AI-based anesthetic dosage personalized prediction system and method

Abstract

The invention relates to the technical field of anesthetic dosage prediction, in particular to an AI-based anesthetic dosage personalized prediction system and method. The data analysis module is in communication connection with the data acquisition module and receives the preprocessed data sent by the data acquisition module; based on the preprocessed data, a machine learning algorithm is used for analysis, and a patient risk assessment result is generated; the prediction module is in communication connection with the data analysis module and receives a patient risk assessment result sent by the data analysis module; according to a patient risk assessment result, combining physiological features, medical history and drug response data characteristics of the patient, and utilizing a machine learning algorithm to predict the maximum safe dose of the personalized drug dose; the output module is in communication connection with the prediction module and receives the personalized drug dosage prediction result sent by the prediction module; and the personalized drug dosage prediction result is output to an anesthetist, so that the intelligence and personalization of the anesthesia process are realized, and the safety and effectiveness of anesthesia are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of anesthesia dosage prediction, and in particular to an AI-based personalized anesthetic drug dosage prediction system and method thereof. Background Art

[0002] With the continuous advancement of medical technology, anesthesia plays an increasingly important role in modern surgery. However, personalized and precise control of anesthetic drug dosage has always been a major challenge in clinical practice. Traditional anesthesia methods mainly rely on the experience of anesthesiologists and standardized dosing regimens. This method is difficult to fully consider the individual differences of patients and can easily lead to problems such as insufficient or excessive anesthesia depth.

[0003] In recent years, with the rapid development of artificial intelligence technology, the application of AI to anesthetic drug dosage prediction has become a new research hotspot. Some existing methods attempt to use machine learning algorithms to predict anesthetic drug dosage, but these methods still have some obvious limitations. First, most existing methods only consider limited patient characteristics, such as age, weight, and gender, and ignore other important factors that may affect the effect of anesthesia, such as the patient's medical history, drug response, and real-time physiological state. Secondly, these methods usually use a single machine learning model, which makes it difficult to capture the complex nonlinear relationships and time series characteristics during anesthesia. Furthermore, existing methods often lack the ability to monitor and adjust in real time, and cannot respond to changes in patient status during surgery in a timely manner. Finally, these methods are also insufficient in the interpretability of prediction results and the quantification of uncertainty, making it difficult to gain full trust and adoption from anesthesiologists.

[0004] In view of these shortcomings of existing technologies, there is an urgent need for a personalized anesthetic drug dosage prediction system that can comprehensively consider individual patient characteristics, achieve accurate prediction and real-time adjustment, and have good interpretability. Summary of the invention

[0005] The present invention aims to solve the above technical problems and provide an AI-based personalized prediction system and method for anesthetic drug dosage. The system integrates multi-source heterogeneous data and combines advanced machine learning and deep learning algorithms to achieve accurate prediction and real-time adjustment of anesthetic drug dosage.

[0006] The present invention proposes an AI-based personalized anesthetic drug dosage prediction system, comprising:

[0007] Data acquisition module for:

[0008] Obtaining patients’ physiological characteristics data, medical history data, and drug response data;

[0009] Preprocess the acquired data to form a unified format and extract key features;

[0010] A data analysis module is connected to the data acquisition module for:

[0011] Receiving the preprocessed data sent by the data acquisition module;

[0012] Based on the preprocessed data, using a machine learning algorithm to analyze and generate a patient risk assessment result;

[0013] The prediction module is in communication with the data analysis module and is used to:

[0014] Receiving the patient risk assessment result sent by the data analysis module;

[0015] Based on the patient risk assessment results, combined with the patient's physiological characteristics, medical history and drug response data characteristics, a machine learning algorithm is used to predict the maximum safe dose of personalized drug dosage;

[0016] An output module, in communication with the prediction module, is used to:

[0017] Receiving the personalized drug dosage prediction result sent by the prediction module;

[0018] The personalized drug dosage prediction result is output to the anesthesiologist.

[0019] Preferably, the data acquisition module comprises:

[0020] A physiological data acquisition unit, used to collect physiological characteristic data of the patient, wherein the physiological characteristic data includes the patient's age, weight, height, body temperature, heart rate, respiratory rate, blood pressure and pre-anesthesia emotional data;

[0021] A medical history data collection unit is used to collect the patient's medical history data, the medical history data including liver function, kidney function, heart function, nervous system diseases, respiratory system diseases, hypertension, diabetes, asthma and drug allergy history;

[0022] A drug reaction data collection unit, used to collect drug reaction data of patients, wherein the drug reaction data includes sedation status after patients use sedative drugs, muscle relaxation status after patients use muscle relaxants, and anesthesia status after patients use anesthetic drugs;

[0023] The data preprocessing unit is used to preprocess the collected data, including data cleaning, missing value processing and feature extraction.

