Intelligent anesthesia monitoring and regulation system

Abstract

The invention discloses an intelligent anesthesia monitoring and regulation system, and relates to the technical field of medical monitoring and automatic control. The system comprises a data acquisition and preprocessing module, a dynamic physiological model construction module, an anesthesia state evaluation and prediction module, a real-time uncertainty quantification module and a layered hybrid closed-loop regulation and control module. The layered mixed closed-loop regulation and control module is used for predicting the change track of the anesthesia depth under different administration strategies in a few minutes in the future through a dynamic physiological model, and calculating an optimal administration rate sequence which can maintain the anesthesia depth in a target range and can minimize the drug dosage and blood pressure fluctuation through online optimization, so as to obtain the optimal administration rate sequence. The confidence score is referenced when controlling the administration rate. The system can realize highly personalized anesthesia monitoring and regulation and control, and has the advantages of quantifiability, high reliability, stable operation, accurate regulation and control and the like.

Description

Technical Field

[0001] This invention relates to the field of medical monitoring and automated control technology, and in particular to an intelligent anesthesia monitoring and control system based on dynamic physiological models and uncertainty quantification. Background Technology

[0002] In clinical practice, precise control of the depth of anesthesia or sedation is of great significance for ensuring patient safety and improving prognosis.

[0003] Current techniques for monitoring the depth of anesthesia, such as the bispectral index (BIS) or entropy index, primarily rely on processing individual EEG signals. These techniques have the following limitations: 1) The accuracy and stability of monitoring results decrease when faced with complex physiological states or interference from electrosurgical procedures; 2) The output is a single index, which cannot reflect the reliability or confidence level of that index at the current moment, often leaving clinicians in a dilemma of "to believe or not to believe" when faced with questionable values; 3) Monitoring is lagging and cannot predict future changes in the patient's condition.

[0004] In terms of anesthesia control, although target-controlled infusion (TCI) technology has been applied, it relies on static pharmacokinetic / pharmacodynamic (PK / PD) models based on population statistical data. This model cannot be adjusted intraoperatively according to the individual patient's real-time response, resulting in actual blood drug concentrations and clinical effects often deviating from expectations in special patients such as the elderly, obese, and those with heart failure, posing a risk of excessively deep or shallow anesthesia. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide an intelligent anesthesia monitoring and control system that can achieve highly personalized anesthesia monitoring and control, and is quantifiable, highly reliable, stable in operation, and accurate in control.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: an intelligent anesthesia monitoring and control system, comprising:

[0007] The data acquisition and preprocessing module is used to synchronously acquire the patient's multimodal physiological signals and drug input information through the monitoring equipment, and to preprocess the acquired multimodal physiological signals and drug input information;

[0008] The dynamic physiological model construction module is used to combine the intraoperative evolutionary physiological model with the LSTM network to construct a nonlinear dynamic physiological model. The intraoperative evolutionary physiological model and the LSTM network are used to predict and fuse the current anesthesia depth index DoA to obtain the fused current anesthesia depth index DoA.

[0009] The anesthesia status assessment and prediction module is used to accurately calculate the current depth of anesthesia index (DoA) based on a dynamic physiological model, and to predict the trend of DoA changes in the near future.

[0010] The real-time uncertainty quantification module receives the DoA index and trend from the anesthesia status assessment and prediction module, and integrates quality assessments from multiple information sources in real time. Through a preset weighted fusion algorithm, it calculates a three-dimensional dynamic quantitative confidence score ranging from 0 to 100%.

[0011] The hierarchical hybrid closed-loop control module is used to predict the trajectory of changes in the depth of anesthesia under different dosing strategies in the next few minutes through a dynamic physiological model. Through online optimization, it calculates the optimal dosing rate sequence that can maintain the depth of anesthesia within the target range while minimizing drug dosage and blood pressure fluctuations. The confidence score is used as a reference when controlling the dosing rate.

[0012] The beneficial effects of adopting the above technical solution are as follows: achieving a high degree of personalization and precision: in the system described in this application, the intraoperative evolutionary physiological model is combined with the LSTM network to construct a nonlinear dynamic physiological model, which can learn and adapt to the unique drug sensitivity of each patient in real time, thus achieving precise anesthesia that is "personalized".

[0013] It possesses effective foresight and stability: the MPC-based control strategy can anticipate and respond to intraoperative stimuli in advance, making the regulation of anesthesia / sedation depth more stable and helping to reduce blood pressure fluctuations.

