An intelligent anesthesia closed-loop control system based on infusion monitoring

Abstract

The application provides an intelligent anesthesia closed-loop control system based on infusion monitoring, comprising a multi-device data fusion acquisition module configured to synchronously acquire key monitoring data of a monitor, an anesthesia machine and an infusion pump through a WIFI communication protocol; an AI algorithm integrated application module integrating a generative AI planning algorithm, a deep learning prediction algorithm and a compound closed-loop control algorithm to form a decision support chain covering the whole cycle; and a whole-cycle closed-loop mechanism module for constructing a whole-cycle closed-loop management system.The application takes infusion monitoring as the core, synchronously acquires key monitoring data of core devices through a unified data acquisition module, combines intelligent AI algorithms and constructs an anesthesia control whole-cycle closed loop of "preoperative prediction-intraoperative regulation-postoperative iteration".

Description

Technical Field

[0001] This invention relates to the field of medical anesthesia equipment control technology, specifically to an intelligent anesthesia closed-loop control system based on infusion monitoring. It is mainly applied in clinical surgical anesthesia scenarios, realizing precise anesthesia management throughout the entire process from preoperative plan formulation to postoperative data traceability, and is suitable for medical scenarios such as operating rooms and intensive care units. Background Technology

[0002] In clinical anesthesia, precise control of anesthetic drugs is directly related to patient safety and postoperative recovery. Traditional anesthesia control mainly relies on the anesthesiologist's clinical experience, manually adjusting the infusion rate and dosage by observing the patient's vital signs (such as blood pressure and heart rate) and anesthesia equipment parameters. This method has the following significant drawbacks: Preoperative planning lacks personalized predictive support: Existing multimodal anesthesia control systems only focus on intraoperative signal acquisition and control, without combining preoperative patient medical history, physical examination data, laboratory results and anesthetic drug response prediction. They still rely on physician experience to formulate initial plans, which is difficult to adapt to individual differences and is prone to initial dose deviation.

[0003] Isolated data from multiple devices leads to incomplete monitoring: Clinical anesthesia involves multiple devices such as monitors, anesthesia machines, and infusion pumps. Currently, data from each device is mostly collected and displayed independently, lacking a unified platform for integrated analysis. Physicians need to monitor parameters from multiple devices simultaneously, making it difficult to quickly capture correlations between data (such as the correlation between changes in airway pressure in the anesthesia machine and the dosage administered via the infusion pump), resulting in delayed adjustments. Manual adjustment is slow to respond and lacks precision: During the operation, the patient's vital signs will fluctuate dynamically with the operation and the depth of anesthesia. The response speed of manually adjusting the infusion pump parameters is difficult to match the real-time changes in demand, and the control precision is affected by factors such as the doctor's skill level and fatigue, which can easily lead to medication deviation. Lack of predictability and closed-loop feedback mechanism: Existing technologies can only be passively adjusted based on real-time monitoring data, and cannot predict the patient's response trend to anesthetic drugs in advance. They also have not formed a closed-loop control system of "setting goals - data collection - analysis and calculation - dynamic adjustment - feedback correction", making it difficult to achieve stable maintenance of anesthesia depth. Therefore, there is an urgent need for an anesthesia control technology that can integrate data from multiple devices and use intelligent algorithms to realize personalized treatment plans and dynamic closed-loop control, in order to solve existing problems such as personalized treatment plan formulation, multi-device parameter linkage, dedicated adaptation of infusion devices, and postoperative algorithm iteration. Summary of the Invention

[0004] To address the various shortcomings of existing technologies, this invention provides an intelligent anesthesia closed-loop control system based on infusion monitoring. With infusion monitoring as the core, it uses a unified data acquisition module to simultaneously collect key monitoring data from three core devices: monitors, anesthesia machines, and infusion pumps. Combined with intelligent AI algorithms, it constructs a closed-loop system for the entire anesthesia control cycle, encompassing "preoperative prediction, intraoperative adjustment, and postoperative iteration."

[0005] The present invention achieves the above objectives through the following technical solutions: An intelligent anesthesia closed-loop control system based on infusion monitoring includes: The multi-device data fusion acquisition module is configured to synchronously acquire key monitoring data from monitors, anesthesia machines, and infusion pumps via WIFI communication protocol. The monitor data includes blood pressure, heart rate, body temperature, electrocardiogram, blood volume, and BIS index; the anesthesia machine data includes inhaled oxygen concentration, exhaled tidal volume, airway pressure, and respiratory rate; and the infusion pump data includes operating status, infusion rate, infusion volume, infusion time, and infusion mode. The AI ​​algorithm integration module integrates generative AI planning algorithms, deep learning prediction algorithms, and composite closed-loop control algorithms to form a decision support chain covering the entire lifecycle. Among them, the generative AI planning algorithm predicts drug sensitivity and formulates personalized initial dosage plans based on first-dimensional data; the deep learning prediction algorithm uses second-dimensional data to predict the trend of changes in anesthesia depth and potential risks; and the composite closed-loop control algorithm combines PID adjustment and model predictive control framework technology to achieve dynamic adjustment of infusion rate. The full-cycle closed-loop mechanism module is used to build a full-cycle closed-loop management system. In the first stage, a personalized anesthesia plan is generated; in the second stage, dynamic infusion control is achieved; and in the third stage, intraoperative data and recovery indicators are automatically linked and uploaded to the cloud database.

[0006] According to the present invention, an intelligent anesthesia closed-loop control system based on infusion monitoring achieves synchronous acquisition of key monitoring data from the monitor, anesthesia machine, and infusion pump through the following steps: It adopts a standardized WIFI communication protocol and uses built-in or external WIFI modules to achieve physical layer connection between the monitor, anesthesia machine, infusion pump and central control unit; The device identifier and data interface specifications are preset in the central control unit to ensure that the three types of devices complete the initial handshake and authentication through a unified communication protocol. Define the key monitoring data fields and transmission formats for three types of equipment, including: The monitor data includes blood pressure, heart rate, body temperature, electrocardiogram, blood volume, and BIS index, encapsulated in structured data packets; Anesthesia machine data, including inhaled oxygen concentration, exhaled tidal volume, airway pressure, and respiratory rate, are transmitted via a real-time streaming protocol. The infusion pump data includes operating status, infusion rate, infusion volume, infusion time, and infusion mode, which are transmitted via timed pulse signals. A data verification mechanism is introduced, in which the central control unit performs CRC verification or hash value comparison on the received data packets, discards erroneous packets and triggers a retransmission request; The NTP network time protocol or the PTP precise time protocol is used to achieve time synchronization between all devices and the central control unit; Set tiered access permissions to allow only authorized devices to read or write critical data, and prohibit non-medical devices from accessing the system.

