Personalized anesthesia scheme generation method

Abstract

The invention relates to the technical field of anesthesia, and discloses a personalized anesthesia scheme generation method, which comprises the following steps: S1, extracting and coding patient features; s2, modeling a pharmacokinetic model; s3, drug concentration prediction; s4, applying a PD model; s5, performing uncertainty quantification; s6, predicting and outputting; s7, updating and feeding back parameters; s8, verifying the model; in the personalized anesthesia scheme generation process, firstly, effective feature extraction and coding are carried out on multi-modal data of a patient; according to the personalized anesthesia scheme generation method, efficient feature extraction is carried out on multi-modal data (such as electronic health records, genome information and real-time monitored vital signs) through VAE and other deep learning methods, and the differences of patients in the aspects of physique, drug metabolism capability, gene background and the like can be fully described; therefore, the optimal anesthetic medication strategy is customized for each patient.

Description

Technical Field

[0001] The present invention relates to the technical field of anesthesia, and specifically to a method for generating personalized anesthesia plans. Background Art

[0002] Anesthesia management is a crucial part in surgical operations, directly related to the safety of patients and the success of surgeries. Traditional anesthesia plans often rely on general models and empirical rules, making it difficult to fully consider individual differences among patients, such as age, weight, genetic background, and changes in real-time physiological indicators.

[0003] This "one-size-fits-all" approach may lead to poor anesthesia effects and even cause serious complications. Therefore, developing accurate and personalized anesthesia plan generation methods has become an important research direction in the medical field. In recent years, with the development of deep learning and physiological models, by combining multi-modal data (such as electronic health records, real-time vital sign monitoring data, genomic information, etc.) with pharmacokinetics (PK) and pharmacodynamics (PD) models, it is possible to accurately model and dynamically predict the patient's state, thereby generating personalized anesthesia plans and improving the accuracy of clinical decisions and the safety of patients. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for generating personalized anesthesia plans to solve the problems raised in the above background art, that is, traditional anesthesia plans often rely on general models and empirical rules, making it difficult to fully consider individual differences among patients, such as age, weight, genetic background, and changes in real-time physiological indicators. This "one-size-fits-all" approach may lead to poor anesthesia effects and even cause serious complications.

[0005] To achieve the above purpose, the present invention provides the following technical solution: A method for generating personalized anesthesia plans, the method comprising:

[0006] S1: Patient feature extraction and encoding;

[0007] S2: Pharmacokinetic model modeling;

[0008] S3: Drug concentration prediction;

[0009] S4: PD model application;

[0010] S5: Uncertainty quantification;

[0011] S6: Prediction output;

[0012] S7: Parameter update and feedback;

[0013] S8: Model verification;

[0014] In the generation of personalized anesthesia plans, it is first necessary to effectively extract and encode the multimodal data of patients. These multimodal data include electronic health records, real-time vital sign monitoring data, genomic information, laboratory test results, etc. Since these data have high dimensions and complex structures, a VAE is used as a feature encoder to map the high-dimensional data to a low-dimensional latent space and extract the key features of the patient. The VAE consists of two parts: an encoder and a decoder. The encoder converts the input data X into the probability distribution of the latent variable Z, which is usually assumed to be a multivariate normal distribution:

[0015]

[0016] In S1, z is the latent vector, representing the compressed features of the patient; μ(x) is the mean vector output by the encoder, representing the center of the latent space; σ 2 (x) is the variance vector output by the encoder, the logarithm of the variance of the diagonal covariance matrix, used to describe the degree of expansion of the latent space; is the multivariate normal distribution; through the reparameterization trick, the latent vector z is generated by the mean vector μ(x) and the variance vector

[0017] σ(x):

[0018]

[0019] where ⊙ represents element-wise multiplication, ∈ represents the noise vector of the standard normal distribution, used to introduce randomness so that the model can generate diverse latent vectors; the decoder reconstructs the input data through the latent variable

[0020] :

[0021]

[0022] μ dec (z) represents the mean vector output by the decoder, used to reconstruct the data; represents the variance vector output by the decoder, corresponding to the distribution of the reconstructed data; the training objective of the VAE is to minimize the reconstruction error and the KL divergence between the latent space distribution and the prior distribution:

[0023]

[0024] where represents the reconstruction loss, measuring the similarity between the reconstructed data and the original data; D KL (q(z|x)||p(z)) represents the KL divergence, measuring the difference between the latent distribution and the prior distribution;

[0025] Through training, the latent vector z generated by the VAE encoder contains the individualized feature information of the patient and is stored in the associative memory matrix M:

[0026] M = {z1, z2,..., z N}

[0027] where M represents the associative memory matrix that stores multiple latent vectors z i ; N represents the number of latent vectors stored in the memory matrix, and each z i corresponds to a patient or a specific clinical state. In addition, we directly predict the vital sign output through the decoder of the VAE as the basis for subsequent uncertainty quantification.

