Anesthesia assessment and risk prediction system based on artificial intelligence

Abstract

The invention discloses an artificial intelligence-based anesthesia evaluation and risk prediction system, which comprises a multi-modal data acquisition and integration module, an artificial intelligence risk prediction model, a personalized anesthesia scheme recommendation module and a dynamic feedback and model optimization module, the multi-modal data acquisition and integration module acquires static data, dynamic data and environmental data of a patient and integrates the data into a unified feature vector, the artificial intelligence risk prediction model adopts a mixed architecture of a deep neural network and a long-short-term memory network, and outputs an overall risk score, a specific risk probability and key risk factors in combination with an attention mechanism, so that the risk prediction accuracy is improved. The personalized anesthesia scheme recommendation module generates and dynamically adjusts the anesthetic dosage based on a reinforcement learning algorithm, and the dynamic feedback and model optimization module continuously optimizes the model performance by utilizing postoperative data and doctor feedback through an online learning technology. According to the method, the anesthesia risk can be accurately evaluated, a personalized anesthesia scheme is provided, and the occurrence rate of intraoperative and postoperative complications is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of anesthesia risk assessment, and in particular to an anesthesia assessment and risk prediction system based on artificial intelligence. Background Art

[0002] Anesthesia is an indispensable and important part of modern medical surgery, and its safety directly affects the patient's life safety and surgical results. With the continuous development of medical technology, the role of anesthesia management in surgery is becoming increasingly important, especially in complex scenarios such as cardiac surgery, laparoscopic surgery, and emergency surgery. However, traditional anesthesia assessment and risk prediction methods have many shortcomings and are difficult to meet the modern medical needs for accuracy and personalization. This provides the necessary technical background for the present invention to provide an anesthesia assessment and risk prediction system based on artificial intelligence.

[0003] In traditional anesthesia management, the assessment of anesthetic risk relies primarily on the physician's clinical experience and subjective judgment. Physicians typically conduct preoperative assessments based on the patient's medical history (e.g., previous medical conditions, surgical history, allergies), physical examination (e.g., height, weight, cardiopulmonary function), and laboratory tests (e.g., blood count, liver and kidney function). However, this approach fails to fully account for individual patient differences, such as the influence of genes on anesthetic drug metabolism, the potential impact of psychological states (e.g., anxiety or depression) on anesthetic efficacy, and the interference of lifestyle habits (e.g., smoking and alcohol consumption) on postoperative recovery. Furthermore, traditional methods have significant shortcomings in predicting intraoperative risk. Although intraoperative monitoring data (e.g., heart rate, blood pressure, and blood oxygen saturation) can be collected in real time by monitoring equipment, analysis of these data often remains at the simple threshold alarm stage and lacks the ability to dynamically predict potential risks (e.g., intraoperative awakening, hypotension, and postoperative respiratory depression). For example, existing technologies struggle to predict potential allergic reactions to specific anesthetics and provide early warning of potential blood pressure fluctuations during surgery.

[0004] When it comes to formulating anesthesia plans, traditional methods are usually based on universal standards, such as determining the initial dose of anesthetic drugs based on the patient's weight and age. However, this "one-size-fits-all" approach fails to fully consider the individual needs of patients, resulting in some patients potentially facing higher anesthesia risks. For example, patients who metabolize certain anesthetic drugs more slowly may experience postoperative respiratory depression due to drug overdose, while patients who metabolize certain drugs more quickly may experience intraoperative awakening due to insufficient dosage. In addition, traditional anesthesia plans lack a dynamic adjustment mechanism during surgery, and doctors can only manually adjust drug dosage and administration timing based on real-time monitoring data and their own experience. This method is inefficient and easily interfered with by human factors.

[0005] Data underutilization is another major issue in traditional anesthesia management. In modern medical scenarios, a large amount of anesthesia-related data has been accumulated, including preoperative assessment data, intraoperative monitoring data, and postoperative recovery data. However, these data are often scattered in different systems (such as electronic medical record systems, monitoring device databases) and are not systematically integrated and analyzed, making it difficult to form effective risk prediction models. For example, a patient's historical anesthesia records may contain information about their response to certain drugs, but existing technologies cannot utilize these data to predict anesthesia risks for current surgeries. In addition, postoperative recovery data (such as recovery time, complication occurrence) and physician feedback data (such as evaluation of anesthesia plans) are also underutilized, resulting in anesthesia management that cannot be continuously improved.