[0007] According to the present invention, an intelligent anesthesia closed-loop control system based on infusion monitoring, comprising a generative AI planning algorithm, includes the following steps: Eight-dimensional feature data of patients were collected before surgery, and a feature vector X = [x1, x2, x3, x4, x5, x6, x7, x8] was constructed, where: x1 is age, x2 is weight, x3 is patient BMI value, x4 is medical history coefficient; x5 is liver and kidney function score, x6 is preoperative hemoglobin, x7 is preoperative albumin, and x8 is preoperative heart rate. The Gaussian process regression algorithm with kernel function optimization is used as the training set model for historical cases in the cloud, and is represented as follows:

[0008] in, Indicates the target value d The predicted value, This represents the input data points or dataset that need to be predicted. It is the training dataset. The calculation is the covariance between the current input data and the training data. Similarly, the kernel function is calculated on the training dataset. Xtrain Its own covariance matrix, Indicates the noise variance. It is an identity matrix, and its dimensions are the same as those of the identity matrix. The matrices have the same dimensions. It is about finding the inverse matrix. These are the output values ​​corresponding to the training dataset; According to the intelligent anesthesia closed-loop control system based on infusion monitoring provided by the present invention, the model of the historical case training set in the cloud is called, and the mixture kernel function is substituted into the calculation, which is expressed as the following formula:

[0009] in, It is the result of the mixed kernel function calculation, used to measure data points. a and data pointsb The similarity between them This represents two data points in the input space. For length scale parameters, For shape parameters, and These two coefficients are used to weight two different parts of the mixed kernel function; Input the patient's feature vector into the cloud-based historical case training model, and output the drug sensitivity prediction value and initial dose suggestion; Dosage adjustment based on BMI deviation: final = ·(1 + 0.015·(x3 24), of which The model outputs the initial dose, and x3 is the patient's BMI value.

[0010] According to the present invention, an intelligent anesthesia closed-loop control system based on infusion monitoring is provided. The deep learning prediction algorithm adopts a hybrid model of Bi-LSTM model and attention mechanism. The construction of this model includes the following steps: Input layer construction: Five key physiological parameters reflecting the patient's anesthesia status—BIS index, airway pressure, blood volume, blood oxygen saturation, and exhaled tidal volume—were collected in real time. A 10-second sliding window of data was constructed according to the time series, with a sampling frequency of 10Hz, meaning each sliding window contained 100 sampling points. This data served as the input data source for the model. Feature extraction layer construction: Channel 1 is used for BIS index processing: For BIS index data, a two-layer convolutional neural network is used to extract spatiotemporal features. The convolution kernel size is set to 3×3 and the stride is 1. Through convolution operation, the feature information of BIS index in time and space is fully explored, and finally a 32-dimensional feature vector is output to represent the potential feature pattern of BIS index. Channel 2 is used for multi-parameter time-series processing: For the four parameters of airway pressure, blood volume, blood oxygen saturation, and exhaled tidal volume, a bidirectional long short-term memory network is used for processing; the number of hidden layer nodes is set to 64 and the dropout rate is 0.2 to prevent the model from overfitting and enhance the model's generalization ability; Attention mechanism feature fusion layer: A self-attention mechanism is introduced to calculate the feature weights of each parameter, where the weights are 0.4 for BIS index, 0.2 for airway pressure, 0.15 for blood volume, 0.15 for blood oxygen saturation, and 0.1 for exhaled tidal volume. Based on the weights, the BIS index feature vector output by CNN and the multi-parameter temporal feature vector output by Bi-LSTM are weighted and fused to generate a 64-dimensional comprehensive feature vector. Multi-task output layer: Parallel output of multiple tasks is achieved by using fully connected layers and the Sigmoid activation function: Anesthesia Depth Prediction: Outputs a predicted anesthesia depth value D, with a range of D∈[0,1], where 0 represents very light anesthesia and 1 represents very deep anesthesia. A mapping relationship is established between this value and the BIS index, with the mapping formula: BIS_pred = 100 90·D; Risk probability prediction: Simultaneously output the probability of occurrence of three types of risks: hypoxemia, delayed muscle relaxation, and hypotension, with probability values ​​P ∈ [0, 1].

[0011] According to the present invention, an intelligent anesthesia closed-loop control system based on infusion monitoring executes a composite closed-loop control algorithm, comprising the following steps: The PID adjustment is calculated using the following formula:

[0012] in, Indicates at time t The adjustment amount of the PID controller at that time. This is the proportionality coefficient. For at any time t The weighted bias, The integral coefficient is... From the initial time 0 to time 1 t The deviation integral, These are the differential coefficients. For deviation e ( t ) on time t The derivative of represents the rate of change of the deviation.

[0013] Real-time acquisition of target parameters and monitoring values, and calculation of weighted deviation e(t):

[0014] in, , , , These are the weighting coefficients for BIS index, heart rate, blood pressure, and airway pressure, respectively. The deviation between the target value and the monitored value of the BIS index. The deviation between the target heart rate value and the monitored value. The deviation between the target blood pressure value and the monitored value. This represents the deviation between the target value and the monitored value of airway pressure.

[0015] According to the present invention, an intelligent anesthesia closed-loop control system based on infusion monitoring calculates the optimal adjustment amount within a model predictive control framework. : min Here, argmin represents the variable value that minimizes the objective function. For time step index, Indicates the first t + k The prediction bias at time t, where λ is the smoothing coefficient. Indicates the first t + k The control quantity and adjustment quantity at any given time. Indicates the first t + k The control quantity adjustment at time 1, where min represents the minimized objective; By combining PID adjustment parameters and MPC optimization parameters, the final infusion rate adjustment command is generated:

[0016] Where α is the fusion coefficient.