[0024] Preferably, the data analysis module comprises:

[0025] A feature selection unit, used for selecting features related to anesthetic drug dosage prediction from the preprocessed data;

[0026] A machine learning model training unit, used to train a machine learning model based on the selected features, wherein the machine learning model includes a support vector machine (SVM), a random forest, and a neural network;

[0027] The risk assessment unit is used to use the trained machine learning model to perform risk assessment on patients and generate risk assessment results.

[0028] Preferably, the prediction module comprises:

[0029] Pharmacokinetic model unit, used to establish a patient-specific pharmacokinetic model;

[0030] Pharmacodynamic model unit, used to establish patient-specific pharmacodynamic models;

[0031] The dosage prediction unit is used to predict the maximum safe dosage of personalized drug dosage based on the drug metabolism kinetic model, the pharmacodynamic model and the patient risk assessment result.

[0032] Preferably, it further comprises a real-time monitoring module, which is in communication with the output module and is used for:

[0033] Receive real-time physiological data of patients during anesthesia;

[0034] Based on the real-time physiological data and the predicted personalized drug dosage, using a deep reinforcement learning algorithm to perform real-time risk monitoring;

[0035] When an abnormality is detected, an alert is sent to the anesthesiologist.

[0036] Preferably, the real-time monitoring module uses a recurrent neural network (RNN) and a long short-term memory (LSTM) network to construct a deep reinforcement learning model, and the deep reinforcement learning model includes:

[0037] a state encoder for encoding the patient's real-time physiological data and predicted drug dosage into a state vector;

[0038] Action generator, used to generate the optimal action for the next step based on the current state vector;

[0039] A reward calculator for calculating a reward value according to the patient's physiological state and anesthesia effect;

[0040] The policy network is used to update the action policy based on the state vector and reward value.

[0041] Preferably, the data acquisition module further includes a multimodal data fusion unit for fusing data from different sources, including:

[0042] Time-align the electronic medical record data in the hospital information system HIS with the physiological data collected in real time;

[0043] Use the attention mechanism to assign different weights to data of different modalities;

[0044] A multimodal deep learning model is used to extract fusion features.

[0045] Preferably, the prediction module further comprises an uncertainty quantification unit, configured to:

[0046] Quantify the uncertainty of prediction results using Bayesian neural networks;

[0047] Generate confidence intervals for predictions;

[0048] When the uncertainty of the prediction result exceeds the preset threshold, the anesthesiologist is prompted to perform manual intervention.

[0049] Preferably, a knowledge graph module is further included, which is in communication connection with the data analysis module and the prediction module and is used for:

[0050] Build a knowledge graph in the field of anesthesia, including information on drug interactions, contraindications, and side effects;

[0051] Associating patient data with the knowledge graph to provide additional semantic information;

[0052] Graph neural networks are used to reason about knowledge graphs to assist in dose prediction and risk assessment.

[0053] The AI-based personalized prediction method for anesthetic drug dosage based on the system includes the following steps:

[0054] S1: Obtain the patient's physiological characteristics data, medical history data and drug reaction data, pre-process the acquired data, form a unified format and extract key features;

[0055] S2: Analyze the pre-processed data obtained in step S1 using a machine learning algorithm to generate a patient risk assessment result;

[0056] S3: Based on the patient risk assessment results obtained in step S2, combined with the patient's physiological characteristics, medical history and drug response data characteristics, a machine learning algorithm is used to predict the maximum safe dose of personalized drug dosage;

[0057] S4: outputting the personalized drug dose prediction result obtained in step S3 to the anesthesiologist;

[0058] S5: During anesthesia, the patient's real-time physiological data is received, and based on the real-time physiological data and the predicted personalized drug dosage, a deep reinforcement learning algorithm is used to perform real-time risk monitoring, and an alarm is issued to the anesthesiologist when an abnormal situation is detected.

[0059] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0060] The AI-based personalized prediction system and method for anesthetic drug dosage of the present invention have significant technical advantages and clinical application value. From a macro perspective, the system realizes the intelligence and personalization of the anesthesia process, greatly improves the safety and effectiveness of anesthesia, and provides a new paradigm for the development of precision medicine.