[0014] Provides quantifiable reliability and safety: The three-dimensional dynamic quantitative confidence score provides clinicians with a basis for decision-making and serves as a control strategy for safe MPC to prevent the system from making erroneous adjustments when monitoring results are unreliable, thereby improving the safety of the automated system.

[0015] The convenience of automatic anesthesia depth adjustment: The MPC's control strategy outputs to the anesthesia pump, automatically adjusting the drug dosage according to the current state, minimizing the workload of medical staff. This system is designed not only for general anesthesia in the operating room but also for long-term sedation management in the ICU, effectively reducing the burden on medical staff and optimizing resource allocation. Attached Figure Description

[0016] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0017] Figure 1 This is a schematic block diagram of the system described in the embodiment of the present invention;

[0018] Figure 2 This is an interface diagram of the system described in an embodiment of the present invention;

[0019] Figure 3 This is a flowchart illustrating the construction and updating of the dynamic physiological model in the system described in this embodiment of the invention;

[0020] Figure 4 This is a logic flowchart of the hierarchical hybrid closed-loop control module according to an embodiment of the present invention;

[0021] Figure 5 This is a diagram of the hybrid prediction model architecture based on LSTM and intraoperative evolutionary physiological model in an embodiment of the present invention.

[0022] Figure 6 This is a graph showing the confidence level versus heart rate variation of system A in this embodiment of the invention.

[0023] Figure 7 This is a comparison chart of the DoA curves of system A and system B in an embodiment of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0025] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0026] like Figure 1 As shown, an embodiment of the present invention discloses an intelligent anesthesia monitoring and control system, comprising:

[0027] The data acquisition and preprocessing module is used to synchronously acquire the patient's multimodal physiological signals and drug input information through the monitoring equipment, and to preprocess the acquired multimodal physiological signals and drug input information;

[0028] The dynamic physiological model construction module is used to combine the intraoperative evolutionary physiological model with the LSTM network to construct a nonlinear dynamic physiological model. The intraoperative evolutionary physiological model and the LSTM network are used to predict and fuse the current anesthesia depth index DoA to obtain the fused current anesthesia depth index DoA.

[0029] The anesthesia status assessment and prediction module is used to accurately calculate the current depth of anesthesia index (DoA) based on a dynamic physiological model, and to predict the trend of DoA changes in the near future.

[0030] The real-time uncertainty quantification module receives the DoA index and trend from the anesthesia status assessment and prediction module, and integrates quality assessments from multiple information sources in real time. Through a preset weighted fusion algorithm, it calculates a three-dimensional dynamic quantitative confidence score ranging from 0 to 100%.

[0031] The hierarchical hybrid closed-loop control module is used to predict the trajectory of changes in the depth of anesthesia under different dosing strategies in the next few minutes through a dynamic physiological model. Through online optimization, it calculates the optimal dosing rate sequence that can maintain the depth of anesthesia within the target range while minimizing drug dosage and blood pressure fluctuations. The confidence score is used as a reference when controlling the dosing rate.

[0032] The user interface diagram of the system described in this invention is as follows: Figure 2 As shown, the above modules will be explained in detail below with reference to specific content:

[0033] Data Acquisition and Preprocessing Module 101: Simultaneously acquires multimodal physiological signals from the patient using standard monitoring equipment, including but not limited to EEG, ECG (for extracting heart rate and heart rate variability HRV), noninvasive blood pressure (NIBP), pulse oxygen saturation (SpO2), end-tidal carbon dioxide (EtCO2), skin conductance (EDA), and anesthetic drug (such as propofol) infusion rate. The acquired modal physiological signals undergo preprocessing such as noise reduction, artifact identification, and removal.

[0034] Dynamic physiological model construction module 102: It is used to combine the intraoperative evolutionary physiological model with the LSTM network to construct a nonlinear dynamic physiological model. The intraoperative evolutionary physiological model and the LSTM network are used to predict and fuse the current anesthesia depth index DoA to obtain the fused current anesthesia depth index DoA.

[0035] Initialization: Based on the patient's baseline information (such as age, weight, ASA classification, etc.) and PK / PD data, an initial physiological model is constructed.