[0017] According to the present invention, an intelligent anesthesia closed-loop control system based on infusion monitoring is provided. In the first stage, a cloud-based historical case training set model is invoked. This model is constructed based on a hybrid kernel function optimization algorithm. By substituting the patient's preoperative multidimensional feature data, the similarity between the current patient and historical cases is calculated, and the most similar case group is selected. Based on the optimal dosage of similar case groups, and adjusted according to the individual characteristics of the current patients, a generative AI planning algorithm is used to predict the patients' drug sensitivity to propofol and remifentanil, and a personalized initial dosage plan is automatically generated.

[0018] According to the present invention, an intelligent anesthesia closed-loop control system based on infusion monitoring is provided, in the second stage, setting the control target values ​​of anesthesia depth and vital signs; The system collects multi-dimensional data from monitors, anesthesia machines, and infusion pumps in real time via WIFI communication protocol. A hybrid model combining Bi-LSTM and attention mechanism was used to analyze the collected data and predict the trend of changes in anesthesia depth and the probability of risks such as hypoxemia and muscle relaxation recovery. Based on a composite closed-loop control algorithm, combined with multi-parameter weighted deviation, the infusion rates of propofol and remifentanil are dynamically adjusted. It receives patient vital sign feedback data in real time, forming a real-time closed loop of target setting → data collection → AI analysis → infusion regulation → vital sign feedback.

[0019] According to the intelligent anesthesia closed-loop control system based on infusion monitoring provided by the present invention, in the third stage, the system automatically associates intraoperative control records with postoperative recovery indicators to form a complete anesthesia management data package; the data package is uploaded to a cloud database to provide data support for model optimization; Based on postoperative data, the parameters of the generative AI planning algorithm model and the deep learning prediction algorithm model are optimized through transfer learning technology.

[0020] Therefore, compared with the prior art, the intelligent anesthesia closed-loop control system based on infusion monitoring proposed in this invention has the following beneficial effects: 1. Multi-device data fusion and acquisition: This invention takes infusion monitoring as the core module and uses a dedicated WIFI communication protocol to adapt to three types of devices: monitors, anesthesia machines, and infusion pumps, to achieve synchronous acquisition of multi-dimensional data and solve the problem of isolated data from multiple devices.

[0021] 2. Integrated application of artificial intelligence algorithms: This invention integrates three types of algorithms: "generative AI preoperative planning + deep learning intraoperative prediction + compound closed-loop control intraoperative regulation", forming a full-cycle decision chain, which improves the personalization and accuracy of anesthesia control.

[0022] 3. Full-cycle closed-loop mechanism: This invention constructs a full-cycle closed loop of "preoperative plan formulation - intraoperative dynamic control - postoperative data traceability - algorithm model iteration", realizing intelligent management of anesthesia control throughout the entire process.

[0023] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of an embodiment of an intelligent anesthesia closed-loop control system based on infusion monitoring according to the present invention.

[0025] Figure 2 This is a schematic diagram illustrating the operating principle of the system modules and algorithms in an embodiment of an intelligent anesthesia closed-loop control system based on infusion monitoring according to the present invention.

[0026] Figure 3 This is a schematic diagram of the structure of the hybrid model of Bi-LSTM model + attention mechanism in an embodiment of the intelligent anesthesia closed-loop control system based on infusion monitoring of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0028] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0029] An embodiment of an intelligent anesthesia closed-loop control system based on infusion monitoring See Figures 1 to 3 This embodiment provides an intelligent anesthesia closed-loop control system based on infusion monitoring, including: The multi-device data fusion acquisition module is configured to synchronously acquire key monitoring data from monitors, anesthesia machines, and infusion pumps via WIFI communication protocol. The monitor data includes blood pressure, heart rate, body temperature, electrocardiogram, blood volume, and BIS index; the anesthesia machine data includes inhaled oxygen concentration, exhaled tidal volume, airway pressure, and respiratory rate; and the infusion pump data includes operating status, infusion rate, infusion volume, infusion time, and infusion mode. The AI ​​algorithm integration module integrates generative AI planning algorithms, deep learning prediction algorithms, and composite closed-loop control algorithms to form a decision support chain covering the entire lifecycle. Among them, the generative AI planning algorithm predicts drug sensitivity and formulates personalized initial dosage plans based on first-dimensional data; the deep learning prediction algorithm uses second-dimensional data to predict the trend of changes in anesthesia depth and potential risks; and the composite closed-loop control algorithm combines PID adjustment and model predictive control framework technology to achieve dynamic adjustment of infusion rate. The full-cycle closed-loop mechanism module is used to build a full-cycle closed-loop management system. In the first stage, it generates personalized anesthesia plans; in the second stage, it realizes dynamic infusion control; and in the third stage, it automatically links intraoperative data and recovery indicators and uploads them to the cloud database.

[0030] In this embodiment, the key monitoring data from the monitor, anesthesia machine, and infusion pump are collected synchronously through the following steps: The system adopts standardized WIFI communication protocols (such as IEEE 802.11a / b / g / n / ac) and uses built-in or external WIFI modules to achieve physical layer connection between the monitor, anesthesia machine, infusion pump and central control unit. The device identifier (MAC address) and data interface specifications are preset in the central control unit to ensure that the three types of devices complete the initial handshake and authentication through a unified communication protocol; Define the key monitoring data fields and transmission formats for three types of equipment, including: The monitor data includes blood pressure, heart rate, body temperature, electrocardiogram, blood volume, BIS index (bispectral index of brainwave), etc., and is encapsulated in structured data packets; Anesthesia machine data, including inhaled oxygen concentration, exhaled tidal volume, airway pressure, and respiratory rate, is transmitted via real-time streaming protocols such as MQTT. The infusion pump data includes operating status (on / paused / sleep / off), infusion rate, infusion volume, infusion time, infusion mode, etc., and is transmitted by a timed pulse signal, such as 1Hz. A data verification mechanism is introduced, in which the central control unit performs CRC verification or hash value comparison on the received data packets, discards erroneous packets and triggers a retransmission request; To handle network fluctuations or momentary device failures and ensure data continuity, a dynamic buffer queue is set up. The length of the buffer queue is dynamically adjusted according to device priority; for example, data from monitors has the highest priority. The NTP network time protocol or the PTP precise time protocol is used to achieve time synchronization between all devices and the central control unit; Data with acquisition delays exceeding a threshold (e.g., 300ms) are marked, and missing values ​​are compensated for in AI analysis using interpolation algorithms; Enable WPA2 or WPA3 encryption protocols at the WIFI communication layer to prevent data eavesdropping or tampering; Set tiered access permissions to allow only authorized devices (such as anesthesia workstations) to read or write critical data, and prohibit non-medical devices from accessing the system.