[0061] From the perspective of system architecture, the present invention adopts a modular design, and each functional module works together to form a complete closed-loop system. The organic combination of data acquisition module, data analysis module, prediction module and output module ensures the intelligence of the whole process from data input to result output. This design not only improves the efficiency of the system, but also enhances its scalability and adaptability.

[0062] In terms of data processing, this invention breaks through the limitation of traditional methods that only consider limited features. Through multimodal data fusion technology, it fully integrates multi-dimensional information such as the patient's physiological characteristics, medical history, and drug response. This comprehensive data collection and processing method lays a solid foundation for subsequent accurate predictions.

[0063] In terms of prediction algorithms, the present invention innovatively adopts a method that combines ensemble learning and deep learning. By integrating multiple machine learning models, such as support vector machines, random forests, and neural networks, the system can better capture the complex nonlinear relationships during anesthesia. At the same time, the introduction of deep learning technology, especially recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), enables the system to effectively process time series data and capture the dynamic changes in patient status.

[0064] Another major innovation of the present invention is the realization of real-time monitoring and dynamic adjustment. Through the deep reinforcement learning algorithm, the system can continuously optimize the anesthesia strategy according to the patient's real-time physiological data and realize closed-loop control. This dynamic adjustment capability greatly improves the accuracy and safety of anesthesia.

[0065] In terms of interpretability, the present invention introduces knowledge graph technology to organically combine domain expert knowledge with machine learning models. This not only improves the accuracy of predictions, but also enhances the interpretability of the system, which helps to gain the trust and acceptance of clinicians.

[0066] In addition, the present invention also innovatively introduces uncertainty quantification technology. Through the Bayesian neural network, the system can provide a confidence interval for each prediction result, helping anesthesiologists better assess risks and make more cautious and reasonable decisions.

[0067] In summary, the AI-based personalized prediction system for anesthetic drug dosage and its method of the present invention have innovative breakthroughs in data processing, algorithm design, real-time adjustment, interpretability, and uncertainty quantification. These innovations complement and synergize each other to jointly build a comprehensive, accurate, and reliable anesthesia decision support system. This system can not only significantly improve the safety and effectiveness of anesthesia, but also is expected to promote anesthesia to develop in a more intelligent and personalized direction, and provide patients with better medical services. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a logic block diagram of the entire system of the present invention.

[0069] Figure 2 It is a logic block diagram of the data acquisition module of the present invention.

[0070] Figure 3 It is a logic block diagram of the data analysis module of the present invention.

[0071] Figure 4 It is a logic block diagram of the prediction module of the present invention.

[0072] Figure 5 It is a logic block diagram of the real-time monitoring module of the present invention.

[0073] Figure 6 This is a logical block diagram of the knowledge graph module of the present invention. DETAILED DESCRIPTION

[0074] See also Figure 1-6 The present invention provides an AI-based personalized prediction system and method for anesthetic drug dosage. The system can accurately predict the optimal dosage of anesthetic drugs according to the individual characteristics and real-time status of the patient, thereby improving the anesthesia effect and patient safety.

[0075] First, the system of the present invention includes a data acquisition module 1, a data analysis module 2, a prediction module 3 and an output module 4. The data acquisition module 1 is used to obtain the patient's physiological characteristic data, medical history data and drug reaction data, and pre-process these data to form a unified format and extract key features. Preferably, the physiological characteristic data includes the patient's age, weight, height, body temperature, heart rate, respiratory rate, blood pressure and pre-anesthesia emotional data. The collection of these data can be completed through the hospital's existing electronic medical record system or special physiological monitoring equipment.

[0076] Data preprocessing is a key step, which includes data cleaning, missing value processing and feature extraction. For example, for missing values, the present invention uses multiple interpolation method for processing:

[0077]

[0078] in, is the estimated value of the missing value, is the mth possible filling value, and k is the number of generated data sets. The feature extraction uses the principal component analysis (PCA) method, which can effectively reduce the data dimension and extract the most representative features.

[0079] This method can generate multiple possible filling values ​​and comprehensively consider the uncertainty of these values. The feature extraction uses the principal component analysis (PCA) method, which can effectively reduce the data dimension and extract the most representative features.