[0036] Real-time learning and evolution: During anesthesia, real-time acquired multimodal physiological data is continuously used as input. Through pre-defined machine learning or state-space estimation algorithms, the core physiological parameters within the model (such as individual drug sensitivity, clearance rate, effect-site concentration, etc.) are updated and optimized in real-time and continuously, thereby generating an intraoperative evolutionary physiological model that dynamically reflects the patient's true physiological response during surgery. Its model structure: This invention integrates the standard three-compartment PK model and the SigmoidEmax PD model into an intraoperative evolutionary model (e.g., Figure 3 (As shown).

[0037] State Vector: State vector x k At time k, it is defined as:

[0038] x k =[C c (k),C p (k),C e (k),k e0 (k),EC 50 (k),γ(k)] T ;

[0039] Among them, C c C p C e These represent the drug concentrations in the central chamber, peripheral chamber, and effect chamber, respectively.

[0040] k e0 EC 50 γ and γ are pharmacodynamic parameters that need to be estimated in real time and reflect individual patient differences. Incorporating them into the state vector is the key to achieving real-time model evolution.

[0041] The state transition equation describes how the state vector evolves over time; its discrete form is x. k =f(x) k-1， u k-1， )+w k-1 , where u k-1， It is the drug infusion rate, w k-1 It is process noise, and the specific equations are obtained by discretizing the differential equations of the three-compartment model:

[0042]

[0043] For the parameter to be estimated, assuming it changes slowly, it is modeled as a random walk:

[0044] k e0 (k)=k e0 (k-1)+wk e0 EC50 The same logic applies to γ.

[0045] The observation equation describes how to derive a measurable physiological signal from a state vector. The observed value z... k The DoA (Depth of Anesthesia) index is obtained from EEG signal processing.

[0046]

[0047] Where h(x) k ) represents the PD model, v k It is observation noise; this equation is nonlinear, so nonlinear filtering methods such as UKF are required.

[0048] Key parameters of the model in this application, such as the drug effect-site half-life k, e0 50% maximum effect concentration EC 50 The Hill coefficient γ is no longer a fixed constant, but is regarded as a state variable to be estimated in the system.

[0049] The algorithm employs either Unscented Kalman Filter (UKF) or Particle Filter. This algorithm uses preprocessed real-time physiological data (especially EEG features) as observations and continuously applies them to k... e0 EC 50 The state vector, including γ, is iteratively estimated and updated. UKF / particle filtering can effectively handle the high nonlinearity of physiological systems, thereby enabling real-time and dynamic personalization of the model.

[0050] The UKF-based implementation process: The UKF algorithm approximates the state distribution through a series of deterministic sampling points (Sigma points), thereby handling nonlinear problems. Its iterative process at each time step k is as follows (e.g., Figure 3 As shown):

[0051] (1) Prediction:

[0052] Based on the state estimate of k-1 at the previous time step Covariance p k-1 Calculate a set of Sigma points.

[0053] Each Sigma point is propagated through the nonlinear state transition equation f(·) to obtain the predicted Sigma point.

[0054] Based on the predicted Sigma points, a weighted prior state estimate is calculated. and prior covariance p k|k-1 .

[0055] (2) Update:

[0056] The predicted Sigma point is propagated through the nonlinear observation equation h(·) to obtain a set of predicted observations.

[0057] The predicted observed mean is calculated using a weighted average.

[0058] Calculate the cross-covariance P of state and observation. xz Covariance P of observations zz .

[0059] Calculate Kalman gain

[0060] Combined with the actual observed value z k (i.e., DoAmeasured), update the state estimate:

[0061]

[0062] Update covariance:

[0063] Through the above cycle, UKF continuously utilizes new observational data z. k To correct the state vector x k The estimate, thus updating k in real time. e0 EC 50 Personalized parameters such as γ enable the model to dynamically approximate the patient's actual physiological condition.

[0064] Anesthesia Status Assessment and Prediction Module 103: Based on the aforementioned dynamic intraoperative evolutionary physiological model (Module 102), it accurately calculates the current depth of anesthesia quantitative index (DoA Index) and predicts the trend of anesthesia depth changes in the near future (e.g., 1-5 minutes).

[0065] Status Assessment: Based on the aforementioned dynamically evolving intraoperative evolutionary physiological model (module 102), the quantitative index of the current depth of anesthesia is accurately calculated, referred to in this invention as the Depth of Anesthesia Index (DoA). This module first obtains the real-time estimated drug effect-site concentration Ce from the intraoperative evolutionary physiological model. Subsequently, the transient effect of the drug is calculated using the Sigmoid Emax pharmacodynamic model in the model:

[0066]

[0067] Among them, EC 50Both the 50% maximum effect concentration and γ (Hill coefficient) are personalized parameters estimated in real time by a dynamic physiological model. The calculated Effect value is transformed linearly or nonlinearly and calibrated into a DoA index ranging from 0 (complete inhibition) to 100 (complete awakening), thereby achieving a quantitative assessment of the depth of anesthesia.