[0028] Preferably, in S2, a pharmacokinetic (PK) model is introduced to describe the absorption, distribution, metabolism, and excretion processes of anesthetic drugs in the body. The PK model adopts a multi-compartment model architecture, and its mathematical expression is:

[0029]

[0030] where is the rate of change of the drug concentration C(t) at time t; C(t) is the drug concentration vector at time t, representing the concentration in each body fluid compartment, with the unit of mg / L; K represents the drug transfer matrix, with the unit of 1 / min, describing the transfer rate and elimination rate of the drug between each body fluid compartment, and D(t) is the drug administration rate vector, with the unit of mg / min, representing the drug administration input to each body fluid compartment;

[0031] To achieve personalized pharmacokinetic description, the parameters of the PK model are dynamically adjusted through the latent vectors in the associative memory matrix

[0032] to adapt to the individual differences of different patients:

[0033] K = g K (z) = W K z + b K

[0034] where W K represents the weight matrix, which is obtained through training and learning and is used to map the latent vector z to the drug transfer matrix K; b K represents the bias vector, which is obtained through training and learning and is used to adjust the mapped transfer matrix K. Through the above mapping, the drug transfer matrix K of the PK model can generate a personalized transfer matrix according to the latent characteristics z of different patients, thereby reflecting the individual differences of patients in aspects such as drug absorption, distribution, metabolism, and excretion.

[0035] Preferably, after pharmacokinetic modeling, to predict the concentration of anesthetic drugs in the patient's body at future time steps, the system can ensure the safe and effective use of drugs. This prediction process closely depends on the aforementioned patient feature extraction and encoding, as well as the personalized adjustment module of the PK model. To predict the concentration of anesthetic drugs C(t + Δt) in the patient's body at future time steps, the system first uses the encoded latent vector and the personalized adjusted PK model parameters, and adopts the fourth-order Runge-Kutta method (Runge-Kutta 4th order, RK4) for numerical integration to calculate the change in drug concentration at future time steps, and gradually approaches the solution of the differential equation through piecewise integration steps. The specific steps are as follows:

[0036] k1 = Δt·(K·C(t) + D(t))

[0037]

[0038] k4 = Δt·(K·(C(t) + k3) + D(t))

[0039]

[0040] Where Δt is the time step, which determines the fineness of the prediction and the stability of the calculation;

[0041] k1, k2, k3, k4 are the intermediate steps in the RK4 method for approaching the solution of the differential equation. By iterating the above steps step by step, the drug concentration C(t + Δt) at multiple future time steps can be predicted, providing reliable data support for the subsequent PD model.

[0042] Preferably, the change curve of the anesthetic drug concentration at each future time step of the patient is obtained through the drug concentration prediction module, and a pharmacodynamics (PD) model is established. Combining the patient's historical vital sign information, the specific impact of the drug concentration change on the patient's current and future vital signs (such as heart rate, blood pressure, respiratory rate, etc.) is predicted. The PD model describes the relationship between the drug concentration C(t + Δt) and the vital sign index S(t). The PD model is constructed by a multi-layer non-linear function f, which can capture the complex impact of the drug on the vital signs:

[0043] S(t + Δt) = f(C(t + Δt), S(t); θ)

[0044] Where S(t) is the vital sign index vector at time t (such as heart rate, blood pressure, respiratory rate, blood oxygen saturation, etc.), C(t + Δt) is the drug concentration vector at time t + Δt, f is a non-linear function, (constructed by a multi-layer perceptron), θ: the parameter vector of the PD model, which determines the relationship between the drug concentration and the vital sign index.

[0045] Preferably, in order for the PD model to accurately predict changes in vital signs, it is necessary to effectively estimate and fit the model parameters θ. The specific steps are as follows:

[0046] Use the least squares method (Least Squares Method) or Bayesian optimization method to optimize the model parameters θ based on the training data set; use the mean squared error (Mean Squared Error, MSE) as the loss function to measure the difference between the model prediction value and the actual value:

[0047]

[0048] where N represents the number of samples; S i (t + Δt) represents the actual vital sign index of the i-th sample;

[0049] represents the predicted vital sign index of the i-th sample.

[0050] Preferably, in clinical applications, accurate vital sign prediction is crucial for patient safety. However, there is uncertainty in the prediction of any model. If this uncertainty is not reasonably quantified and managed, it may pose risks to clinical decisions. Therefore, the physical model layer particularly emphasizes uncertainty quantification. By methods such as Bayesian inference and Monte Carlo simulation, comprehensively evaluate and manage the uncertainty of the prediction to ensure the reliability and credibility of the prediction results. The Bayesian inference method calculates the posterior distribution of the model parameters by combining prior information and observed data, thereby modeling the uncertainty of the parameters. The specific steps are as follows:

[0051] 1. Define the prior distribution:

[0052] According to the latent vector z in the associative memory matrix, define the prior distributions of the PK model transition matrix K and the PD model parameters θ:

[0053]

[0054] where μ K (z) is the mean vector of the PK model transition matrix K, ∑ K (z): the covariance matrix of the PK model transition matrix K; μ θ (z) is the mean vector of the PD model parameters θ; ∑ θ (z): the covariance matrix of the PD model parameters θ

[0055] 2. Calculate the posterior distribution:

[0056] Combine the observed data C(t) and S(t) and calculate the posterior distribution through the Bayesian formula:

[0057] p(K, θ|C(t), S(t), z) ∝ p(C(t), S(t)|K, θ)·p(K|z)·p(θ|z)

[0058] where p(C(t), S(t)|K, θ) is the likelihood function, calculated based on the PK / PD model equation, which describes the probability of observing drug concentration and vital sign data given the parameters K and θ. Using the Markov chain Monte Carlo method, multiple parameter instances (K (m) , θ (m) ) are sampled from the posterior distribution, where m = 1, 2,..., M and M is the number of samplings; K (m) is the PK model transition matrix for the m-th sampling; θ (m) is the PD model parameter for the m-th sampling.