[0031] In this embodiment, the generative AI planning algorithm, based on the patient's preoperative multidimensional data (age, weight, medical history, liver and kidney function, laboratory indicators), uses an improved Gaussian process regression algorithm to predict the patient's sensitivity to drugs such as propofol and remifentanil, and formulates a personalized initial dose and drug combination plan, including the following steps: Eight-dimensional feature data of patients were collected before surgery, and a feature vector X = [x1, x2, x3, x4, x5, x6, x7, x8] was constructed, where: x1 is age (years), x2 is weight (kg), x3 is BMI (kg / m²). 2x4 is the medical history coefficient, for example, no underlying disease = 0, hypertension = 0.3, diabetes = 0.4, abnormal liver and kidney function = 0.5, and multiple underlying diseases are summed; x5 is the liver and kidney function score, such as Child-Pugh classification: A = 1, B = 2, C = 3; x6 is the preoperative hemoglobin (g / L); x7 is the preoperative albumin (g / L); x8 is the preoperative heart rate (beats / minute). The Gaussian process regression algorithm with kernel function optimization is used as the training set model for historical cases in the cloud, and is represented as follows:

[0032] in, Indicates the target value d The predicted value, This represents the input data points or dataset that need to be predicted. It is the training dataset. The calculation is the covariance between the current input data and the training data. Similarly, the kernel function is calculated on the training dataset. Xtrain Its own covariance matrix, Indicates the noise variance. It is an identity matrix, and its dimensions are the same as those of the identity matrix. The matrices have the same dimensions. It is about finding the inverse matrix. ≥100,000 feature data examples These are the output values ​​(labels) corresponding to the training dataset, corresponding to the optimal dose vector, with propofol in mg / (kg·h) and remifentanil in μg / (kg·min). The model, trained on a cloud-based historical case dataset, is used for calculation via a hybrid kernel function, as shown in the following formula:

[0033] in, It is the result of the mixed kernel function calculation, used to measure data points. a and data points b The similarity between them This represents two data points in the input space. In applications such as anesthetic drug sensitivity prediction, these can be vectors containing multidimensional preoperative features of the patient (such as age, weight, medical history coefficients, etc.). The data is calculated... a and b The kernel function values ​​between the two patients can be used to assess the degree of similarity between their characteristics. For length scale parameters, This is a shape parameter with a value of 1.2, primarily used to adjust the shape of the kernel function. and These two coefficients are used to weight two different parts of the mixed kernel function; In Gaussian process regression algorithms, kernel functions define the covariance structure between data points, describing the similarity between different data points in the input space. Data points with higher similarity are more correlated in their probability distributions. Through kernel functions, Gaussian processes can map input data to a high-dimensional feature space, thus better capturing complex patterns and nonlinear relationships in the data. Hybrid kernel functions are an important method for kernel function selection and optimization in Gaussian process regression algorithms with kernel function optimization. By reasonably combining and optimizing hybrid kernel functions, the performance and predictive ability of Gaussian process regression models on complex data can be significantly improved.

[0034] Input the patient feature vector into the cloud-based historical case training model, and output the predicted drug sensitivity value (high / medium / low) and initial dose suggestion; Setting upper and lower limit constraints for propofol: min=2.0 mg / (kg·h) max=6.0 mg / (kg·h); remifentanil min=0.05 μg / (kg·min), max = 0.2 μg / (kg·min); Dosage adjustment based on BMI deviation: final = ·(1 + 0.015·(x3 24), of which The model outputs the initial dose, where x3 is the patient's BMI value; Doctors can manually adjust the initial dose within ±20%, and the system automatically records the adjustment log (including the basis for adjustment, timestamp, and doctor's identification). Based on the sensitivity prediction results and the corrected dosage, a personalized drug combination regimen of propofol and remifentanil is generated to optimize the effects of anesthesia induction and maintenance.

[0035] As can be seen, this embodiment solves the bias problem of traditional empirical medication by fusing multidimensional preoperative data through Gaussian process regression algorithm, combined with kernel function optimization and individualized correction, and significantly improves the accuracy and safety of initial anesthesia dosage.

[0036] In this embodiment, the deep learning prediction algorithm employs a hybrid model combining a Bi-LSTM model and an attention mechanism, integrating multi-dimensional data such as the BIS index, airway pressure, and blood volume to predict trends in anesthesia depth and the risk of hypoxemia and muscle relaxation recovery, thereby improving predictive foresight. Figure 3 As shown, the construction of this model includes the following steps: Input layer construction: Five key physiological parameters reflecting the patient's anesthesia status—BIS index, airway pressure, blood volume, blood oxygen saturation, and exhaled tidal volume—were collected in real time. A 10-second sliding window of data was constructed according to the time series, with a sampling frequency of 10Hz, meaning each sliding window contained 100 sampling points. This data served as the input data source for the model. Feature extraction layer construction: Channel 1 is used for BIS index processing: For BIS index data, a 2-layer convolutional neural network (CNN) is used to extract spatiotemporal features. The convolution kernel size is set to 3×3 and the stride is 1. The convolution operation fully explores the feature information of BIS index in time and space, and finally outputs a 32-dimensional feature vector to represent the potential feature pattern of BIS index. Channel 2 is used for multi-parameter time-series processing: For four parameters—airway pressure, blood volume, blood oxygen saturation, and exhaled tidal volume—a bidirectional long short-term memory network (Bi-LSTM) is employed. The number of hidden layer nodes is set to 64, and the dropout rate is 0.2 to prevent overfitting and enhance the model's generalization ability. Bi-LSTM can fully consider the context of the time-series data, effectively capturing the changing patterns of these four parameters over time, and outputting a 64-dimensional time-series feature vector. Attention mechanism feature fusion layer: A self-attention mechanism is introduced to calculate the feature weights of each parameter, where the weights are 0.4 for BIS index, 0.2 for airway pressure, 0.15 for blood volume, 0.15 for blood oxygen saturation, and 0.1 for exhaled tidal volume. Based on the weights, the BIS index feature vector output by CNN and the multi-parameter temporal feature vector output by Bi-LSTM are weighted and fused to generate a 64-dimensional comprehensive feature vector. Multi-task output layer: Parallel output of multiple tasks is achieved by using fully connected layers and the Sigmoid activation function: Anesthesia Depth Prediction: Outputs a predicted anesthesia depth value D, with a range of D∈[0,1], where 0 represents very light anesthesia and 1 represents very deep anesthesia. A mapping relationship is established between this value and the BIS index, with the mapping formula: BIS_pred = 100 90·D, thus intuitively mapping the model's output of the predicted depth of anesthesia into the clinically commonly used BIS index, making it easier for medical staff to understand and apply; Risk probability prediction: Simultaneously outputs the probability of occurrence of three types of risks: hypoxemia, delayed muscle relaxation, and hypotension, with probability values ​​P ∈ [0, 1]. This provides a quantitative basis for medical staff to take preventive measures in advance, effectively ensuring the safety of patients during anesthesia.