[0006] In recent years, artificial intelligence technology has gradually penetrated into the medical field, especially in disease diagnosis, risk prediction, and personalized treatment plan development. For example, deep learning technology can integrate multi-modal data to extract complex patient features for accurate disease risk prediction; reinforcement learning technology can optimize treatment plans for dynamic adjustment. However, current artificial intelligence systems for anesthesia evaluation and risk prediction still have technical gaps. Existing technologies have not yet had a system that can comprehensively integrate patients' static data (such as medical history, genetic information), dynamic data (such as intraoperative heart rate, blood pressure), and environmental data (such as operating room temperature), and lack risk prediction models based on deep learning and attention mechanisms. In addition, existing technologies fail to introduce reinforcement learning technology into anesthesia plan recommendation, achieving personalized and dynamic anesthesia management. More importantly, existing systems lack dynamic feedback and online learning mechanisms, and cannot utilize postoperative data and physician feedback to continuously optimize model performance, limiting their applicability in different patient populations and medical scenarios.

[0007] To address the above problems, an artificial intelligence-based anesthesia evaluation and risk prediction system is needed to solve these problems. SUMMARY

[0008] The present application aims to solve the technical problems raised in the background art and provides an artificial intelligence-based anesthesia evaluation and risk prediction system.

[0009] The present application achieves the above-mentioned objectives through the following solutions:

[0010] An artificial intelligence-based anesthesia evaluation and risk prediction system includes the following modules:

[0011] A multi-modal data acquisition and integration module is used to acquire static data S, dynamic data D(t) and other data O of a patient, and integrate multi-source heterogeneous data into a unified feature vector X, wherein: the static data S = [S1, S2, S3] includes patient history data S1, genetic data S2 and laboratory examination data S3, which are derived from an electronic medical record system EMR, a genetic test report and a preoperative examination report, and are all standardized dimensionless vectors; the dynamic data D(t) = [HR(t), BP(t), SpO2(t)] includes heart rate HR(t), blood pressure BP(t) and blood oxygen saturation SpO2(t); and the other data O includes psychological state scores and environmental factors;

[0012] An artificial intelligence risk prediction model is used to predict anesthesia risks based on deep learning technology, and outputs an overall risk score R and a specific risk category probability P k and key risk factors;

[0013] A personalized anesthesia regimen recommendation module is used to generate and dynamically adjust anesthesia regimens according to risk prediction results;

[0014] A dynamic feedback and model optimization module is used to continuously optimize model performance using postoperative data and doctor feedback; wherein the system realizes precise assessment of anesthesia risks and recommendation of personalized anesthesia regimens through multi-modal data integration, risk prediction and dynamic feedback mechanisms.

[0015] As a preferred technical solution of the present application, the multi-modal data acquisition and integration module comprises:

[0016] A static data acquisition unit is used to acquire patient history data S1, genetic data S2 and laboratory examination data S3, wherein S1, S2 and S3 are standardized dimensionless vectors;

[0017] A dynamic data acquisition unit is used to acquire intraoperative real-time monitoring data, including heart rate HR(t), blood pressure BP(t) and blood oxygen saturation SpO2(t), wherein t represents time data integration unit, which fuses static data S, dynamic data D(t) and other data O into a feature vector X, and the formula is:

[0018] X = S1, S2, S3, O, HR(t1), BP(t1), SpO2(t1), …, HR(t n ), BP(t n ), SpO2((t n )

[0019] Wherein, t1, t2, …, t n are sampling time points, and n is the number of sampling points. The feature vector X is processed into a dimensionless vector through standardization.

[0020] As a preferred technical solution of the present application, the artificial intelligence risk prediction model adopts a hybrid architecture of deep neural network DNN and long short-term memory network LSTM, and introduces an attention mechanism, specifically including: a DNN sub-model for processing static data S=[S1, S2, S3] and outputting static feature representation F s , and its calculation formula is:

[0021] F s =σ(W s ·S+b s )

[0022] Where W s is a weight matrix (dimensionless), b s is a bias vector (dimensionless), σ is a ReLU activation function, and F s is a dimensionless vector.