[0080] The data analysis module 2 is in communication with the data acquisition module 1, and is used to receive the pre-processed data, and analyze it using a machine learning algorithm to generate a patient risk assessment result. The present invention adopts an integrated learning method, combining multiple machine learning algorithms, such as support vector machines (SVM), random forests, and neural networks, to improve the accuracy and robustness of the analysis.

[0081] Taking the random forest algorithm as an example, its basic principle can be expressed as:

[0082]

[0083] in, is the final prediction result, B is the number of decision trees, T b (x) is the prediction result of the bth decision tree.

[0084] The prediction module 3 is in communication with the data analysis module 2, and is used to receive the patient risk assessment results, and based on these results, combined with the patient's physiological characteristics, medical history and drug reaction data characteristics, use a machine learning algorithm to predict the maximum safe dose of personalized drug dosage. In one embodiment of the present invention, the prediction module 3 uses a deep neural network model, and the structure of the model is as follows:

[0085] h1=σ(W1x+b1),

[0086] h2=σ(W2h1+b2),

[0087] y=W3h2+b3,

[0088] Among them, x is the input feature, h1 and h2 are hidden layers, y is the output layer (predicted drug dosage), W and b are the weight matrix and bias vector respectively, and σ is the activation function (such as ReLU function).

[0089] The output module 4 is in communication with the prediction module 3, and is used to receive the personalized drug dosage prediction results and output these results to the anesthesiologist. Preferably, the output module 4 also includes a visualization unit, which can intuitively display the prediction results in the form of charts, so that the anesthesiologist can quickly understand and make decisions.

[0090] Furthermore, the system of the present invention also includes a physiological data acquisition unit 11, a medical history data acquisition unit 12, a drug reaction data acquisition unit 13 and a data preprocessing unit 14 in the data acquisition module 1. These units work together to ensure the comprehensiveness and accuracy of the data.

[0091] The physiological data acquisition unit 11 is specifically used to collect the physiological characteristic data of the patient. For example, for the collection of body temperature, an infrared thermometer can be used, and the measurement accuracy can reach ±0.1°C. The collection of heart rate and blood pressure can use a multi-parameter monitor to monitor the changes of these physiological indicators in real time.

[0092] The medical history data collection unit 12 is responsible for collecting the patient's medical history data. These data usually come from the hospital's electronic medical record system, including the patient's past disease records, surgical experience, etc. In one embodiment of the present invention, the unit also includes a natural language processing module that can extract key information from unstructured medical record text.

[0093] The drug response data acquisition unit 13 is used to collect the drug response data of the patient. These data can be obtained by an anesthesia depth monitoring instrument, such as an electroencephalogram (EEG) monitoring device. For example, the bispectral index (BIS) can be used to evaluate the patient's anesthesia depth, and the BIS value range is 0-100, where 40-60 is generally considered to be an appropriate general anesthesia depth.

[0094] The data preprocessing unit 14 preprocesses the collected data. In addition to the missing value processing and feature extraction mentioned above, this unit also performs data standardization. For example, the Z-score standardization method can be used:

[0095]

[0096] Among them, x is the original data, μ is the mean, and σ is the standard deviation.

[0097] The data analysis module 2 includes a feature selection unit 21, a machine learning model training unit 22 and a risk assessment unit 23. The feature selection unit 21 uses a recursive feature elimination (RFE) algorithm, which repeatedly builds a model (such as SVM) and selects the best features (or feature combinations) to ultimately determine the optimal feature subset. The machine learning model training unit 22 is responsible for training multiple machine learning models. Taking the support vector machine (SVM) as an example, its objective function can be expressed as:

[0098]

[0099] subjecttoy i (w T x i +b)≥1-ξ i and i ≥0, i=1,...,n, where w is the weight vector, b is the bias term, ξ i is the slack variable and C is the penalty parameter.

[0100] By solving this optimization problem, SVM can find an optimal separating hyperplane for classification or regression tasks.

[0101] The risk assessment unit 23 uses the trained model to perform risk assessment on the patient. In one embodiment of the present invention, the risk assessment results are divided into three levels: low, medium, and high. For example, if the predicted probability of adverse reactions is less than 5%, it is rated as low risk; if it is between 5% and 15%, it is medium risk; if it is greater than 15%, it is rated as high risk. This grading method can help anesthesiologists quickly determine the risk status of patients and take corresponding measures.

[0102] Through the above detailed description, it can be seen that the system of the present invention makes full use of artificial intelligence and machine learning technology to achieve personalized and accurate prediction of anesthetic drug dosage. This can not only improve the anesthesia effect and reduce adverse drug reactions, but also provide decision support for anesthesiologists and improve surgical safety. At the same time, the modular design of this system also makes it have good scalability and adaptability, and can be adjusted and optimized according to actual needs.