[0059] Preferably, the Monte Carlo simulation is used to generate a set of prediction results through a large number of random samplings, and then the statistical distribution of the prediction results is estimated to quantify the uncertainty. The specific steps are as follows:

[0060] 1. Parameter sampling:

[0061] A set of parameters (K (m) , θ (m) ) is randomly sampled from the posterior distribution obtained from Bayesian inference, where m = 1, 2,..., M and M is the number of samplings;

[0062] 2. Multiple predictions:

[0063] For each set of sampled parameters, the PK model is used to predict the drug concentration C (m) (t + Δt), and the PD model is used to predict the vital sign S (m) (t + Δt)

[0064] 3. Uncertainty estimation:

[0065] Based on the multiple prediction results, the mean and variance of the vital sign prediction are calculated to obtain the uncertainty distribution of the prediction results:

[0066]

[0067] where is the mean vector of the vital sign prediction, representing the average prediction result under all sampled parameters;

[0068] Var(S(t + Δt)) is the variance vector of the vital sign prediction, representing the uncertainty of the prediction results; Confidence interval construction:

[0069] Based on the mean and variance of the prediction results, a confidence interval for the vital sign prediction is constructed:

[0070]

[0071] Among them, 1.96 corresponds to a 95% confidence level; is the standard deviation of the prediction result, measuring the degree of dispersion of the prediction result.

[0072] Preferably, the vital sign prediction results directly obtained by using VAE are weighted and fused, and at the same time combined with uncertainty quantification to adjust the fusion weights. The specific method is as follows:

[0073] Weighted fusion:

[0074] S(t + Δt) = λ·S VAE (t + Δt) + (1 - λ)·S PD (t + ·Δt)

[0075] where λ is the fusion weight, which determines the proportion of the prediction results of the deep learning model and the PD model in the final output; S VAE (t + Δt) is the vital sign index predicted by the deep learning model; S PD (t + Δt) is the vital sign index predicted by the PD model.

[0076] Preferably, the vital sign prediction provides the predicted values and confidence intervals of the vital sign indexes at future time steps to help doctors identify potential physiological abnormalities in advance; the anesthesia plan generation is based on the predicted drug concentration and vital sign changes, and the system generates a specific anesthesia plan. The anesthesia plan includes the dosage, administration time, administration frequency of each anesthetic drug, and necessary adjustment strategies. The plan is presented in the form of a detailed schedule or operation instructions for anesthesiologists to refer to and execute in real time. These plans ensure the physiological stability and anesthesia effect of the patient during the entire surgical process, and provide personalized and dynamically adjusted treatment suggestions.

[0077] Preferably, during the surgical process, the vital signs and drug concentration data of the patient will change continuously, and it is necessary to dynamically adjust the PK / PD model parameters according to these real-time data to maintain the accuracy and adaptability of the prediction. During the surgical process, the vital signs of the patient (such as heart rate, blood pressure, respiratory rate, blood oxygen saturation, etc.) are collected in real time through monitoring devices, including multi-parameter monitors and continuous blood drug concentration monitors. In addition, the administration process of anesthetic drugs is controlled by an intelligent infusion pump, which can accurately record the administered amount and time of the drugs. These real-time data are transmitted to the central data processing system through wired or wireless networks to ensure the immediacy and accuracy of the data. Specifically, the drug concentration

[0078] C(t) and vital sign data S(t) of the patient are collected as the basis for updating the model parameters;

[0079] Using the Bayesian inference method, combined with real-time data C(t) and S(t), update the transition matrix K of the PK model and the parameter θ of the PD model to ensure that the model parameters always reflect the latest patient status:

[0080] p(K,θ|C(t),S(t),z) ∝ p(C(t),S(t)|K,θ)·p(K|z)·p(θ|z)

[0081] where p(C(t), S(t)|K,θ is the likelihood function, which describes the probability of observing drug concentration and vital sign data given the parameters K and θ; p(K|z) is the prior distribution of the PK model transition matrix, generated based on the latent vector z; p(θ|z) is the prior distribution of the PD model parameters, generated based on the latent vector z; through Markov chain Monte Carlo, multiple instances of the parameters K and θ are sampled from the posterior distribution for the prediction of drug concentration and vital signs.

[0082] Compared with the prior art, the beneficial effects of the present invention are:

[0083] The method for generating a personalized anesthesia plan extracts high - efficiency features from multi - modal data (such as electronic health records, genomic information, real - time monitored vital signs, etc.) through deep - learning methods such as VAE, which can fully characterize the differences in patients' physical constitution, drug metabolism ability, genetic background, etc., so as to customize the optimal anesthesia drug - using strategy for each patient. Dynamically adjust the PK / PD model parameters based on the individual latent vector, effectively reflecting the individual differences of different patients in the processes of drug absorption, distribution, metabolism, and excretion, making the prediction of the anesthesia plan more targeted and accurate; uncertainty quantification and safety improvement; use methods such as Bayesian inference and Monte Carlo simulation to quantify the uncertainty of model parameters and prediction results, helping clinicians intuitively understand the credibility and potential risks of the anesthesia plan, being more reliable in the early prediction of possible abnormalities, thus escorting the surgical safety and reducing the complication risks caused by overdose or under - dose of drugs; the organic combination of deep learning and physical models. Deep - learning models (VAE, neural networks, etc.) are good at capturing high - dimensional complex relationships, while the PK / PD model provides a solid pharmacological and physiological theoretical basis. The two complement each other, taking into account the interpretability and flexibility of the model. Compared with the practice of simply relying on machine learning or only based on traditional empirical formulas, this plan can not only provide higher - accuracy predictions, but also utilize the interpretability of the PK / PD model to help doctors better understand the transport and action mechanisms of drugs in patients' bodies; clinical decision - making support and efficiency improvement. The system can automatically generate a personalized anesthesia plan, including dosage, time, and frequency planning, significantly reducing the decision - making burden of anesthesiologists in complex situations. By providing real - time monitoring, prediction, and plan suggestions, it can enable medical staff to make correct anesthesia decisions faster, improve surgical efficiency, and show better clinical effects in the postoperative recovery period of patients and anesthesia quality management. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 It is a schematic structural diagram of the process of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0085] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0086] Please refer to Figure 1 , the present invention provides a technical solution: a method for generating a personalized anesthesia plan, the method includes:

[0087] S1: Patient feature extraction and encoding;

[0088] S2: Pharmacokinetic model building;

[0089] S3: Drug concentration prediction;

[0090] S4: PD model application;

[0091] S5: Uncertainty quantification;

[0092] S6: Prediction output;

[0093] S7: Parameter update and feedback;

[0094] S8: Model validation;

[0095] In the generation of personalized anesthesia plans, it is first necessary to effectively extract and encode the multimodal data of patients. These multimodal data include electronic health records, real-time vital sign monitoring data, genomic information, laboratory test results, etc. Due to the high dimensionality and complex structure of these data, VAE is used as a feature encoder to map the high-dimensional data to a low-dimensional latent space and extract the key features of the patient. VAE consists of two parts: an encoder and a decoder. The encoder converts the input data X into the probability distribution of the latent variable Z, which is usually assumed to be a multivariate normal distribution:

[0096]

[0097] In S1, z is the latent vector representing the compressed features of the patient; μ(x) is the mean vector output by the encoder, representing the center of the latent space; σ 2 (x) is the variance vector output by the encoder, the logarithm of the variance of the diagonal covariance matrix, used to describe the degree of expansion of the latent space; is the multivariate normal distribution; through the reparameterization trick, the latent vector z is generated by the mean vector μ(x) and the variance vector

[0098] σ(x):

[0099]

[0100] where ⊙ represents element-wise multiplication, ∈ represents the noise vector of the standard normal distribution, used to introduce randomness so that the model can generate diverse latent vectors; the decoder reconstructs the input data through the latent variable

[0101] :

[0102]

[0103] μ dec (z) represents the mean vector output by the decoder, used to reconstruct the data; Denotes the variance vector of the decoder output, corresponding to the distribution of the reconstructed data; the training objective of the VAE is to minimize the reconstruction error and the KL divergence between the latent space distribution and the prior distribution:

[0104]

[0105] where denotes the reconstruction loss, measuring the similarity between the reconstructed data and the original data; D KL (q(z|x)||p(z)) denotes the KL divergence, measuring the difference between the latent distribution and the prior distribution;

[0106] Through training, the latent vector z generated by the VAE encoder contains the individual characteristic information of the patient and is stored in the associative memory matrix M:

[0107] M = {z1, z2,..., z N}

[0108] where M denotes the associative memory matrix, storing multiple latent vectors Z i ; N denotes the number of latent vectors stored in the memory matrix, each Z i corresponds to a patient or a specific clinical state. In addition, we also directly predict the vital sign output through the decoder of the VAE as the basis for subsequent uncertainty quantification.

[0109] Furthermore, in S2, a pharmacokinetic (PK) model is introduced to describe the absorption, distribution, metabolism, and excretion processes of anesthetic drugs in the body. The PK model adopts a multi-compartment model architecture, and its mathematical expression is:

[0110]

[0111] where is the change rate of the drug concentration C(t) at time t; C(t) is the drug concentration vector at time t, representing the concentration in each body fluid compartment, with the unit of mg / L; K represents the drug transfer matrix, with the unit of 1 / min, describing the transfer rate and elimination rate of the drug between each body fluid compartment, and D(t) is the drug administration rate vector, with the unit of mg / min, representing the drug administration input to each body fluid compartment;

[0112] To achieve personalized pharmacokinetic description, the parameters of the PK model are dynamically adjusted through the latent vector z in the associative memory matrix to adapt to the individual differences of different patients:

[0113] K = g K (z) = W K z + b K

[0114] where W KDenotes the weight matrix, which is obtained through training and learning and is used to map the latent vector z to the drug transfer matrix K; b K Denotes the bias vector, which is obtained through training and learning and is used to adjust the mapped transfer matrix K. Through the above mapping, the drug transfer matrix K of the PK model can generate a personalized transfer matrix according to the latent characteristics z of different patients, thereby reflecting the individual differences of patients in aspects such as drug absorption, distribution, metabolism, and excretion.

[0115] Furthermore, after pharmacokinetic modeling, to predict the concentration of anesthetic drugs in the patient's body at future time steps, the system can ensure the safe and effective use of drugs. This prediction process closely depends on the aforementioned patient feature extraction and encoding and the personalized adjustment module of the PK model. To predict the concentration of anesthetic drugs C(t+Δt) in the patient's body at future time steps, the system first uses the encoded latent vector and the personalized adjusted PK model parameters, and adopts the fourth-order Runge-Kutta method (Runge-Kutta4thorder, RK4) for numerical integration to calculate the change in drug concentration at future time steps, and gradually approximates the solution of the differential equation through segmented integration steps. The specific steps are as follows:

[0116] k1 = Δt·(K·C(t)+D(t))

[0117]

[0118] k4 = Δt·(K·(C(t)+k3)+D(t))

[0119]

[0120] Where Δt is the time step, which determines the fineness of the prediction and the stability of the calculation; k1, k2, k3, k4 are the intermediate steps in the RK4 method for approximating the solution of the differential equation. By gradually iterating the above steps, the drug concentration C(t+Δt) at multiple future time steps can be predicted, providing reliable data support for the subsequent PD model.