[0037] Early warning levels are classified based on risk probability thresholds: Level 1 warning (0.3≤P<0.6): Triggers a yellow alert, prompting doctors to pay attention; Level II warning (0.6≤P<0.9): Triggers an orange alert; adjustment of anesthesia parameters is recommended. Level 3 warning (P≥0.9): Triggers a red alert and automatically initiates emergency control procedures; It simultaneously outputs the trend curve of anesthesia depth change in the next 5 seconds to help doctors predict the anesthesia status.

[0038] As can be seen, this embodiment, through a hybrid architecture of Bi-LSTM and attention mechanisms, effectively integrates multi-dimensional time-series data, solving the problem of traditional models lagging in predicting dynamic changes in anesthesia depth and complex risks, and significantly improving the foresight and accuracy of intraoperative safety monitoring.

[0039] In this embodiment, a composite closed-loop control algorithm is executed, combining PID and model predictive control (MPC) fusion algorithms. Using the target parameter deviation as input, a rapid response is first achieved through PID, followed by optimization through MPC. Anesthesia machine parameter weights are introduced to dynamically adjust the infusion rate, resulting in more accurate anesthesia control. Specifically, the following steps are included: The PID adjustment is calculated using the following formula:

[0040] in, Indicates at time t The adjustment amount of the PID controller is used in the anesthesia closed-loop control system to adjust the infusion rate of anesthetic drugs (such as propofol and remifentanil) so that the actual monitored physiological parameters are close to the target parameters. The PID parameters are shown in Table 1. This is the proportionality coefficient. For at any time t The weighted bias, The integral coefficient is... From the initial time 0 to time 1 t The deviation integral, These are the differential coefficients. For deviation e ( t ) on time t The derivative of represents the rate of change of the deviation.

[0041] Table 1: PID Parameters

[0042] Real-time acquisition of target parameters and monitoring values, and calculation of weighted deviation e(t):

[0043] in, , , , These are the weighting coefficients for BIS index, heart rate, blood pressure, and airway pressure, respectively. =0.4、 =0.15、 =0.25、 =0.2; The deviation between the target value and the monitored value of the BIS index. The deviation between the target heart rate value and the monitored value. The deviation between the target blood pressure value and the monitored value. This represents the deviation between the target value and the monitored value of airway pressure. Within the model predictive control framework, the optimal adjustment amount is calculated. : min Here, argmin represents the variable value that minimizes the objective function. For time step index, Indicates the first t + k In the context of anesthesia control, the prediction bias at a given moment is the difference between the target physiological parameter value and the predicted physiological parameter monitoring value at that moment, such as the difference between the target value and the predicted value of the BIS index, the difference between the target value and the predicted value of the heart rate, etc. λ is the smoothing coefficient, λ=0.15, and a future state prediction model is constructed with a prediction step size of 5 seconds (k=1~5). Indicates the first t + k The control quantity and adjustment quantity at any given time. Indicates the first t + k In the control of anesthetic drug infusion, the control variables at time 1 are the adjustments made to the drug infusion rate at different times; min represents the target to be minimized. Set constraints: single adjustment amount ≤ 15% of the initial dose to avoid sudden changes in infusion rate; By integrating PID adjustment parameters and MPC optimization parameters, a final infusion rate adjustment command is generated and sent to the infusion pump.

[0044] Where α is the fusion coefficient, 0.3≤α≤0.7, which is dynamically adjusted according to the system response delay.

[0045] As can be seen, this embodiment, through the fusion architecture of PID fast response and MPC optimized prediction, takes into account both the real-time performance and accuracy of anesthesia control. Combined with multi-parameter weighting and infusion constraint mechanisms, it effectively solves the problems of easy overshoot and response lag in traditional single-model control, and significantly improves the safety and stability of anesthesia control.

[0046] In the full-cycle closed-loop mechanism module, a full-cycle closed loop of "preoperative plan formulation - intraoperative dynamic control - postoperative data tracking - algorithm model iteration" is constructed. In this embodiment, the first stage is the preoperative plan formulation stage, the second stage is the intraoperative dynamic control stage, and the third stage is the postoperative data tracking and algorithm model iteration stage.

[0047] In the first stage, the cloud-based historical case training set model is invoked. This model is built based on a hybrid kernel function optimization algorithm. By substituting the patient's preoperative multidimensional feature data, including but not limited to age, weight, medical history, liver and kidney function, and laboratory indicators, the model predicts the patient's drug sensitivity to propofol and remifentanil, calculates the similarity between the current patient and historical cases, and selects the most similar case group. Based on the optimal dosage of similar case groups, and adjusted according to the individual characteristics of the current patients (such as BMI, liver and kidney function scores), a generative AI planning algorithm is used to predict the patient's drug sensitivity (high / medium / low) to propofol and remifentanil, and a personalized initial dosage plan is automatically generated.