[0068] Trend prediction: Using this model to perform forward extrapolation, the changing trend of anesthesia depth in the short term (e.g., 1-5 minutes) can be predicted, providing a basis for prospective regulation.

[0069] The specific implementation process of trend prediction is as follows:

[0070] Obtain the current state: Obtain the current optimally estimated state vector from the dynamic physiological model.

[0071] Define future input: Define a drug infusion sequence within a short future time period (e.g., Np time steps). In the absence of control commands, it can be assumed that the current infusion rate will be maintained, i.e., u k+1 =u k-1 .

[0072] Forward iterative simulation: Using the established state transition equation f(·), from the current state... Begin iterative calculations:

[0073]

[0074] Generate predicted trajectories: This involves generating the future state sequence obtained from forward simulation. The observation equation h(·) is mapped to the future DoA prediction trajectory:

[0075] The predicted trajectory represents the trend of anesthesia depth, which will be sent to the MPC controller in the hierarchical hybrid closed-loop control module 105 as the basis for its optimization decision.

[0076]

[0077] Real-time uncertainty quantification module 104: Based on the current state and predicted evaluation value obtained from the anesthesia state assessment and prediction module 103, the confidence level is calculated using the following method. A 0-100% "three-dimensional dynamic quantification confidence score" is calculated, which comprehensively reflects the reliability of the output index of the anesthesia state assessment and prediction module 103.

[0078] The real-time uncertainty quantification module 104 receives the DoA index and trend from the anesthesia status assessment and prediction module 103, and integrates quality assessments from multiple information sources in real time. Through a preset weighted fusion algorithm, it calculates a confidence score ranging from 0% to 100%. This score comprehensively evaluates the reliability of the output index of the anesthesia status assessment and prediction module (103).

[0079] The score was calculated by taking into account information from the following three dimensions:

[0080] (1) Multi-source signal quality (SQI): Real-time signal quality score of each sensor (EEG, ECG, etc.).

[0081] (2) Model Fit Residual: The residual between the observed and predicted values ​​generated by the UKF / Particle Filter algorithm in the update step. The smaller the residual, the better the model fit and the higher the confidence level.

[0082] (3) Physiological System Consistency Score: This score uniquely assesses the consistency between the central nervous system (from EEG) and the autonomic nervous system (from HRV and EDA). For example, when the EEG shows deep anesthesia, but the HRV and EDA show severe stress, it is judged as a "physiological conflict," and the score is reduced, thus lowering the overall confidence level. These three scores are integrated into a final "three-dimensional dynamic quantitative confidence score" through a weighted algorithm or fuzzy logic reasoning system.

[0083] These three scores are ultimately fused into a final three-dimensional dynamic quantitative confidence score using a weighted fusion algorithm. For example, a linear weighted model can be used:

[0084] Confidence score = w1·Score SQI +w2·Score 残差 +w3·Score 一致性

[0085] Among them, w1, w2, and w3 are preset weight coefficients that satisfy w1+w2+w3=1. These weights can be calibrated based on clinical experience data to reflect the importance of different dimensions of information in specific clinical scenarios.

[0086] Layered hybrid closed-loop control module 105: Automatically adjusts the infusion rate of the anesthesia pump based on the current value and future trend obtained from the anesthesia status assessment and prediction module.

[0087] Model Predictive Control (MPC) is used for control: MPC, as the core, utilizes an "intraoperative evolutionary physiological model" to predict the trajectory of changes in anesthetic depth under different dosing strategies over the next few minutes. Through online optimization, it calculates the optimal dosing rate sequence that maintains anesthetic depth within the target range while minimizing drug dosage and blood pressure fluctuations.

[0088] The online optimization process of MPC, in each control cycle k, is essentially solving a finite-time optimization problem. The specific implementation is as follows:

[0089] (1) Objective Function: The controller MPC calculates the optimal dosing rate sequence by minimizing an objective function J. This objective function comprehensively considers the accuracy of anesthesia depth control and the smoothness of dosing:

[0090]

[0091] in:

[0092] N p It refers to the prediction time domain, that is, the time it takes for the controller to anticipate the future.