[0121] Furthermore, the change curve of the anesthetic drug concentration at each future time step of the patient is obtained through the drug concentration prediction module, and a pharmacodynamics (PD) model is established. Combining the patient's historical vital sign information, the specific impact of the drug concentration change on the patient's current and future vital signs (such as heart rate, blood pressure, respiratory rate, etc.) is predicted. The PD model describes the relationship between the drug concentration C(t+Δt) and the vital sign index S(t). The PD model is constructed through a multi-layer non-linear function f and can capture the complex impact of the drug on the vital signs:

[0122] S(t+Δt) = f(C(t+Δt),S(t);θ)

[0123] Among them, S(t) is the vital sign index vector at time t (such as heart rate, blood pressure, respiratory rate, blood oxygen saturation, etc.), C(t+Δt) is the drug concentration vector at time t+Δt, f is a non-linear function (constructed by a multi-layer perceptron), and θ is the parameter vector of the PD model, which determines the relationship between the drug concentration and the vital sign index.

[0124] Furthermore, in order to enable the PD model to accurately predict the changes in vital signs, it is necessary to effectively estimate and fit the model parameter θ. The specific steps are as follows:

[0125] Use the least squares method (Least Squares Method) or the Bayesian optimization method to optimize the model parameter θ based on the training data set; use the mean squared error (Mean Squared Error, MSE) as the loss function to measure the difference between the model prediction value and the actual value:

[0126]

[0127] where N represents the number of samples; Si(t+Δt) represents the actual vital sign index of the i-th sample;

[0128] represents the predicted vital sign index of the i-th sample.

[0129] Furthermore, in clinical applications, accurate vital sign prediction is crucial for patient safety. However, there is uncertainty in the prediction of any model. If this uncertainty is not reasonably quantified and managed, it may pose risks to clinical decisions. Therefore, the physical model layer particularly emphasizes uncertainty quantification. By methods such as Bayesian inference and Monte Carlo simulation, the uncertainty of the prediction is comprehensively evaluated and managed to ensure the reliability and credibility of the prediction results. The Bayesian inference method calculates the posterior distribution of the model parameters by combining prior information and observed data, thereby modeling the uncertainty of the parameters. The specific steps are as follows:

[0130] 1. Define the prior distribution:

[0131] According to the latent vector z in the associative memory matrix, define the prior distribution of the PK model transition matrix K and the PD model parameter θ:

[0132]

[0133] where μ K (z) is the mean vector of the PK model transition matrix K, ∑ K (z): the covariance matrix of the PK model transition matrix K; μ θ (z) is the mean vector of the PD model parameter θ; ∑θ (z): The covariance matrix of the PD model parameter θ

[0134] 2. Calculate the posterior distribution:

[0135] Combining the observed data C(t) and S(t), calculate the posterior distribution through Bayes' formula:

[0136] p(K, θ|C(t), S(t), z) ∝ p(C(t), S(t)|K, θ)·p(K|z)·p(θ|z)

[0137] where p(C(t), S(t)|K, θ) is the likelihood function, calculated based on the PK / PD model equation, describing the probability of observing drug concentration and vital sign data given the parameters K and θ. Using the Markov chain Monte Carlo method, multiple parameter instances (K (m) , θ (m) ) are sampled from the posterior distribution, where m = 1, 2,..., M and M is the number of sampling times; K (m) is the PK model transition matrix for the m-th sampling; θ (m) is the PD model parameter for the m-th sampling.

[0138] Furthermore, the use of Monte Carlo simulation generates a set of prediction results through a large number of random samplings, and then estimates the statistical distribution of the prediction results to achieve the quantification of uncertainty. The specific steps are as follows:

[0139] 1. Parameter sampling:

[0140] Randomly sample a set of parameters (K (m) , θ (m) ) from the posterior distribution obtained by Bayesian inference, where m = 1, 2,..., M and M is the number of sampling times;

[0141] 2. Multiple predictions:

[0142] For each set of sampled parameters, use the PK model to predict the drug concentration C (m) (t + Δt), and predict the vital signs S (m) (t + Δt)

[0143] 3. Uncertainty estimation:

[0144] Based on the multiple prediction results, calculate the mean and variance of the vital sign predictions to obtain the uncertainty distribution of the prediction results:

[0145]

[0146] where is the mean vector of the vital sign predictions, representing the average prediction result under all sampled parameters;

[0147] Var(S(t+Δt)) is the variance vector of the vital sign prediction, representing the uncertainty of the prediction result; Confidence interval construction:

[0148] Based on the mean and variance of the prediction result, construct the confidence interval of the vital sign prediction:

[0149]

[0150] where 1.96: corresponds to a 95% confidence level; is the standard deviation of the prediction result, measuring the dispersion degree of the prediction result.

[0151] Furthermore, the vital sign prediction result directly obtained by using VAE is weighted and fused, and at the same time, uncertainty quantification is combined to adjust the fusion weight. The specific method is as follows:

[0152] Weighted fusion:

[0153] S(t+Δt) = λ·S VAE (t+Δt)+(1 - λ)·S PD (t+Δt)

[0154] where λ is the fusion weight, which determines the proportion of the prediction result of the deep learning model and the prediction result of the PD model in the final output; S VAE (t+Δt) is the vital sign index predicted by the deep learning model; S PD (t+Δt) is the vital sign index predicted by the PD model.