[0048] In the second phase, target values ​​for the control of anesthesia depth (BIS index) and vital signs (heart rate, blood pressure, blood oxygen saturation, etc.) are set. The system uses a WIFI communication protocol to collect multi-dimensional data from monitors, anesthesia machines, and infusion pumps in real time, including but not limited to BIS index, airway pressure, blood volume, blood oxygen saturation, and exhaled tidal volume. A hybrid model combining Bi-LSTM and attention mechanism was used to analyze the collected data and predict the trend of changes in anesthesia depth and the probability of risks such as hypoxemia and muscle relaxation recovery. Based on a composite closed-loop control algorithm, combined with multi-parameter weighted deviations (e.g., BIS index weight 0.4, heart rate 0.15, blood pressure 0.2, airway pressure 0.2), the infusion rate of propofol and remifentanil is dynamically adjusted with a control accuracy of ±0.1 mg / (kg). h) (propofol), ±0.01 μg / (kg) min) (remifentanil); It receives patient vital signs feedback data in real time, forming a real-time closed loop of target setting → data collection → AI analysis → infusion control → vital signs feedback, iterating once every 300ms to ensure stable anesthesia status.

[0049] In the third stage, the system automatically links intraoperative control records (infusion rate changes, risk event responses) with postoperative recovery indicators (awakening time, complication status) to form a complete anesthesia management data package; The data package is uploaded to the cloud database to provide data support for subsequent model optimization. Based on postoperative data, the parameters of the generative AI planning algorithm model and the deep learning prediction algorithm model are optimized through transfer learning technology to improve the accuracy of subsequent patient drug sensitivity prediction and the precision of regulation. The third stage includes the following execution steps: Data upload: Intraoperative data (control records, risk events, equipment parameters) + Postoperative recovery data (awakening time, extubation time, VAS pain score at 24 hours postoperatively, complication status); Model optimization: By updating the parameters of the generative AI and deep learning models through transfer learning, the model prediction accuracy improves by ≥1% for every 1000 new cases. Traceability query: Supports data retrieval by patient ID, surgery type, and date, and generates visual reports (data is retained for 3 years).

[0050] As can be seen, this embodiment achieves intelligent control of anesthesia from preoperative planning to postoperative optimization through closed-loop management throughout the entire cycle, which significantly improves anesthesia safety and control efficiency.

[0051] In practical applications, this embodiment also provides an application method for an intelligent anesthesia closed-loop control system based on infusion monitoring, including the following steps: System deployment and preoperative planning: Equipment installation and connection: First, connect the monitor, anesthetic, and infusion pump to the infusion system via WIFI protocol. The system detects the connection status of the equipment. If there is no fault prompt, the data transmission delay test is <=300ms to ensure smooth data communication between the devices.

[0052] Preoperative data entry: Key patient information (age, weight, BMI, medical history, liver and kidney function, preoperative physiological indicators, etc.) is entered through the user interaction module, generating an 8-dimensional feature vector, which serves as input data for generative AI, such as: X = [62, 78, 28.3, 0.8, 2, 105, 32, 76].

[0053] Generative AI-based preoperative planning: a) Call the cloud-based historical case training set model, substitute it into the mixed kernel function to calculate, and predict the patient's drug sensitivity to propofol and remifentanil.

[0054] b) Based on the calculation results in a, find the "most similar case group" between the current patient and historical cases. Then, based on the optimal dose for similar cases and adjusted for the patient's own characteristics (such as BMI, liver and kidney function), automatically generate the initial dose, such as: i. Patient sensitivity: Propofol "moderate sensitivity", remifentanil "moderate sensitivity" ii. Initial dose of anesthetic drugs: propofol prop=3.8 mg / (kg·h), remifentanil remi = 0.13 μg / (kg·min); c) Based on the calculation results in b, adjust the dosage using the BMI correction formula to ensure it is appropriate for individual patient differences. Calculate the corrected dosing rate, as shown in the following results: i. Propofol: inal_prop =3.8×(1+0.015×(28.3 24))=3.8×1.0645≈4.04 mg / (kg·h); ii. Remifentanil: final_remi=0.13×1.0645≈0.138; d) Anesthesiologists can fine-tune the dosage within ±20% based on the type of surgery, and the system records the adjustment log: i. Type of surgery: For example, in orthopedic surgery with greater trauma, the dosage of analgesic drugs (remifentanil) can be appropriately increased; for minimally invasive surgery, the dosage of sedative drugs (propofol) can be appropriately reduced. ii. Special patient conditions: If the patient has preoperative anxiety, the initial dose can be slightly increased to ensure smooth induction; if the patient is elderly (e.g., >75 years old), even if the AI ​​determines "moderate sensitivity", the dose can be appropriately reduced to reduce the risk of cyclic inhibition. iii. Drug interactions: If the patient is using other drugs that affect the depth of anesthesia (such as preoperative sedatives), the physician may adjust the initial dose to avoid additive drug effects.

[0055] IV. Target Parameter Setting: Set target ranges for heart rate, systolic blood pressure, and BIS index as benchmarks for intraoperative control.

[0056] Intraoperative dynamic control: 1. Phase One: Anesthesia Induction (0-5 minutes after the start of surgery) a) Data acquisition: Collect real-time data from monitors (BIS index, heart rate, blood pressure, etc.), anesthesia machines (airway pressure, tidal volume, etc.), and infusion pumps (infusion rate, etc.) at a data acquisition frequency of 1 time / second. b) Deep learning prediction: i. Take the data collected in a 10-second sliding window and input it into the Bi-LSTM + attention mechanism hybrid model.

[0057] ii. The model outputs the predicted value of anesthesia depth (D) and the risk probability of hypoxemia, delayed muscle relaxation, and hypotension, and determines whether to trigger an early warning (if there is no risk, it enters the control phase).

[0058] c) Composite closed-loop control: i. According to the weighted deviation formula Calculate the deviation between real-time data and target parameters; ii. Substitute into the PID formula Calculate the fast adjustment amount, and then combine it with the MPC formula to optimize the adjustment amount to avoid parameter abrupt changes.

[0059] iii. Output the final infusion rate adjustment command to the infusion pump, collect feedback data after 1 minute, and confirm whether the parameters have returned to the target range.