[0103] The prediction module 3 of the present invention further includes a pharmacokinetic model unit 31, a pharmacodynamic model unit 32 and a dosage prediction unit 33. The coordinated work of these units enables the system to more accurately predict personalized drug dosages, taking into full account the impact of individual patient differences on drug metabolism and effects.

[0104] The pharmacokinetic model unit 31 is responsible for establishing a pharmacokinetic model for individual patients. In a preferred embodiment of the present invention, the unit uses a three-compartment model to describe the distribution and elimination process of the drug in the body. The model can be represented by the following differential equations:

[0105]

[0106] Among them, A1, A2 and A3 represent the amount of drug in the central compartment, shallow peripheral compartment and deep peripheral compartment respectively, k ij represents the transfer rate constant from chamber i to chamber j, and R(t) represents the drug administration rate.

[0107] The pharmacodynamic model unit 32 is used to establish a pharmacodynamic model for individual patients. max Model to describe the relationship between drug concentration and effect:

[0108]

[0109] Among them, E represents the drug effect, E0 is the baseline effect, and E m ax is the maximum effect, C is the drug concentration, EC 50 is the drug concentration that produces 50% of the maximum effect, and n is the Hill coefficient. By analyzing the patient's drug response data, individualized values ​​for these parameters can be determined.

[0110] The dose prediction unit 33 predicts the maximum safe dose of personalized drug dose based on the above two models and the patient risk assessment results. Preferably, the unit uses a Bayesian optimization algorithm to find the optimal dose. Its objective function can be expressed as:

[0111]

[0112] Among them, D opt is the optimal dose, E(D) is the expected effect function, R(D) is the risk function, and λ is the trade-off coefficient. By adjusting the λ value, a balance can be achieved between effect and safety.

[0113] The system of the present invention further comprises a real-time monitoring module 5, which is in communication connection with the output module 4. The introduction of the real-time monitoring module 5 enables the system to dynamically adjust the anesthesia scheme, further improving the accuracy and safety of anesthesia.

[0114] The real-time monitoring module 5 mainly performs three tasks: receiving the real-time physiological data of the patient during anesthesia, performing risk monitoring based on the real-time data and the predicted personalized drug dosage, and alerting the anesthesiologist when an abnormal situation is detected. In one embodiment of the present invention, the module uses a deep reinforcement learning algorithm for real-time risk monitoring.

[0115] Specifically, the real-time monitoring module 5 uses a recurrent neural network (RNN) and a long short-term memory network (LSTM) to construct a deep reinforcement learning model. The model includes a state encoder 51, an action generator 52, a reward calculator 53 and a strategy network 54.

[0116] The state encoder 51 is used to encode the patient's real-time physiological data and predicted drug dosage into a state vector. Assume that the input sequence is x1, x2, ..., x T , the core formula of LSTM is as follows:

[0117]

[0118] o t =σ(W o ·[h t-1 , x t ]+b o ),

[0119] h t =o t *tanh(C t ),

[0120] Among them, f t 、i t and t They are forget gate, input gate and output gate, C t is the cell state, h t is a hidden state.

[0121] The action generator 52 generates the optimal action for the next step based on the current state vector. In the scenario of adjusting the dose of anesthetic drugs, the action can be to increase, decrease or maintain the current dose. Preferably, the present invention adopts an Actor-Critic architecture, in which the Actor network is responsible for generating actions and the Critic network is responsible for evaluating the value of actions.

[0122] The reward calculator 53 calculates the reward value according to the patient's physiological state and the anesthesia effect. For example, the following reward function can be defined:

[0123] R=w1·S BIS +w2·S HR +w3·S Bp -w4 ·|DD target |,

[0124] Among them, S BIS , S HR and S Bp D and D represent the stability scores of BIS index, heart rate and blood pressure, respectively. target represent the actual dose and target dose respectively, wi is the weight coefficient.

[0125] The policy network 54 updates the action strategy according to the state vector and the reward value. The policy gradient method is used, and the update formula is:

[0126]

[0127] Among them, θ is the policy network parameter, α is the learning rate, and J(θ) is the performance objective function.

[0128] In this way, the real-time monitoring module 5 can continuously learn and optimize the anesthesia strategy to adapt to the individual differences of different patients and the dynamic changes during the operation.