[0023] An LSTM sub-model for processing dynamic data sequence D(t)=[HR(t), BP(t), SpO2(t)] and outputting dynamic feature representation F d , and its calculation formula is:

[0024]

[0025] Where h t and c t are hidden state and cell state respectively (both dimensionless), t n is the last time step, F d is a dimensionless vector, and input D(t) has been processed by standardization to be dimensionless.

[0026] As a preferred technical solution of the present application, the artificial intelligence risk prediction model introduces an attention mechanism to calculate the weight of each feature in the feature vector F=[F s , F d ], and its calculation formula is:

[0027]

[0028] Where F i is the i-th component of the feature vector F, m is the total number of features, W a is the attention weight matrix, α i is the dimensionless attention weight, and F att is the weighted feature representation.

[0029] As a preferred technical solution of the present application, the artificial intelligence risk prediction model outputs an overall risk score R, and its calculation formula is:

[0030] R=sigmoid(Wr ·F att +b r )×100

[0031] where W r is the weight matrix, b r is the bias vector, R ranges from 0 to 100; at the same time, the model output specific risk category probability P k , whose formula is:

[0032] P k = softmax(W k ·F att +b k ),

[0033] where k represents the kth risk category, K is the total number of risk categories, W k is the weight matrix, b k is the bias vector, and P k is the dimensionless probability.

[0034] As a preferred technical solution of the present application, the individualized anesthesia regimen recommendation module is based on a reinforcement learning algorithm, specifically including: a state space S t defined as the feature vector X of the current patient and the risk score R, i.e. S t =[X, R];

[0035] an action space A t including an anesthetic drug dose D drug (unit: mg), a drug administration time T drug and an auxiliary intervention measure;

[0036] a reward function Reward defined as the weighted sum of intraoperative stability and postoperative recovery effect, whose formula is:

[0037] Reward = w1·Stability + w2·Recovery

[0038] wherein, BP(t) and BP target are in mmHg, n is the number of sampling points, Stability is dimensionless; Recovery is the postoperative recovery score, w1, w2 are weights, w1 + w2 = 1, and Reward is dimensionless.

[0039] As a preferred technical solution of the present application, the reinforcement learning algorithm adopts a deep deterministic policy gradient DDPG method to update the policy network π and the value network Q, whose update formula is:

[0040]

[0041] wherein θ is the policy network parameter, α is the learning rate, γ is the discount factor, Q(S t ,A t ) is the state-action value.

[0042] As a preferred technical solution of the present application, the dynamic feedback and model optimization module adopts online learning technology, specifically including: a data collection unit collects postoperative recovery data R post , including recovery time T rec and complication occurrence C comp , and doctor feedback data F doc ;

[0043] a model updating unit updates the risk prediction model using new data, and the loss function is:

[0044]

[0045] wherein R pred,i and P k,i are the predicted risk score and risk probability respectively, R true,i and P true,k,i are the true values, λ is the regularization coefficient, N is the sample number, and L is dimensionless.

[0046] As a preferred technical solution of the present application, the system includes a doctor interaction interface for displaying risk prediction results and recommended solutions to an anesthesiologist, specifically including: a risk prediction visualization unit displaying the overall risk score R and specific risk probability P k ;

[0047] a anesthesia regimen recommendation unit displaying the recommended anesthesia drug dose D drug , administration timing T drug and auxiliary intervention measures V fluid ;

[0048] a real-time warning unit issuing an alarm when the risk score R exceeds a preset threshold R threshold .

[0049] As a preferred technical solution of the present application, the system supports multi-scene application, including but not limited to cardiac surgery, laparoscopic surgery and emergency surgery, characterized in that: the system adjusts the model parameters W s , W r , W k to adapt to the risk characteristics of different types of surgery; the system supports multi-language interface and data interface, and is compatible with electronic medical record systems EMR of different medical institutions, making the data acquisition and model deployment portable.

