite nodes and environmental nodes, the initial weights of which are set based on experimental data on the stability of metabolites from environmental factors and dynamically adjusted during model training using a gradient descent algorithm.

3. The method for predicting anesthetic drug dosage based on machine learning according to claim 1, characterized in that, The feature learning process of the dynamic hypergraph convolutional network includes: hierarchical feature aggregation of the hyperedge set to which each node belongs; transformation of node features within the hyperedge using a multilayer perceptron (MLP); the input of the MLP is the sum of feature vectors of all nodes within the hyperedge, and the output is the feature vector of the current node after hypergraph convolution; and dynamic weights between nodes within the hyperedge are calculated through an attention mechanism to adaptively adjust the contribution of different nodes to the feature of the target node. The calculation process of the attention mechanism includes linearly transforming the feature vectors of the target node and its neighboring nodes and inputting them into the LeakyReLU activation function, and then normalizing them through the softmax function to obtain the attention weights between nodes.

4. The method for predicting anesthetic drug dosage based on machine learning according to claim 1, characterized in that, The input data of the causal Transformer network includes multimodal feature vectors of the current time and the past n time steps. The kernel size of the causal convolutional layer is n+1. Padding operation is used to ensure that the length of the output sequence is consistent with the length of the input sequence. The multi-head self-attention mechanism includes at least 8 attention heads. Each attention head independently calculates the query vector, key vector and value vector, and the output results of each attention head are integrated through concatenation operation.

5. The method for predicting anesthetic drug dosage based on machine learning according to claim 3, characterized in that, The attention weights between nodes in the dynamic hypergraph convolutional network are calculated as follows: ； in, Indicates the superedge Middle node With nodes Attention weights and The weight matrix is ​​a learnable matrix. and They represent the first Nodes in hypergraph convolution and nodes eigenvectors, Indicates the superedge All nodes within, This is the activation function.

6. The method for predicting anesthetic drug dosage based on machine learning according to claim 4, characterized in that, The causal convolutional layer output features of the causal Transformer network are calculated as follows: ； in, Indicates the current time Causal convolution output features The kernel size is [size]. For the first Each convolutional kernel weight, Indicates lag Input characteristics at time step The preset drug infusion lag time constant is determined using pharmacokinetic experimental data.

7. The method according to claim 1, characterized in that, The reward function expression of the reinforcement learning algorithm is: ； in, Indicates the current time The reward value, , , The preset weighting coefficients, This is the current bispectral index measurement value. The bispectral index of the electroencephalogram (EEG) corresponding to the target depth of anesthesia. This represents the difference between the mean arterial pressure at the current time and the mean arterial pressure at the previous time. This represents the current dosage of anesthetic drugs. Within the clinically safe dosage range, This is an indicator function that outputs 1 when the dose exceeds the safe range, and 0 otherwise.

8. A machine learning-based anesthetic drug dosage prediction system, applied to the machine learning-based anesthetic dosage prediction method as described in any one of claims 1-7, characterized in that, include: The data acquisition module is configured to acquire preoperative genomic data, metabolomics data, intraoperative real-time vital signs data, and operating room environmental data of the patient. The genomic data is acquired through a gene sequencer, the metabolomics data is obtained through mass spectrometry, the intraoperative real-time vital signs data is acquired in real time through a monitor, and the operating room environmental data is acquired through a sensor network. The dynamic hypergraph construction module is configured to construct a dynamic hypergraph model containing drug nodes, gene nodes, metabolite nodes, and environmental nodes. The drug nodes and gene nodes are connected by hyperedges, and the metabolite nodes and environmental nodes are connected by dynamic weighted edges. The feature learning module is configured to perform feature learning on the dynamic hypergraph model through a dynamic hypergraph convolutional network to extract high-order interaction features between multimodal data; temporal sequence The modeling module is configured to input the higher-order interaction features and intraoperative real-time vital signs data into a causal Transformer network. The causal Transformer network outputs anesthesia depth prediction features through causal convolutional layers and multi-head self-attention mechanisms. The dose adjustment module is configured to generate anesthesia drug dose adjustment strategies based on a reinforcement learning algorithm. The reinforcement learning algorithm achieves dynamic adjustment of the anesthesia dose by optimizing a reward function that includes anesthesia depth deviation penalty term, hemodynamic fluctuation penalty term, and clinical safety constraint term.

9. The machine learning-based anesthetic drug dosage prediction system according to claim 8, characterized in that, The dynamic hypergraph convolutional network includes at least three hypergraph convolutional layers. Each hypergraph convolutional layer contains a hyperedge attention calculation unit and a feature aggregation unit. The hyperedge attention calculation unit is used to calculate the dynamic attention weights between nodes within the hyperedge. The feature aggregation unit aggregates the node features within the hyperedge using a multilayer perceptron. The causal Transformer network includes at least two encoder layers. Each encoder layer contains a causal convolutional sublayer and a multi-head self-attention sublayer. The output of the causal convolutional sublayer serves as the input of the multi-head self-attention sublayer.

10. The machine learning-based anesthetic drug dosage prediction system according to claim 8, characterized in that, It also includes an interpretability analysis module, configured to calculate the contribution of each input feature to the anesthetic dose prediction result using the SHAP value analysis method, and generate a visualized feature importance ranking chart. The input features include gene polymorphism data of liver function indicators and kidney function indicators; metabolic concentration data of anesthetic drugs such as remifentanil citrate injection and propofol; real-time intraoperative vital signs data including BIS value, MAP value, SpO2 value, BP value, HR value, body temperature value, muscle relaxation monitoring value, and inhaled anesthetic drug concentration (MAC); operating room temperature and humidity parameters; the hardware architecture of the system includes a server cluster, which contains a graphics processing unit (GPU) and a central processing unit (CPU), and the data acquisition module is connected to the server cluster via a wired network and / or a wireless network.

Images