[0093] N c It refers to the control time domain, specifically the number of times the controller plans to change future dosing patterns (N). c ≤N p ).

[0094] It is the DoA value predicted at time k+i in the future from time k.

[0095] r(k+i) is a preset DoA target value (e.g., 50), which forms the reference trajectory.

[0096] Δu(k+i)=u(k+i)-u(k+i-1) is the change in the drug administration rate.

[0097] Q and R are weighting coefficients. A larger Q means that the controller focuses more on accurately tracking the target DoA; a larger R means that the controller focuses more on the stability of control and avoids drastic changes in dosage.

[0098] (2) Constraints: The optimization process must be carried out within clinical safety and physical limitations:

[0099] DoA range constraints: (For example, 40≤DoA≤60).

[0100] Dosing rate constraint: u min ≤u(k+i)≤u max (For example, 0 ≤ propofol rate ≤ 20 ml / h).

[0101] Dosing rate variation constraint: Δu min ≤Δu(k+i)≤Δu max (To prevent sudden changes in the infusion pump rate.)

[0102] (3) Online solution: In each control cycle, MPC uses a dynamic physiological model to make predictions and solves the above-mentioned constrained optimization problem:

[0103]

[0104] Where ΔU=[Δu(k),…,Δu(k+N c [-1)] is the control sequence to be optimized. Solving this problem (usually using numerical optimization algorithms such as quadratic programming) yields an optimal control sequence.

[0105] (4) Rolling optimization: MPC executes only the first control action u(k) = u(k-1) + Δu*(k) in the optimal control sequence. In the next time step, the system collects new measurements, updates the model, and repeats the entire prediction and optimization process. This rolling optimization method allows the control to adapt to changes in patient status and disturbances in real time.

[0106] The optimal dosing rate output by the MPC is not directly sent to the anesthesia pump; it is first processed by this safety regulator. This regulator modulates the MPC commands based on the "three-dimensional dynamic quantitative confidence score" output by module 104, such as... Figure 4 As shown:

[0107] 1) If the confidence level of the three-dimensional dynamic quantization is >90%, then the instructions of MPC are fully trusted to achieve fully automated and precise control.

[0108] 2) If the confidence level of the three-dimensional dynamic quantization is between 70% and 90%, adopt the direction of the MPC instruction, but limit the magnitude of its change and carry out conservative control.

[0109] 3) If the confidence level of the three-dimensional dynamic quantification is less than 70%, the system determines that the current state is unreliable, suspends automatic control, maintains the current dosing rate, and immediately issues a high-priority alarm to the clinician, suggesting manual evaluation.

[0110] Prediction Enhancement: To further improve prediction accuracy, especially in capturing complex dynamics that are difficult to describe by mechanistic models, a data-driven model (LSTM) network can be introduced. Its implementation and its integration with MPC are as follows: Figure 5 As shown:

[0111] LSTM model architecture: The LSTM used is a typical sequence-to-sequence (Seq2Seq) model.

[0112] Input layer: Receives a multidimensional time series containing the past Th time steps, with features including: historical DoA values, heart rate, blood pressure, drug infusion rate, etc.

[0113] Encoder: Composed of several stacked LSTM layers, it encodes the input time series into a fixed-dimensional context vector, which contains the essence of historical information.

[0114] Decoder: Also composed of LSTM layers, it receives the context vector and gradually generates the DoA prediction sequence for the next Np time steps.

[0115] Output layer: A fully connected layer that maps the decoder's output to the final DoA prediction.

[0116] Working mechanism with MPC: The prediction results of LSTM do not directly replace the intraoperative evolutionary physiological model, but are deeply integrated with it through one of the following two methods to jointly provide predictive input for MPC:

[0117] Method 1: Residual Prediction In this mode, LSTM does not directly predict DoA, but learns the prediction error (residual) of the prediction mechanism model.

[0118] 1) Intraoperative evolutionary physiological models provide a basic predictive mechanism for DoA prediction.

[0119] 2) The LSTM network predicts the possible bias of the basic prediction, DoA prediction residual, based on historical data.

[0120] 3) The final fused prediction value fed into MPC is: DoA prediction fusion (k+i) = DoA prediction_mechanism (k+i) + DoA prediction_residual (k+i). This method preserves the interpretability of the mechanistic model while using the data model to correct its shortcomings.