[0050] Compared with the prior art, the present application has the following beneficial effects:

[0051] The present application provides an artificial intelligence-based anesthesia evaluation and risk prediction system, which has significant beneficial effects. First, through the multi-modal data acquisition and integration module, the system can comprehensively integrate the patient's static data, dynamic data and other data, build a comprehensive anesthesia evaluation foundation, overcome the limitations of traditional methods of incomplete evaluation, significantly improve the accuracy and comprehensiveness of anesthesia risk assessment, and provide more reliable decision-making basis for doctors.

[0052] Secondly, the system realizes dynamic prediction of anesthesia risk and recommendation of individualized anesthesia plan through the artificial intelligence risk prediction model and the individualized anesthesia plan recommendation module, combining deep learning and reinforcement learning technology. The model can accurately identify key risk factors and dynamically adjust the anesthesia plan, effectively reduce the incidence of intraoperative and postoperative complications, and at the same time, reduce the work burden of anesthesiologists, improve the efficiency and safety of surgery.

[0053] Finally, the dynamic feedback and model optimization module uses online learning technology to continuously optimize the model performance using postoperative data and doctor feedback, ensuring the applicability of the system in different patient groups and medical scenarios. The design of the doctor interaction interface further improves the practicality of the system, supports multiple languages and data interface compatibility, so that the system can seamlessly integrate into existing medical processes and has a wide application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0054] In order to more clearly illustrate the technical solutions in the present application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0055] Figure 1 is a system block diagram of an artificial intelligence-based anesthesia evaluation and risk prediction system of the present application. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0057] Reference Figure 1The present application provides an artificial intelligence-based anesthesia evaluation and risk prediction system. The technical content of the system is described in detail below in conjunction with the specific embodiments. The system aims to achieve precise evaluation of anesthesia risk and personalized anesthesia regimen recommendation through multi-modal data integration, artificial intelligence risk prediction, personalized anesthesia regimen recommendation, and dynamic feedback mechanism. It is suitable for various scenarios such as cardiac surgery, laparoscopic surgery, and emergency surgery. The implementation process of the system starts with data collection, gradually completes risk prediction, regimen recommendation, and model optimization, and finally provides support for anesthesiologists through a doctor interaction interface.

[0058] Firstly, the system collects various data of the patient through the multi-modal data acquisition and integration module to build a comprehensive anesthesia evaluation foundation. Static data S = [S1, S2, S3] includes patient medical history data S1, genetic data S2, and laboratory examination data S3. These data come from the electronic medical record system (EMR), genetic test reports, and preoperative examination reports. For example, medical history data S1 includes information such as the patient's history of hypertension and diabetes, genetic data S2 includes information about genetic variations related to anesthesia drug metabolism, and laboratory examination data S3 includes blood routine, liver and kidney function indicators. These data are standardized by Z-score to become dimensionless vectors.

[0059] Dynamic data D(t) = [HR(t), BP(t), SpO2(t)] includes intraoperative real-time monitoring of heart rate HR(t) (unit: beats / min), blood pressure BP(t) (unit: mmHg), and oxygen saturation SpO2(t) (unit: %). These data are collected by intraoperative monitoring equipment such as an electrocardiogram monitor, and the time t is in seconds. Other data O includes psychological state score (anxiety level assessed by preoperative questionnaire, range [0, 1], dimensionless) and environmental factors (such as operating room temperature, unit: °C).

[0060] The data integration unit fuses static data S, dynamic data D(t), and other data O into a feature vector X, whose formula is: X = [S1, S2, S3, O, HR(t1), BP(t1), SpO2(t1), …, HR(t n ), BP(t n ), SpO2(t n )]

[0061] Where t1, t2, …, t n are sampling time points, and n is the number of sampling points. Dynamic data is standardized (such as Z-score standardization) to make the feature vector X a dimensionless vector, providing high-quality input for subsequent models.

[0062] Next, the system predicts the anesthesia risk using an artificial intelligence risk prediction model. The model adopts a hybrid architecture of deep neural network (DNN) and long short-term memory network (LSTM), and introduces an attention mechanism.

[0063] The DNN sub-model processes static data S = [S1, S2, S3] and outputs static feature representation F s , whose calculation formula is:

[0064] F s = σ(W s · S + b s )

[0065] where W s is the weight matrix, b s is the bias vector, σ is the ReLU activation function, and F s is a dimensionless vector.