[0060] 2. Phase Two: Surgical Procedure Period (approximately 30 minutes after the start of surgery) a) Sudden Abnormal Response: If the surgical procedure causes a sudden change in the patient's physiological parameters (heart rate, blood pressure, BIS index) or anesthesia machine parameters (airway pressure): i. The deep learning model re-predicts the risk, and if the airway pressure is abnormal, a Level 1 warning is triggered. ii. Repeat the compound closed-loop control steps, calculate the weighted deviation and adjustment amount, and send auxiliary instructions (such as adjusting the inhaled oxygen concentration) to the anesthesia machine to optimize the patient's condition. b) Collect feedback data after 2 minutes to confirm that the parameters have returned to stability.

[0061] Risk warning and intervention 1. Level II Warning (Risk of Hypoxemia) a) Triggering conditions: abnormal blood oxygen saturation and exhaled tidal volume, with the probability of hypoxemia output by the deep learning model being between 50% and 70%. b) Early warning and intervention: i. The system triggers an orange alert box, a 2Hz buzzer, and flashing indicator lights, displaying the warning type, risk probability, and recommended intervention measures (such as increasing the inhaled oxygen concentration or checking the airway). ii. After confirmation by the anesthesiologist, the system automatically executes the recommended measures to keep the infusion pump parameters stable. Data is collected after 1.5 minutes to confirm that the risk has been eliminated.

[0062] 2. Level 3 warning (infusion pump malfunction + insufficient anesthesia) a) Triggering conditions: Abnormal infusion pump flow rate, and BIS index and heart rate deviating too much from the target range (e.g., heart rate deviating by +50%). b) Early warning and intervention: i. The system triggers a red warning box, a 3Hz high-decibel alarm, and rapidly flashing lights, indicating the type of emergency risk and the necessary intervention measures (such as temporary manual dose adjustment). ii. The anesthesiologist clicks "Confirm Intervention," manually enters the adjustment parameters, the system resumes closed-loop control, and data is collected after 5 minutes to confirm that the parameters are stable.

[0063] Postoperative data tracking and algorithm iteration 1. Data Upload: After the surgery, the system automatically uploads the preoperative plan, intraoperative data (physiological parameters, control records, early warning events), and postoperative recovery indicators (awakening time, extubation time, pain score, and complication status) to the cloud database. 2. Model Iteration: New case data is integrated into the training set via transfer learning in the cloud, optimizing the kernel function parameters of the generative AI and the attention weights of the deep learning model to improve subsequent prediction and control accuracy. Performance Verification: After adding a certain number of similar cases, the changes in the model's prediction error and accuracy are statistically analyzed to verify the iterative effect.

[0064] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0065] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. An intelligent anesthesia closed-loop control system based on infusion monitoring, characterized in that, include: The multi-device data fusion acquisition module is configured to synchronously acquire key monitoring data from monitors, anesthesia machines, and infusion pumps via WIFI communication protocol. The monitor data includes blood pressure, heart rate, body temperature, electrocardiogram, blood volume, and BIS index; the anesthesia machine data includes inhaled oxygen concentration, exhaled tidal volume, airway pressure, and respiratory rate; and the infusion pump data includes operating status, infusion rate, infusion volume, infusion time, and infusion mode. The AI ​​algorithm integration module integrates generative AI planning algorithms, deep learning prediction algorithms, and composite closed-loop control algorithms to form a decision support chain covering the entire lifecycle. Among them, the generative AI planning algorithm predicts drug sensitivity and formulates personalized initial dosage plans based on first-dimensional data; the deep learning prediction algorithm uses second-dimensional data to predict the trend of changes in anesthesia depth and potential risks; and the composite closed-loop control algorithm combines PID adjustment and model predictive control framework technology to achieve dynamic adjustment of infusion rate. The full-cycle closed-loop mechanism module is used to build a full-cycle closed-loop management system. In the first stage, a personalized anesthesia plan is generated; in the second stage, dynamic infusion control is achieved; and in the third stage, intraoperative data and recovery indicators are automatically linked and uploaded to the cloud database.

2. The system according to claim 1, characterized in that, The following steps enable the synchronous acquisition of key monitoring data from monitors, anesthesia machines, and infusion pumps: It adopts a standardized WIFI communication protocol and uses built-in or external WIFI modules to achieve physical layer connection between the monitor, anesthesia machine, infusion pump and central control unit; The device identifier and data interface specifications are preset in the central control unit to ensure that the three types of devices complete the initial handshake and authentication through a unified communication protocol. Define the key monitoring data fields and transmission formats for three types of equipment, including: The monitor data includes blood pressure, heart rate, body temperature, electrocardiogram, blood volume, and BIS index, encapsulated in structured data packets; Anesthesia machine data, including inhaled oxygen concentration, exhaled tidal volume, airway pressure, and respiratory rate, are transmitted via a real-time streaming protocol. The infusion pump data includes operating status, infusion rate, infusion volume, infusion time, and infusion mode, which are transmitted via timed pulse signals. A data verification mechanism is introduced, in which the central control unit performs CRC verification or hash value comparison on the received data packets, discards erroneous packets and triggers a retransmission request; The NTP network time protocol or the PTP precise time protocol is used to achieve time synchronization between all devices and the central control unit; Set tiered access permissions to allow only authorized devices to read or write critical data, and prohibit non-medical devices from accessing the system.

3. The system according to claim 1, characterized in that, Generative AI planning algorithms include the following steps: Eight-dimensional feature data of patients were collected before surgery, and a feature vector X = [x1, x2, x3, x4, x5, x6, x7, x8] was constructed, where: x1 is age, x2 is weight, x3 is patient BMI value, x4 is medical history coefficient; x5 is liver and kidney function score, x6 is preoperative hemoglobin, x7 is preoperative albumin, and x8 is preoperative heart rate. The Gaussian process regression algorithm with kernel function optimization is used as the training set model for historical cases in the cloud, and is represented as follows: in, Indicates the target value d The predicted value, This represents the input data points or dataset that need to be predicted. It is the training dataset. The calculation is the covariance between the current input data and the training data. Similarly, the kernel function is calculated on the training dataset. Xtrain Its own covariance matrix, Indicates the noise variance. It is an identity matrix, and its dimensions are the same as those of the identity matrix. The matrices have the same dimensions. It is about finding the inverse matrix. It is the output value corresponding to the training dataset.