[0121] Method 2: Confidence-Weighted Fusion. In this mode, the intraoperative evolutionary physiological model and the LSTM model make independent DoA predictions in parallel.

[0122] 1) Mechanism of DoA prediction output from intraoperative evolutionary physiological model.

[0123] LSTM model output prediction DoA prediction data.

[0124] The fusion module dynamically adjusts the weight α of the two modules based on the confidence score output by the real-time uncertainty quantification module. k DoA prediction_fusion(k+i)=α k DoA prediction mechanism (k+i)+(1-α) k DoA prediction_data(k+i) When the confidence score is high, αk When the confidence score is close to 1, the system trusts the mechanistic model more; when the confidence score is low (e.g., due to strong interference or physiological conflict), α... k As the data decreases, the system will refer more to the prediction results of the LSTM model, which learns patterns from massive amounts of data.

[0125] To verify the beneficial effects of this invention, simulation tests were conducted in this application:

[0126] Simulation Scenario: A 65-year-old male patient undergoes laparoscopic cholecystectomy. During the anesthesia maintenance phase, a severe nociceptive event is simulated, caused by the surgeon performing peritoneal traction.

[0127] Comparison System:

[0128] System A: The intelligent control system described in this invention.

[0129] System B: Traditional TCI system, which uses a fixed population PK / PD model and sets the target depth of anesthesia (simulated by BIS) to 50.

[0130] Simulation process and results: Refer to Figure 6 and Figure 7 .

[0131] Before stimulation (T = 0-10 minutes): Both System A and System B could maintain DoA at around 50. However, the "intraoperative evolutionary physiological model" module of System A had fine-tuned the personalized parameters based on the patient's steady-state response during this period, and found that the patient was more sensitive to propofol.

[0132] Strong stimulus occurs (T=10 minutes):

[0133] System B (Traditional TCI): Approximately 45-60 seconds after stimulation, the DoA index begins to rise passively, reaching a peak of 64 at T=11.5 minutes (insufficient anesthesia). At this point, the TCI system begins to significantly increase the drug delivery rate to "compensate," causing the DoA to subsequently drop to 40 (insufficient anesthesia), resulting in a significant drop in blood pressure.

[0134] System A (this invention):

[0135] Just after T=10 minutes, the patient's heart rate (HR) and skin conductance (EDA) immediately showed a transient increase. At this time, the DoA index was still 51 and had not changed.

[0136] However, the uncertainty quantification module of this invention immediately detected the physiological conflict between the EEG state and the autonomic nervous system state, and the confidence score rapidly dropped from 95% to 75% (see...). Figure 6 It also sends a basic message to the doctor that "physiological signals are inconsistent".

[0137] Almost simultaneously, the MPC controller used an "intraoperative evolutionary physiological model" for forward prediction, which showed that the current dosing rate was insufficient to suppress the impending stimulus and predicted that DoA would exceed 60 in 1 minute. The MPC immediately and proactively increased the propofol infusion rate.

[0138] Due to timely intervention, System A's DoA index dropped slightly to 46, then quickly stabilized back to around 50 (e.g., Figure 7 As shown in the figure, the anesthesia was neither too light nor too deep throughout the entire process, and the blood pressure fluctuations were minimal.

[0139] Conclusion of the embodiments: The system of the present invention, through its two core innovations, demonstrates excellent forward-looking control capabilities and safety, which are significantly superior to the existing technologies.

[0140] Clinical application scenarios:

[0141] General anesthesia in the operating room:

[0142] This invention is an ideal choice for various types of general anesthesia surgeries, especially suitable for special patients such as those with cardiovascular insufficiency, the elderly, and obese patients, as well as surgeries with extremely high requirements for depth of anesthesia and hemodynamic stability (such as cardiothoracic surgery and neurosurgery). The system can automatically and smoothly navigate the intense stimulation stages such as intubation, skin incision, and abdominal closure, reducing the incidence of intraoperative awareness and excessive anesthesia.

[0143] ICU sedation management:

[0144] Prolonged sedation of patients on ventilators in the ICU is a major challenge. This system can achieve:

[0145] Stable sedation: Maintaining the patient's sedation level (such as RASS score) precisely within the target range over a long period of time to avoid problems such as prolonged ventilator dependence and extended ICU stay caused by excessive sedation.

[0146] Automated daily wake-up: Daily sedation interruption can be executed programmatically, smoothly reducing dosage, waking up, assessing, and then automatically resuming sedation, greatly reducing the workload of nurses.