[0066] The LSTM sub-model processes dynamic data sequence D(t) = [HR(t), BP(t), SpO2(t)] and outputs dynamic feature representation F d , whose calculation formula is:

[0067]

[0068] where h t and c t are the hidden state and cell state, respectively, t n is the last time step, F d is a dimensionless vector, and the input D(t) has been processed by standardization to be dimensionless.

[0069] The feature vector F = [F s , F d ] is weighted by the attention mechanism to calculate the weight of each feature:

[0070]

[0071] where F i is the i-th component of the feature vector F, m is the total number of features, W a is the attention weight matrix, α i is the dimensionless attention weight, and F att is the weighted feature representation (dimensionless). Based on F att , the model outputs the overall risk score R, whose formula is:

[0072] R = sigmoid(W r · F att + b r ) × 100

[0073] where W r is the weight matrix, b r is the bias vector, R ranges from 0 to 100.

[0074] Meanwhile, the model outputs the specific risk category probability P k (such as intraoperative awareness risk, postoperative respiratory depression risk), whose formula is:

[0075] P k = softmax(W k · F att + b k ),

[0076] where k denotes the kth risk category, K is the total number of risk categories, W k is the weight matrix, b k is the bias vector, and P k is the dimensionless probability.

[0077] The model also identifies key risk factors such as genetic variations or intraoperative blood pressure fluctuations through attention weights a i .

[0078] Subsequently, the system generates and dynamically adjusts the anesthesia regimen based on the risk prediction results through the personalized anesthesia regimen recommendation module, which is based on a reinforcement learning algorithm. The state space S t is defined as the current patient's feature vector X and risk score R, i.e., S t = [X, R], both of which are dimensionless.

[0079] The action space A t includes the anesthetic drug dose D drug (mg), the administration time T drug (seconds), and auxiliary interventions (such as fluid support volume V fluid (mL)).

[0080] The reward function Reward is defined as the weighted sum of intraoperative stability and postoperative recovery effect, whose formula is:

[0081] Reward = w1 · Stability + w2 · Recovery

[0082] where, BP(t) and BP targetin mmHg, n is the number of sampling points, Stability is dimensionless; Recovery is the postoperative recovery score (range [0, 1], dimensionless), w1, w2 are weights (dimensionless, w1 + w2 = 1), Reward is dimensionless. The reinforcement learning algorithm uses the Deep Deterministic Policy Gradient (DDPG) method to update the policy network p and the value network Q, whose update formula is:

[0083]

[0084] where p is the policy network parameter, a is the learning rate, g is the discount factor (range [0, 1]), Q(S t ,A t ) is the state-action value, and all are dimensionless.

[0085] Through reinforcement learning, the system generates the optimal anesthesia scheme, for example, recommending a propofol dose of 2 mg / kg, an etomidate dose of 0.2 mg / kg, and administering at the 300th second of the operation, while suggesting a fluid support amount of 500 mL.

[0086] The system also continuously optimizes model performance using postoperative data and physician feedback through the dynamic feedback and model optimization module. The data collection unit collects postoperative recovery data R post , including recovery time T rec (unit: hours) and complication occurrence C comp (0 or 1, dimensionless), as well as physician feedback data F doc (score, range [0, 1], dimensionless).

[0087] The model update unit updates the risk prediction model using new data, and its loss function is:

[0088]

[0089] where R pred,i and P k,i are the predicted risk score and risk probability, respectively, R true,i and P true,k,i are the true values, l is the regularization coefficient, N is the number of samples, and L is dimensionless.

[0090] Through online learning techniques, the system continuously optimizes model parameters W s , W r , and W k to adapt to the risk characteristics of different types of surgery, such as focusing more on hypotension risk in cardiac surgery and more on intraoperative wakefulness risk in laparoscopic surgery. The system is equipped with a physician interaction interface to display risk prediction results and recommended schemes to anesthesiologists. The risk prediction visualization unit displays the overall risk score R and the specific risk probability P.k and key risk factors, e.g. "15% risk of intraoperative hypotension, key factor: sensitivity to propofol".