[0129] In another embodiment of the present invention, the data acquisition module 1 further includes a multimodal data fusion unit 15. The introduction of this unit enables the system to capture the health status of the patient more comprehensively and improve the accuracy of the prediction.

[0130] The multimodal data fusion unit 15 mainly completes three tasks: time-aligning the electronic medical record data in the hospital information system (HIS) with the physiological data collected in real time, using the attention mechanism to assign different weights to data of different modalities, and using a multimodal deep learning model to extract fusion features.

[0131] Time alignment is the basis of data fusion. The present invention adopts the dynamic time warping (DTW) algorithm to solve the time inconsistency problem of different modal data. The goal of the DTW algorithm is to find the best alignment between two time series so that the distance between them is minimized.

[0132] The introduction of the attention mechanism enables the system to automatically learn the importance of data of different modalities. Assuming there are M modalities, the feature representation of the i-th modality is h i , then the fused feature h fusion It can be expressed as:

[0133]

[0134] Among them, α i is the attention weight of the ith modality, satisfying ∑ i =1 M α i =1.

[0135] Finally, the multimodal deep learning model is used to extract fusion features. The present invention adopts a multimodal variational autoencoder (MVAE), whose objective function is:

[0136]

[0137] Where x is the multimodal input, z is the potential representation, q(z|x) is the encoder, p(x|z) is the decoder, and D K L is the KL divergence.

[0138] In this way, the multimodal data fusion unit 15 can effectively integrate data from different sources, providing a richer and more reliable information basis for subsequent analysis and prediction. The prediction module 3 of the present invention also includes an uncertainty quantification unit 34, the introduction of which enables the system to better handle the uncertainty in the prediction process and provide more comprehensive decision support for anesthesiologists.

[0139] The uncertainty quantification unit 34 mainly completes three tasks: quantifying the uncertainty of the prediction results using the Bayesian neural network, generating a confidence interval for the prediction results, and prompting the anesthesiologist to perform manual intervention when the uncertainty of the prediction results exceeds a preset threshold.

[0140] In a preferred embodiment of the present invention, the Bayesian neural network uses a variational inference method to approximate the posterior distribution. Specifically, assuming that the prior distribution of the network weight w is p(w), given the data D, the posterior distribution p(w|D) can be approximated by minimizing the following variational free energy:

[0141]

[0142] where q(w) is the approximate posterior distribution and KL represents the Kullback-Leibler divergence. The optimal approximate posterior distribution q can be obtained * (w) ∈ Based on the obtained approximate posterior distribution, the uncertainty quantification unit 34 can generate a confidence interval for the prediction result. For a given input x, the mean and variance of the predicted output y can be estimated by the Monte Carlo sampling method:

[0143]

[0144] Where T is the number of sampling times, f(x, w t ) represents the output of the neural network, is the variance of each prediction.

[0145] Preferably, the present invention sets an uncertainty threshold. When the uncertainty of the prediction result (such as the prediction variance) exceeds this threshold, the system will automatically prompt the anesthesiologist to perform manual intervention. For example, the threshold can be set to 20% of the predicted mean. This mechanism can effectively reduce the risk in high uncertainty situations.

[0146] The system of the present invention further includes a knowledge graph module 6, which is in communication connection with the data analysis module 2 and the prediction module 3. The introduction of the knowledge graph module 6 enables the system to combine domain expert knowledge and improve the accuracy and interpretability of predictions.

[0147] The knowledge graph module 6 mainly completes three tasks: constructing a knowledge graph in the field of anesthesia, associating patient data with the knowledge graph, and using graph neural networks to reason about the knowledge graph.

[0148] In one embodiment of the present invention, the knowledge graph in the field of anesthesia contains information such as drug interactions, contraindications, and side effects. The knowledge graph can be represented as a set of triples (h, r, t), where h represents the head entity, r represents the relationship, and t represents the tail entity. For example, (propofol, inhibition, central nervous system) is a triple.

[0149] The process of associating patient data with the knowledge graph is actually an entity linking task. The present invention adopts an entity linking method based on the attention mechanism. Given a mention m in the patient data and a candidate entity e in the knowledge graph, the similarity between them can be expressed as:

[0150]

[0151] Among them, v m is the vector representation of mention, is the i-th related vector of entity e, α i is the attention weight, n is the number of related vectors, and cos represents the cosine similarity.