[0155] Furthermore, the vital sign prediction provides the predicted value of the vital sign index at future time steps and its confidence interval, helping doctors to identify potential physiological abnormalities in advance; The anesthesia plan generation is based on the predicted drug concentration and vital sign changes. The system generates a specific anesthesia plan, which includes the dosage, administration time, administration frequency of each anesthetic drug, and necessary adjustment strategies. The plan is presented in the form of a detailed schedule or operation instructions for anesthesiologists to refer to and execute in real time. These plans ensure the physiological stability and anesthesia effect of the patient during the entire surgical process, and provide personalized and dynamically adjusted treatment suggestions.

[0156] Furthermore, during the operation, the patient's vital signs and drug concentration data will continuously change, and it is necessary to dynamically adjust the PK / PD model parameters according to these real-time data to maintain the accuracy and adaptability of the prediction. During the operation, the patient's vital signs (such as heart rate, blood pressure, respiratory rate, blood oxygen saturation, etc.) are collected in real time through monitoring devices, including multi-parameter monitors and continuous blood drug concentration monitors. In addition, the administration process of anesthetic drugs is controlled by an intelligent infusion pump, which can accurately record the administered drug amount and time. These real-time data are transmitted to the central data processing system through wired or wireless networks to ensure the immediacy and accuracy of the data. Specifically, the patient's drug concentration C(t) and vital sign data S(t) are collected in real time as the basis for updating the model parameters;

[0157] Using the Bayesian inference method, combined with the real-time data C(t) and S(t), update the transition matrix K of the PK model and the parameter θ of the PD model to ensure that the model parameters always reflect the latest patient status:

[0158] p(K,θ|C(t),S(t),z)∝p(C(t),S(t)|K,θ)·p(K|z)·p(θ|z)

[0159] where p(C(t), S(t)|K, θ) is the likelihood function, which describes the probability of observing drug concentration and vital sign data given the parameters K and θ; p(K|z) is the prior distribution of the PK model transition matrix, generated based on the latent vector z; p(θ|z) is the prior distribution of the PD model parameters, generated based on the latent vector z; through Markov chain Monte Carlo, multiple instances of the parameters K and θ are sampled from the posterior distribution for the prediction of drug concentration and vital signs.

[0160] Working principle: Collect comprehensive patient information to provide basic data for personalized anesthesia plans, including the patient's basic information (such as age, gender, weight, height), medical history, allergy history, previous surgical records, heart rate, blood pressure, respiratory rate, blood oxygen saturation, body temperature, etc. Analyze the patient's genetic background, evaluate drug metabolism ability and potential drug reactions, blood tests (blood glucose, potassium, liver and kidney function, etc.), electrolyte levels, etc. Transform high-dimensional and multi-modal patient data into low-dimensional latent features for easy model processing, including data cleaning, standardization, missing value handling, and outlier detection to ensure data quality. Use a variational autoencoder (VAE) to encode multi-modal data into low-dimensional latent vectors and extract the patient's key features. These latent features can effectively represent the patient's individualized physiological and drug response characteristics. According to the patient's individual characteristics, adjust the kinetic parameters of anesthetic drugs in the body to ensure accurate prediction of drug concentration. Use the encoded latent vectors to generate a personalized drug transfer matrix through a pre-trained mapping function. According to the generated transfer matrix, adjust the PK model parameters to reflect the individual differences in drug absorption, distribution, metabolism, and excretion of the patient. Predict the anesthetic drug concentration in the patient's body at future time steps to ensure the safety and effectiveness of drug use. Input the current drug administration rate and initial drug concentration to generate the drug concentration curve at future time points to ensure that the drug concentration is maintained within a safe and effective range. Predict the impact of drug concentration changes on the patient's vital signs to ensure the stability and safety of the anesthesia effect. Combine the predicted drug concentration with the current and historical vital sign data to form an input vector, and through a deep neural network model, predict the vital sign indicators at future time steps to generate the trend of vital sign changes at future time points and give early warnings of possible abnormalities. Evaluate the uncertainty of the prediction results to improve the reliability of the prediction and provide a basis for clinical decision-making. Combine prior information and observed data to calculate the posterior distribution of model parameters, quantify parameter uncertainty, and based on the posterior distribution, perform multiple parameter samplings and predictions to estimate the mean and variance of the prediction results and construct a confidence interval to provide a confidence interval for vital sign prediction to help clinicians evaluate the credibility of the prediction results.

[0161] Finally, it should be noted that the above content is only used to illustrate the technical solution of the present invention, rather than a limitation on the protection scope of the present invention. Any simple modification or equivalent replacement of the technical solution of the present invention by those of ordinary skill in the art shall not depart from the essence and scope of the technical solution of the present invention.