[0091] The anesthesia regimen recommendation unit displays the recommended anesthetic drug dosage D drug , timing of administration t drug and auxiliary interventions V fluid . The real-time warning unit issues an alert when the risk score R exceeds a preset threshold R threshold , e.g. 80, dimensionless, e.g. "hypotension likely within 5 minutes, adjust anesthesia depth". The interface supports multilingual display and is compatible with different medical institutions' electronic medical record systems (EMR) through a data interface, ensuring the portability of data collection and model deployment.

[0092] In practical applications, the system can be applied to various surgical scenarios. Taking a 65-year-old male patient (a patient with a history of hypertension and diabetes, impaired liver function, undergoing non-cardiac surgery) as an example, the system analyzes the preoperative data, and the patient's ASA is 2, predicting a 15% risk of intraoperative hypotension and a 20% risk of postoperative respiratory depression, and identifying the key risk factor as "sensitivity to propofol, impaired liver metabolism". The system recommends reducing the propofol dosage to 1.5 mg / kg, increasing the etomidate proportion to 0.3 mg / kg, recommending the use of cisatracurium which is not metabolized by the liver and kidneys, and strengthening fluid support during the operation to maintain urine output (1 mL / kg / h). Intraoperative monitoring shows that the patient's blood pressure is stable, and postoperative recovery is good without delayed awakening or respiratory depression. The system further optimizes the model through postoperative data feedback to ensure its applicability in different patient populations and medical scenarios.

[0093] In summary, the system realizes precise assessment of anesthesia risk and recommendation of individualized anesthesia regimen through multi-modal data collection, artificial intelligence risk prediction, individualized anesthesia regimen recommendation, and dynamic feedback mechanism, significantly improving anesthesia safety and surgical outcomes, and has broad application prospects.

[0094] Finally, it should be noted that: the above examples are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent substitutions for some technical features; and these modifications or substitutions do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. An artificial intelligence-based anesthesia assessment and risk prediction system, characterized in that: Includes the following modules: The multimodal data acquisition and integration module is used to collect the patient's static data S, dynamic data D(t) and other data O, and integrate the multi-source heterogeneous data into a unified feature vector X, where: static data S = [S1, S2, S3], including the patient's medical history data S1, genetic data S2, and laboratory test data S3, which are derived from the electronic medical record system EMR, genetic test reports and preoperative examination reports, and are all standardized dimensionless vectors; dynamic data D(t) = HR(t), BP(t), SpO2(t)], including heart rate HR(t), blood pressure BP(t), and blood oxygen saturation SpO2(t); other data O, including psychological status scores and environmental factors; Artificial intelligence risk prediction model, used to predict anesthesia risk based on deep learning technology, outputting the overall risk score R and the probability of specific risk category P k ; Personalized anesthesia plan recommendation module, used to generate and dynamically adjust anesthesia plans based on risk prediction results; The dynamic feedback and model optimization module is used to continuously optimize model performance using postoperative data and physician feedback. The system achieves accurate assessment of anesthesia risk and personalized anesthesia plan recommendation through multimodal data integration, risk prediction and dynamic feedback mechanism.

2. The artificial intelligence-based anesthesia assessment and risk prediction system according to claim 1, characterized in that: The multimodal data acquisition and integration module includes: Static data collection unit, used to collect the patient's medical history data S1, genetic data S2, and laboratory test data S3, where S1, S2, and S3 are standardized dimensionless vectors; The dynamic data acquisition unit is used to collect real-time monitoring data during surgery, including heart rate HR(t), blood pressure BP(t), and blood oxygen saturation SpO2(t). t represents the time data integration unit, which fuses static data S, dynamic data D(t), and other data O into a feature vector X. The formula is: X=S1,S2,S3,O,HR(t1),BP(t1),SpO2(t1),…,HR(t n ),BP(t n ),SpO2((t n ) Among them, t1, t2, …, t n is the sampling time point, n is the number of sampling points, and the feature vector X is normalized to a dimensionless vector.