4. The system according to claim 3, characterized in that: The model, trained on a cloud-based historical case dataset, is used for calculation via a hybrid kernel function, as shown in the following formula: in, It is the result of the mixed kernel function calculation, used to measure data points. a and data points b The similarity between them This represents two data points in the input space. For length scale parameters, For shape parameters, and These two coefficients are used to weight two different parts of the mixed kernel function; Input the patient's feature vector into the cloud-based historical case training model, and output the drug sensitivity prediction value and initial dose suggestion; Dosage adjustment based on BMI deviation: final = ·(1 + 0.015·(x3 24), of which The model outputs the initial dose, and x3 is the patient's BMI value.

5. The system according to claim 1, characterized in that, The deep learning prediction algorithm uses a hybrid model of Bi-LSTM and attention mechanism. The construction of this model includes the following steps: Input layer construction: Five key physiological parameters reflecting the patient's anesthesia status—BIS index, airway pressure, blood volume, blood oxygen saturation, and exhaled tidal volume—were collected in real time. A 10-second sliding window of data was constructed according to the time series, with a sampling frequency of 10Hz, meaning each sliding window contained 100 sampling points. This data served as the input data source for the model. Feature extraction layer construction: Channel 1 is used for BIS index processing: For BIS index data, a two-layer convolutional neural network is used to extract spatiotemporal features. The convolution kernel size is set to 3×3 and the stride is 1. Through convolution operation, the feature information of BIS index in time and space is fully explored, and finally a 32-dimensional feature vector is output to represent the potential feature pattern of BIS index. Channel 2 is used for multi-parameter time-series processing: For the four parameters of airway pressure, blood volume, blood oxygen saturation, and exhaled tidal volume, a bidirectional long short-term memory network is used for processing; the number of hidden layer nodes is set to 64 and the dropout rate is 0.2 to prevent the model from overfitting and enhance the model's generalization ability; Attention mechanism feature fusion layer: A self-attention mechanism is introduced to calculate the feature weights of each parameter, where the weights are 0.4 for BIS index, 0.2 for airway pressure, 0.15 for blood volume, 0.15 for blood oxygen saturation, and 0.1 for exhaled tidal volume. Based on the weights, the BIS index feature vector output by CNN and the multi-parameter temporal feature vector output by Bi-LSTM are weighted and fused to generate a 64-dimensional comprehensive feature vector. Multi-task output layer: Parallel output of multiple tasks is achieved by using fully connected layers and the Sigmoid activation function: Anesthesia Depth Prediction: Outputs a predicted anesthesia depth value D, with a range of D∈[0,1], where 0 represents very light anesthesia and 1 represents very deep anesthesia. A mapping relationship is established between this value and the BIS index, with the mapping formula: BIS_pred = 100 90·D; Risk probability prediction: Simultaneously output the probability of occurrence of three types of risks: hypoxemia, delayed muscle relaxation, and hypotension, with probability values ​​P ∈ [0, 1].

6. The system according to claim 1, characterized in that, Executing the composite closed-loop control algorithm includes the following steps: The PID adjustment is calculated using the following formula: in, Indicates at time t The adjustment amount of the PID controller at that time. This is the proportionality coefficient. For at any time t The weighted bias, The integral coefficient is... From the initial time 0 to time 1 t The deviation integral, These are the differential coefficients. For deviation e ( t ) on time t The derivative of represents the rate of change of the deviation; Real-time acquisition of target parameters and monitoring values, and calculation of weighted deviation e(t): in, , , , These are the weighting coefficients for BIS index, heart rate, blood pressure, and airway pressure, respectively. The deviation between the target value and the monitored value of the BIS index. The deviation between the target heart rate value and the monitored value. The deviation between the target blood pressure value and the monitored value. This represents the deviation between the target value and the monitored value of airway pressure.

7. The system according to claim 6, characterized in that: Within the model predictive control framework, the optimal adjustment amount is calculated. : min Here, argmin represents the variable value that minimizes the objective function. For time step index, Indicates the first t + k The prediction bias at time t, where λ is the smoothing coefficient. Indicates the first t + k The control quantity and adjustment quantity at any given time. Indicates the first t + k The control quantity adjustment at time 1, where min represents the minimized objective; By combining PID adjustment parameters and MPC optimization parameters, the final infusion rate adjustment command is generated: Where α is the fusion coefficient.

8. The system according to any one of claims 1 to 7, characterized in that: In the first stage, the cloud-based historical case training set model is invoked. This model is built based on a hybrid kernel function optimization algorithm. By substituting the patient's preoperative multidimensional feature data, the similarity between the current patient and historical cases is calculated, and the most similar case group is selected. Based on the optimal dosage of similar case groups, and adjusted according to the individual characteristics of the current patients, a generative AI planning algorithm is used to predict the patients' drug sensitivity to propofol and remifentanil, and a personalized initial dosage plan is automatically generated.

9. The system according to any one of claims 1 to 7, characterized in that: In the second phase, target values ​​for the depth of anesthesia and vital signs are set. The system collects multi-dimensional data from monitors, anesthesia machines, and infusion pumps in real time via WIFI communication protocol. A hybrid model combining Bi-LSTM and attention mechanism was used to analyze the collected data and predict the trend of changes in anesthesia depth and the probability of risks such as hypoxemia and muscle relaxation recovery. Based on a composite closed-loop control algorithm, combined with multi-parameter weighted deviation, the infusion rates of propofol and remifentanil are dynamically adjusted. It receives patient vital sign feedback data in real time, forming a real-time closed loop of target setting → data collection → AI analysis → infusion regulation → vital sign feedback.

10. The system according to any one of claims 1 to 7, characterized in that: In the third stage, the system automatically links intraoperative control records with postoperative recovery indicators to form a complete anesthesia management data package; the data package is then uploaded to the cloud database to provide data support for model optimization. Based on postoperative data, the parameters of the generative AI planning algorithm model and the deep learning prediction algorithm model are optimized through transfer learning technology.

Images