Claims

1. A method for generating a personalized anesthesia plan, characterized in that: The method includes: S1: Patient feature extraction and encoding; S2: Pharmacokinetic model building; S3: Drug concentration prediction; S4: PD model application; S5: Uncertainty quantification; S6: Prediction output; S7: Parameter update and feedback; S8: Model validation; In the generation of personalized anesthesia protocols, it is first necessary to effectively extract and encode the multimodal data of the patient. These multimodal data include electronic health records, real-time vital sign monitoring data, genomic information, laboratory test results, etc. Since these data have high dimensions and complex structures, a VAE is used as the feature encoder to map the high-dimensional data into a low-dimensional latent space and extract the key features of the patient. The VAE consists of two parts: an encoder and a decoder. The encoder converts the input data X into the probability distribution of the latent variable Z, which is usually assumed to be a multivariate normal distribution: In S1, z is the latent vector, representing the compressed features of the patient; μ(x) is the mean vector output by the encoder, representing the center of the latent space; σ 2 (x) is the variance vector output by the encoder, the logarithm of the variance of the diagonal covariance matrix, used to describe the extent of the latent space; is the multivariate normal distribution; through the reparameterization trick, the latent vector z is generated from the mean vector μ(x) and the variance vector σ(x): where ⊙ represents element-wise multiplication, ∈ represents the noise vector of the standard normal distribution, which is used to introduce randomness so that the model can generate diverse latent vectors; the decoder reconstructs the input data through the latent variable z: μ dec (z) represents the mean vector output by the decoder and is used to reconstruct data; represents the variance vector output by the decoder, corresponding to the distribution of the reconstructed data; the training objective of the VAE is to minimize the reconstruction error and the KL divergence between the latent space distribution and the prior distribution: where represents the reconstruction loss, which measures the similarity between the reconstructed data and the original data; D KL (q(z|x)||p(z)) represents the KL divergence, which measures the difference between the latent distribution and the prior distribution; Through training, the latent vector z generated by the VAE encoder contains the individualized feature information of the patient and is stored in the associative memory matrix M: M = {z1, z2,..., z N} Where M represents an associative memory matrix that stores multiple latent vectors z i ; N represents the number of latent vectors stored in the memory matrix, and each z i corresponds to a patient or a specific clinical state.

2. The personalized anesthesia plan generation method according to claim 1, wherein: In S2, a pharmacokinetic (PK) model is introduced to describe the absorption, distribution, metabolism, and excretion processes of anesthetic drugs in the body. The PK model adopts a multi-compartment model architecture, and its mathematical expression is: where is the rate of change of the drug concentration C(t) at time t; C(t) is the drug concentration vector at time t, representing the concentration in each body fluid compartment, with the unit of mg / L; K represents the drug transfer matrix, with the unit of 1 / min, describing the transfer rate and elimination rate of the drug between body fluid compartments, and D(t) is the drug administration rate vector, with the unit of mg / min, representing the drug administration input to each body fluid compartment; To achieve personalized pharmacokinetic description, the parameters of the PK model are dynamically adjusted through the latent vector z in the associative memory matrix to adapt to the individual differences of different patients: K = g K (z) = W K z + b K Among which W K represents a weight matrix, obtained through training and learning, and is used to map the latent vector z to the drug transfer matrix K; b K represents a bias vector, obtained through training and learning, and is used to adjust the mapped transfer matrix K.

3. The personalized anesthesia plan generation method according to claim 1, characterized in that: After pharmacokinetic modeling, to predict the concentration of anesthetic drugs in the patient's body at future time steps, the system can ensure the safe and effective use of the drugs. This prediction process closely depends on the aforementioned patient feature extraction and encoding and the personalized adjustment module of the PK model. To predict the concentration of anesthetic drugs C(t + Δt) in the patient's body at future time steps, the system first uses the encoded latent vector and the personalized adjusted PK model parameters, and adopts the fourth-order Runge-Kutta method (Runge-Kutta4thorder, RK4) for numerical integration to calculate the change in drug concentration at future time steps, and gradually approximates the solution of the differential equation through piecewise integration steps. The specific steps are as follows: k1 = Δt·(K·C(t) + D(t)) k4 = Δt·(K·(C(t) + k3) + D(t)) where Δt is the time step, which determines the fineness of the prediction and the stability of the calculation; k1, k2, k3, k4 are the intermediate steps used in the RK4 method to approximate the solution of the differential equation.

4. The personalized anesthesia plan generation method according to claim 1, wherein: The anesthetic drug concentration change curve of the patient at each future time step is obtained through the drug concentration prediction module, a pharmacodynamics (PD) model is established, and combined with the patient's historical vital sign information, the specific impact of drug concentration changes on the patient's current and future vital signs (such as heart rate, blood pressure, respiratory rate, etc.) is predicted. The PD model describes the relationship between the drug concentration C(t+Δt) and the vital sign index S(t). The PD model is constructed through a multi-layer non-linear function f and can capture the complex impact of drugs on vital signs: S(t+Δt) = f(C(t+Δt), S(t); θ) where S(t) is the vital sign index vector at time t (such as heart rate, blood pressure, respiratory rate, blood oxygen saturation, etc.), C(t+Δt) is the drug concentration vector at time t+Δt, f is a non-linear function (constructed by a multi-layer perceptron), and θ: the parameter vector of the PD model, which determines the relationship between the drug concentration and the vital sign index.

5. The personalized anesthesia plan generation method according to claim 4, characterized in that: In order to enable the PD model to accurately predict vital sign changes, it is necessary to effectively estimate and fit the model parameter θ. The specific steps are as follows: Adopt the least squares method or the Bayesian optimization method to optimize the model parameter θ based on the training data set; use the mean squared error (MSE) as the loss function to measure the difference between the model prediction value and the actual value: where N represents the number of samples; S i (t + Δt) represents the actual vital sign index of the i-th sample; Represents the predicted vital sign index of the i-th sample.