[0147] Delirium prevention: Maintaining a more physiological state of sedation may help reduce the incidence of delirium in ICU patients.

[0148] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An intelligent anesthesia monitoring and control system, characterized in that... include: The data acquisition and preprocessing module is used to synchronously acquire the patient's multimodal physiological signals and drug input information through the monitoring equipment, and to preprocess the acquired multimodal physiological signals and drug input information; The dynamic physiological model construction module is used to combine the intraoperative evolutionary physiological model with the LSTM network to construct a nonlinear dynamic physiological model. The intraoperative evolutionary physiological model and the LSTM network are used to predict and fuse the current anesthesia depth index DoA to obtain the fused current anesthesia depth index DoA. The anesthesia status assessment and prediction module is used to accurately calculate the current depth of anesthesia index (DoA) based on a dynamic physiological model, and to predict the trend of DoA changes in the near future. The real-time uncertainty quantification module receives the DoA index and trend from the anesthesia status assessment and prediction module, and integrates quality assessments from multiple information sources in real time. Through a preset weighted fusion algorithm, it calculates a three-dimensional dynamic quantitative confidence score ranging from 0 to 100%. The hierarchical hybrid closed-loop control module is used to predict the trajectory of changes in the depth of anesthesia under different dosing strategies in the next few minutes through a dynamic physiological model. Through online optimization, it calculates the optimal dosing rate sequence that can maintain the depth of anesthesia within the target range while minimizing drug dosage and blood pressure fluctuations. The confidence score is used as a reference when controlling the dosing rate.

2. The intelligent anesthesia monitoring and control system as described in claim 1, characterized in that: The multimodal physiological signals include EEG signals, ECG signals, NIBP signals, SpO2 signals, EtCO2 signals, EDA signals, and anesthetic infusion rate signals.

3. The intelligent anesthesia monitoring and control system as described in claim 1, characterized in that: The intraoperative evolutionary physiological model integrates the standard three-compartment PK model and the Sigmoid Emax PD model into a nonlinear intraoperative evolutionary physiological model.

4. The intelligent anesthesia monitoring and control system as described in claim 3, characterized in that: In the intraoperative evolutionary physiological model: State vector: State vector x k At time k, it is defined as: x k =[C c (k),C p (k),C e (k),k e0 (k),EC 50 (k),γ(k)] T ; Among them, C c C p C e These represent the drug concentrations in the central chamber, peripheral chamber, and effect chamber, respectively. k e0 EC 50 γ and γ are pharmacodynamic parameters that need to be estimated in real time and reflect individual patient differences; State transition equation: used to describe how the state vector evolves over time, its discrete form is: x k =f(x) k-1， u k-1， )+w k-1 , where u k-1， It is the drug infusion rate, w k-1 It is process noise, and the specific equations are obtained by discretizing the differential equations of the three-compartment model: For the parameter to be estimated, assuming it changes slowly, it is modeled as a random walk: k e0 (k)=k e0 (k-1)+wk e0 EC 50 The same applies to γ; Observation equation: used to describe how to obtain a measurable physiological signal from the state vector, the observed value z. k It is the DoA index, which is obtained from EEG signal processing; Where h(x) k ) represents the PD model, v k It is observation noise.

5. The intelligent anesthesia monitoring and control system as described in claim 1, characterized in that, The LSTM model architecture is a sequence-to-sequence model, including: Input layer: Receives a multidimensional time series including the past Th time steps; Encoder: Composed of several stacked LSTM layers, it encodes the input time series into a fixed-dimensional context vector, which includes historical information; Decoder: Composed of several LSTM layers, it receives the context vector and gradually generates the DoA prediction sequence for the next Np time steps; Output layer: Includes a fully connected layer that maps the decoder output to the final DoA prediction.

6. The intelligent anesthesia monitoring and control system as described in claim 1, characterized in that: The anesthesia status assessment and prediction module is implemented through the following method: Based on a dynamic physiological model of dynamic evolution, the real-time estimated drug concentration C in the effect compartment is first obtained from the dynamic physiological model. e Subsequently, the transient effect of the drug was calculated using the Sigmoid Emax PD model in the model: Among them, EC 50 The concentration is 50% of the maximum effect concentration, and γ is the Hill coefficient. The calculated Effect value is transformed linearly or nonlinearly and calibrated as the DoA index, which ranges from 0 to 100, to achieve a quantitative assessment of the depth of anesthesia.