3. The artificial intelligence-based anesthesia assessment and risk prediction system according to claim 1, characterized in that: The artificial intelligence risk prediction model adopts a hybrid architecture of deep neural network DNN and long short-term memory network LSTM, and introduces an attention mechanism, specifically including: DNN sub-model for processing static data S = [S1, S2, S3], outputting static feature representation F s , and its calculation formula is: F s =σ(W s ·S+b s ) Among them, W s is the weight matrix, b s is the bias vector, σ is the ReLU activation function, F s is a dimensionless vector; LSTM sub-model, used to process dynamic data sequence D(t) = HR(t), BP(t), SpO2(t)], and output dynamic feature representation F d , and its calculation formula is: Among them, h t and c t are hidden state and cell state respectively, t n is the last time step, F d is a dimensionless vector. The input D(t) has been normalized to be dimensionless.

4. The artificial intelligence-based anesthesia assessment and risk prediction system according to claim 3, characterized in that: The artificial intelligence risk prediction model introduces an attention mechanism to calculate the feature vector F = [F s ,F d The weight of each feature in ] is calculated as follows: Among them, F i is the i-th component of the feature vector F, m is the total number of features, W a is the attention weight matrix, α i is the dimensionless attention weight, F att is the weighted feature representation.

5. The artificial intelligence-based anesthesia assessment and risk prediction system according to claim 4, characterized in that: The artificial intelligence risk prediction model outputs an overall risk score R, which is calculated as follows: R=sigmoid(W r ·F att +b r )×100 Among them, W r is the weight matrix, b r is the bias vector, The range of R is 0-100; at the same time, the model outputs the probability of a specific risk category P k , the formula is: P k =softmax(W k ·F att +b k ), Where k represents the kth risk category, K is the total number of risk categories, and W k is the weight matrix, b k is the bias vector, P k is a dimensionless probability.

6. The artificial intelligence-based anesthesia assessment and risk prediction system according to claim 1, characterized in that: The personalized anesthesia plan recommendation module is based on the reinforcement learning algorithm and specifically includes: state space S t , defined as the current patient's feature vector X and risk score R, that is, S t =[X,R]; Action space A t , including anesthetic drug doses drug , administration time T drug and adjunctive interventions; The reward function Reward is defined as the weighted sum of intraoperative stability and postoperative recovery effect, and its formula is: Reward=w1·Stability+w2·Recovery in, BP(t) and BP target The unit is mmHg, n is the number of sampling points, Stability is dimensionless; Recovery is the postoperative recovery score, w1 and w2 are weights, w1+w2=1), and Reward is dimensionless.

7. The artificial intelligence-based anesthesia assessment and risk prediction system according to claim 6, characterized in that: The reinforcement learning algorithm uses the deep deterministic policy gradient DDPG method to update the policy network π and the value network Q. The update formula is: Among them, θ is the policy network parameter, α is the learning rate, γ is the discount factor, Q(S t ,A t ) is the state-action value.

8. The artificial intelligence-based anesthesia assessment and risk prediction system according to claim 1, characterized in that: The dynamic feedback and model optimization module adopts online learning technology, specifically including: a data collection unit, which collects postoperative recovery data R post , including the recovery time T rec and complications C comp , and doctor feedback data F doc ; The model updating unit uses new data to update the risk prediction model, and its loss function is: Among them, R pred,i and P k,i are the predicted risk score and risk probability, R true,i and P true,k,i is the true value, λ is the regularization coefficient, N is the number of samples, and L is dimensionless.

9. The artificial intelligence-based anesthesia assessment and risk prediction system according to claim 1, characterized in that: The system includes a doctor interaction interface for displaying risk prediction results and recommended solutions to anesthesiologists, specifically including: a risk prediction visualization unit that displays the overall risk score R, the specific risk probability P k ; Anesthesia plan recommendation unit, showing the recommended anesthetic drug dosage D drug , administration time T drug and auxiliary interventions V fluid ; Real-time warning unit, when the risk score R exceeds the preset threshold R threshold When the alarm is sounded.

10. The artificial intelligence-based anesthesia assessment and risk prediction system according to claim 1, characterized in that: The system supports multiple scenarios, including but not limited to cardiac surgery, laparoscopic surgery and emergency surgery, characterized in that: the system adjusts the model parameter W s ,W r ,W k Adapt to the risk characteristics of different surgical types; the system supports multi-language interfaces and data interfaces, and is compatible with the electronic medical record systems (EMRs) of different medical institutions, making data collection and model deployment portable.

Images