6. The personalized anesthesia plan generation method according to claim 1, characterized in that: Through methods such as Bayesian inference and Monte Carlo simulation, comprehensively evaluate and manage the uncertainty of the prediction to ensure the reliability and credibility of the prediction results. The Bayesian inference method calculates the posterior distribution of the model parameters by combining prior information and observed data, thereby modeling the uncertainty of the parameters. The specific steps are as follows:

1. Define the prior distribution: According to the latent vector z in the associative memory matrix, define the prior distributions of the PK model transition matrix K and the PD model parameter θ: where μ K (z) is the mean vector μ of the PK model transition matrix K, K (z): the covariance matrix Σ of the PK model transition matrix K; μ θ (z) is the mean vector of the PD model parameter θ; Σ θ (z): the covariance matrix of the PD model parameter θ 2. Calculate the posterior distribution: Combine the observed data C(t) and S(t) and calculate the posterior distribution through Bayes' formula: p(K, θ|C(t), S(t), z) ∝ p(C(t), S(t)|K, θ) · p(K|z) · p(θ|z) where \(p(C(t), S(t)|K, \theta)\) is the likelihood function, calculated based on the PK / PD model equation, which describes the probability of observing drug concentration and vital sign data given the parameters \(K\) and \(\theta\). Using the Markov chain Monte Carlo method, multiple parameter instances \((K (m) , \theta (m) )\) are sampled from the posterior distribution, where \(m = 1, 2, \ldots, M\) and \(M\) is the number of samplings; \(K (m) \) is the PK model transition matrix for the \(m\)-th sampling; \(\theta (m) \) is the PD model parameter for the \(m\)-th sampling.

7. The personalized anesthesia plan generation method according to claim 6, characterized in that: Use Monte Carlo simulation to generate a set of prediction results through a large number of random samplings, and then estimate the statistical distribution of the prediction results to achieve the quantification of uncertainty. The specific steps are as follows:

1. Parameter sampling: From the posterior distribution obtained by Bayesian inference, randomly sample a set of parameters (K (m) , θ (m) ), where m = 1, 2,..., M and M is the number of samplings; 2. Multiple predictions: For each set of sampling parameters, use the PK model to predict the drug concentration C (m) (t + Δ), and predict the vital signs S (m) (t + Δt) 3. Uncertainty estimation: Based on the multiple prediction results, calculate the mean and variance of the vital sign prediction to obtain the uncertainty distribution of the prediction results: Among them is the mean vector for vital sign prediction, representing the average prediction result under all sampling parameters; Var(S(t+Δt)) is the variance vector of the vital sign prediction, indicating the uncertainty of the prediction results; Confidence interval construction: Based on the mean and variance of the prediction results, construct the confidence interval of the vital sign prediction: Among them, 1.96 corresponds to a 95% confidence level; is the standard deviation of the prediction result, which measures the degree of dispersion of the prediction result.

8. The personalized anesthesia plan generation method according to claim 1, wherein: Directly use the vital sign prediction results obtained by VAE prediction for weighted fusion, and at the same time combine uncertainty quantification to adjust the fusion weights. The specific method is as follows: Weighted fusion: S(t + Δt) = λ·S(t) + (1 - λ)·S(t + Δt) VAE (t + Δt) PD (t + Δt) It should be noted that the original formula in seems to be incorrect. The corrected formula is provided in the translation. If this is not what you intended, please double-check the original content. where λ is the fusion weight, which determines the proportion of the prediction results of the deep learning model and the PD model in the final output; S VAE (t + Δt) is the vital sign index predicted by the deep learning model; S PD (t + Δt) is the vital sign index predicted by the PD model.

9. The personalized anesthesia plan generation method according to claim 1, wherein: Vital sign prediction provides predicted values of vital sign indicators and their confidence intervals at future time steps, helping doctors identify potential physiological abnormalities in advance; Anesthesia plan generation is based on the predicted drug concentration and changes in vital signs. The system generates a specific anesthesia plan, which includes the dosage, administration time, administration frequency of each anesthetic drug, and necessary adjustment strategies. The plan is presented in the form of a detailed schedule or operation instructions for anesthesiologists to refer to and execute in real time.

10. The personalized anesthesia plan generation method according to claim 1, wherein: During the operation, the vital signs and drug concentration data of the patient will change continuously. It is necessary to dynamically adjust the PK / PD model parameters according to these real-time data to maintain the accuracy and adaptability of the prediction. During the operation, the patient's vital signs (such as heart rate, blood pressure, respiratory rate, blood oxygen saturation, etc.) are collected in real time through monitoring devices, including multi-parameter monitors and continuous blood drug concentration monitors. In addition, the administration process of anesthetic drugs is controlled by an intelligent infusion pump, which can accurately record the administered amount and time of the drugs. These real-time data are transmitted to the central data processing system through wired or wireless networks to ensure the immediacy and accuracy of the data. Specifically, the drug concentration C(t) and vital sign data S(t) of the patient are collected in real time as the basis for updating the model parameters. Using the Bayesian inference method, combined with the real-time data C(t) and S(t), update the transition matrix K of the PK model and the parameter θ of the PD model to ensure that the model parameters always reflect the latest patient status: p(K, θ|C(t), S(t), z) ∝ p(C(t), S(t)|K, θ) · p(K|z) · p(θ|z) where p(C(t), S(t)|K, θ) is the likelihood function, which describes the probability of observing drug concentration and vital sign data given the parameters K and θ; p(K|z) is the prior distribution of the PK model transition matrix, generated based on the latent vector z; p(θ|z) is the prior distribution of the PD model parameters, generated based on the latent vector z; Through Markov chain Monte Carlo, multiple instances of the parameters K and θ are sampled from the posterior distribution for the prediction of drug concentration and vital signs.

Images