7. The intelligent anesthesia monitoring and control system as described in claim 1, characterized in that: The trend prediction method for the Depth of Anesthesia (DoA) index includes the following steps: Obtain the current state: Obtain the current optimally estimated state vector from the dynamic physiological model. Set future input: Set a drug infusion sequence for a short future period. In the absence of control commands, assume the current infusion rate is maintained, i.e., u k+1 =u k-1 ; Forward iterative simulation: Using the established state transition equation f(·), from the current state... Begin iterative calculations: Generate predicted trajectories: This involves generating the future state sequence obtained from forward simulation. The observation equation h(·) is mapped to the future DoA prediction trajectory: The predicted trajectory represents the trend of anesthesia depth. This trend is fed into the MPC controller in the hierarchical hybrid closed-loop control module as the basis for its optimization decisions.

8. The intelligent anesthesia monitoring and control system as described in claim 1, characterized in that: The confidence score in the real-time uncertainty quantification module is used to comprehensively evaluate the reliability of the output index of the anesthesia status assessment and prediction module. Its implementation methods include: 1) Multi-source signal quality index (SQI): Real-time signal quality score for each sensor; 2) Model fit residuals: The residuals between the observed and predicted values ​​generated in the update step. The smaller the residuals, the better the model fit and the higher the confidence level. 3) Physiological system consistency score; assesses the consistency between the central nervous system (EEG) and the autonomic nervous system (HRV, EDA). When the EEG shows deep anesthesia, but the HRV and EDA show strong stress, it is judged as a physiological conflict, and this score is reduced, thus lowering the overall confidence level. The above scores are combined using a weighted fusion algorithm to form a final three-dimensional dynamic quantitative confidence score: Confidence score = w1·Score SQI +w2·Score 残差 +w3·Score 一致性 ; Among them, w1, w2, and w3 are preset weight coefficients, satisfying w1 + w2 + w3 = 1.

9. The intelligent anesthesia monitoring and control system as described in claim 1, characterized in that, The hierarchical hybrid closed-loop control module, including the model prediction controller (MPC), is specifically implemented as follows: (1) Objective function: The MPC controller calculates the optimal dosing rate sequence by minimizing an objective function J, which comprehensively considers the accuracy of anesthesia depth control and the smoothness of dosing: Where: N p This is the prediction time domain, where N represents the duration predicted by the controller. c This is the control time domain, where N represents the number of future dosing changes planned by the controller. c ≤N p ; It is the DoA value predicted at time k+i in the future at time k; r(k+i) is the preset DoA target value, which constitutes the reference trajectory; Δu(k+i)=u(k+i)-u(k+i-1) is the change in the dosing rate; Q and R are weighting coefficients. The larger Q is, the more the controller focuses on accurately tracking the target DoA; the larger R is, the more it focuses on the stability of the control and avoids drastic changes in the dosage. (2) Constraints: The optimization process is conducted within clinical safety and physical limitations. DoA range constraints: Dosing rate constraint: u min ≤u(k+i)≤u max ; Dosing rate variation constraint: Δu min ≤Δu(k+i)≤Δu max ; (3) Online solution: In each control cycle, the Model Predictive Controller (MPC) uses the dynamic physiological model to make predictions and solve the above-mentioned constrained optimization problem: Among them, ΔU=[Δu(k),…,Δu(k+N c [-1)] is the control sequence to be optimized. After solving this problem, an optimal control sequence is obtained. (4) Rolling optimization: The Model Predictive Controller (MPC) executes only the first control action u(k) = u(k-1) + Δu*(k) in the optimal control sequence. At the next time step, new measurements are collected, the model is updated, and the entire prediction and optimization process is repeated.

10. The intelligent anesthesia monitoring and control system as described in claim 9, characterized in that, The optimal dosing rate output by the model predictive controller (MPC) is first processed by the MPC. The MPC modulates its instructions based on a three-dimensional dynamic quantization confidence score. If the confidence level of the three-dimensional dynamic quantization is >90%, then the instructions of MPC are fully trusted to achieve fully automated and precise control. If the confidence level of the three-dimensional dynamic quantization is between 70% and 90%, then the direction of the MPC instruction will be adopted, but the magnitude of its change will be limited for conservative control. If the confidence level of the three-dimensional dynamic quantification is less than 70%, the current state is deemed unreliable. Automated control is suspended, the current dosing rate is maintained, and a high-priority alert is immediately issued to the clinician, with manual evaluation recommended.

Images