[0152] Reasoning on the knowledge graph is the core task of the knowledge graph module 6. The present invention uses a graph attention network (GAT) to achieve this function. For each node i in the knowledge graph, its updated representation h i It can be calculated by the following formula:

[0153]

[0154] in, is the neighbor set of node i, α i j is the attention coefficient, W is the weight matrix, and σ is the activation function. Attention coefficient α i The calculation formula for j is:

[0155]

[0156] Among them, a is a learnable attention vector and || represents the concatenation operation.

[0157] In this way, the knowledge graph module 6 can effectively integrate domain knowledge and provide more comprehensive and reliable support for personalized prediction of anesthetic drug dosage.

[0158] Finally, the present invention also provides an AI-based personalized prediction method for anesthetic drug dosage, which includes the following steps:

[0159] S1: Obtain the patient's physiological characteristics data, medical history data and drug reaction data, pre-process the acquired data, form a unified format and extract key features;

[0160] S2: Analyze the pre-processed data obtained in step S1 using a machine learning algorithm to generate a patient risk assessment result;

[0161] S3: Based on the patient risk assessment results obtained in step S2, combined with the patient's physiological characteristics, medical history and drug response data characteristics, a machine learning algorithm is used to predict the maximum safe dose of personalized drug dosage;

[0162] S4: outputting the personalized drug dose prediction result obtained in step S3 to the anesthesiologist;

[0163] S5: During anesthesia, the patient's real-time physiological data is received, and based on the real-time physiological data and the predicted personalized drug dosage, a deep reinforcement learning algorithm is used to perform real-time risk monitoring, and an alarm is issued to the anesthesiologist when an abnormal situation is detected.

[0164] In step S1, the present invention adopts a variety of data preprocessing techniques. For example, for missing values, multiple interpolation can be used for processing. The basic idea of ​​this method is to generate multiple possible filling values ​​and then comprehensively consider the uncertainty of these values. Specifically, assuming that there are m missing values ​​to be filled, k complete data sets can be generated, and the missing values ​​in each data set are estimated by a certain model (such as linear regression). Then, these k data sets can be analyzed separately, and finally the results can be merged.

[0165] In step S2, the present invention uses an ensemble learning method to perform patient risk assessment. Specifically, multiple base learners such as random forest, gradient boosting tree and support vector machine can be used, and then the final risk assessment result is obtained by voting or weighted averaging. This method can effectively reduce the deviation of a single model and improve the stability and accuracy of the prediction.

[0166] In step S3, a deep neural network model is used to predict personalized drug dosage. The input of the model includes the patient's physiological characteristics, medical history, drug response data, and risk assessment results, and the output is the predicted maximum safe dose. During the model training process, batch normalization and dropout techniques are used to improve the generalization ability of the model.

[0167] In step S4, the output of the prediction results uses visualization technology to present the predicted dose range, confidence interval and other information to the anesthesiologist in an intuitive way. This method can help anesthesiologists quickly understand the prediction results and make better decisions.

[0168] In step S5, real-time risk monitoring uses a LSTM-based recurrent neural network model. This model can effectively capture the temporal variation characteristics of the patient's physiological indicators, thereby timely discovering potential risks. When an abnormality is detected, the system will immediately alert the anesthesiologist and give corresponding suggestions.

[0169] Through the above steps, the method of the present invention can achieve personalized and accurate prediction of anesthetic drug dosage, greatly improving the safety and effectiveness of anesthesia. At the same time, each step of the method has strong interpretability, which helps anesthesiologists understand and trust the decision-making process of the AI ​​system.

[0170] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. AI-based personalized prediction system for anesthetic drug dosage, characterized by: include: Data acquisition module for: Obtaining patients’ physiological characteristics data, medical history data, and drug response data; Preprocess the acquired data to form a unified format and extract key features; A data analysis module is connected to the data acquisition module for: Receiving the preprocessed data sent by the data acquisition module; Based on the preprocessed data, using a machine learning algorithm to analyze and generate a patient risk assessment result; A prediction module, which is in communication with the data analysis module, is used to: Receiving the patient risk assessment result sent by the data analysis module; Based on the patient risk assessment results, combined with the patient's physiological characteristics, medical history and drug response data characteristics, a machine learning algorithm is used to predict the maximum safe dose of personalized drug dosage; An output module, in communication with the prediction module, is used to: Receiving the personalized drug dosage prediction result sent by the prediction module; The personalized drug dosage prediction result is output to the anesthesiologist.

2. The system according to claim 1, characterized in that The data acquisition module comprises: A physiological data acquisition unit, used to collect physiological characteristic data of the patient, wherein the physiological characteristic data includes the patient's age, weight, height, body temperature, heart rate, respiratory rate, blood pressure and pre-anesthesia emotional data; A medical history data collection unit is used to collect the patient's medical history data, the medical history data including liver function, kidney function, heart function, nervous system diseases, respiratory system diseases, hypertension, diabetes, asthma and drug allergy history; A drug reaction data collection unit, used to collect drug reaction data of patients, wherein the drug reaction data includes sedation status after patients use sedative drugs, muscle relaxation status after patients use muscle relaxants, and anesthesia status after patients use anesthetic drugs; The data preprocessing unit is used to preprocess the collected data, including data cleaning, missing value processing and feature extraction.

3. The system according to claim 1, characterized in that The data analysis module includes: A feature selection unit, used for selecting features related to anesthetic drug dosage prediction from the preprocessed data; A machine learning model training unit, used to train a machine learning model based on the selected features, wherein the machine learning model includes a support vector machine (SVM), a random forest, and a neural network; The risk assessment unit is used to use the trained machine learning model to perform risk assessment on patients and generate risk assessment results.

4. The system according to claim 1, characterized in that The prediction module comprises: Pharmacokinetic model unit, used to establish a patient-specific pharmacokinetic model; Pharmacodynamic model unit, used to establish patient-specific pharmacodynamic models; The dosage prediction unit is used to predict the maximum safe dosage of personalized drug dosage based on the drug metabolism kinetic model, the pharmacodynamic model and the patient risk assessment result.

5. The system according to claim 1, characterized in that It also includes a real-time monitoring module, which is in communication with the output module and is used to: Receive real-time physiological data of patients during anesthesia; Based on the real-time physiological data and the predicted personalized drug dosage, using a deep reinforcement learning algorithm to perform real-time risk monitoring; When an abnormality is detected, an alert is sent to the anesthesiologist.

6. The system according to claim 5, characterized in that The real-time monitoring module uses a recurrent neural network (RNN) and a long short-term memory (LSTM) network to construct a deep reinforcement learning model, which includes: a state encoder for encoding the patient's real-time physiological data and predicted drug dosage into a state vector; Action generator, used to generate the optimal action for the next step based on the current state vector; A reward calculator for calculating a reward value according to the patient's physiological state and anesthesia effect; The policy network is used to update the action policy based on the state vector and reward value.

7. The system according to claim 1, characterized in that The data acquisition module also includes a multimodal data fusion unit, which is used to fuse data from different sources, including: Time-align the electronic medical record data in the hospital information system HIS with the physiological data collected in real time; Use the attention mechanism to assign different weights to data of different modalities; A multimodal deep learning model is used to extract fusion features.

8. The system according to claim 1, characterized in that The prediction module further includes an uncertainty quantification unit, which is used to: Quantify the uncertainty of prediction results using Bayesian neural networks; Generate confidence intervals for predictions; When the uncertainty of the prediction result exceeds the preset threshold, the anesthesiologist is prompted to perform manual intervention.

9. The system according to claim 1, characterized in that It also includes a knowledge graph module, which is in communication with the data analysis module and the prediction module and is used to: Build a knowledge graph in the field of anesthesia, including information on drug interactions, contraindications, and side effects; Associating patient data with the knowledge graph to provide additional semantic information; Graph neural networks are used to reason about knowledge graphs to assist in dose prediction and risk assessment.

10. The AI-based personalized prediction method for anesthetic drug dosage based on the system according to any one of claims 1 to 9, characterized in that: The following steps are involved: S1: Obtain the patient's physiological characteristics data, medical history data and drug response data, pre-process the acquired data, form a unified format and extract key features; S2: Analyze the pre-processed data obtained in step S1 using a machine learning algorithm to generate a patient risk assessment result; S3: Based on the patient risk assessment results obtained in step S2, combined with the patient's physiological characteristics, medical history and drug response data characteristics, a machine learning algorithm is used to predict the maximum safe dose of personalized drug dosage; S4: outputting the personalized drug dose prediction result obtained in step S3 to the anesthesiologist; S5: During anesthesia, the patient's real-time physiological data is received, and based on the real-time physiological data and the predicted personalized drug dosage, a deep reinforcement learning algorithm is used to perform real-time risk monitoring, and an alarm is issued to the anesthesiologist when an abnormal situation is detected.

Images