Perioperative anesthesia method and system based on prospective early warning and regulation

Abstract

The invention provides a perioperative anesthesia method and system based on prospective early warning and regulation. The method comprises the following steps: S1, carrying out data acquisition and preprocessing; s2, a step of introducing a prospective early warning model is carried out; s3, a step of constructing an intelligent agent based on the prospective early warning model is carried out; s4, performing paradigm training based on reinforcement learning; and S5, carrying out personalized anesthesia control. According to the automatic anesthesia system in the perioperative period, prospective prediction and intelligent drug regulation and control can be combined, and precise control over the anesthesia depth is achieved; meanwhile, dosage decisions are optimized in different patient groups by utilizing a reinforcement learning strategy, and the drug dosage is reduced; the online learning module can realize personalized adjustment to adapt to individual differences of old people, obese patients and high-risk patients; the system can early warn risks in an operation in advance, and operation safety and postoperative recovery quality are ensured.

Description

Technical Field

[0001] This invention specifically relates to a perioperative anesthesia method and system based on proactive early warning and control. Background Technology

[0002] Automated anesthesia technology began in the 1980s and has been developed and widely used over the past few decades. This technology improves the quality of care for patients by reducing the repetitive operations of anesthesiologists in drug control, allowing them to focus on the most critical decisions in the case. In addition, automated anesthesia has shown great potential in optimizing drug infusion protocols and achieving precise anesthesia. For example, the study by Pasin et al. [1] showed that total intravenous anesthesia guided by bispectral index (BIS) of electroencephalogram can significantly reduce the amount of drugs used during induction, better maintain the depth of anesthesia and shorten the patient's recovery time. The multicenter study by Puri et al. [2] further verified the stability and superiority of automated anesthesia systems in various surgical settings. The retrospective study by Brogi et al. [3] found that automated anesthesia systems can effectively reduce the overshoot or undershoot of physiological indicators and significantly prolong the time that indicators are maintained in the target range.

[0003] Currently, the development of automated anesthesia infusion technology is often closely related to the progress of artificial intelligence. Traditional methods such as proportional-integral-derivative (PID) controllers [4] and model predictive controllers (MPC) [5] have been applied to the infusion control of anesthetic drugs, but these methods usually rely on linear assumptions and are difficult to adapt to complex nonlinear physiological systems. To this end, Moore et al. first proposed a method based on discrete reinforcement learning (RL) for anesthesia control, which achieved better results than the traditional PID controller [6]. Subsequently, Lowery et al. [7] extended reinforcement learning to continuous space, while Schamberg et al. [8] further used an advanced actor-critic algorithm to train the automated anesthesia agent. Recently, Yun et al. [9] proposed a hierarchical reinforcement learning algorithm, which uses high-level policies to generate target BIS trajectories and low-level policies to achieve more stable infusion control.

[0004] While existing research has demonstrated the broad application prospects of automated anesthesia technology, several problems remain to be addressed in current systems. Most existing methods rely heavily on historical physiological data as input, failing to fully utilize the predictive and early warning capabilities for future surgical characteristics. Furthermore, traditional control methods are often limited to simple linear modeling, making it difficult to handle the variability in complex physiological systems. To address these issues, we propose a perioperative automated anesthesia system based on prospective early warning and control. Compared to existing methods, this system can further improve the accuracy of anesthesia strategies through the prediction of future information and employs offline reinforcement learning to avoid dependence on environmental modeling, thereby enhancing the safety and stability of automated anesthesia.

[0005] Despite significant advancements in automated anesthesia control technology over the past few decades, including PID controllers, model predictive control (MPC), closed-loop infusion systems, and the rapidly developing reinforcement learning methods, existing technologies still have several shortcomings in practical applications, mainly in the following aspects.

[0006] First, there is a lack of predictive ability regarding future physiological states. Most existing closed-loop anesthesia control systems rely primarily on current and historical physiological parameters as input, such as BIS values, blood pressure, heart rate, or drug infusion rates. These systems often only adjust based on instantaneous feedback and cannot predict future physiological changes in the patient. For example, traditional algorithms typically cannot anticipate when surgical stimulation is imminent or when the patient is about to enter the awakening phase, which can easily lead to delayed adjustment of anesthesia depth and slow response to drug dosage, potentially resulting in excessively deep anesthesia or delayed awakening. This lack of forward-looking predictive ability is a significant limitation of existing automated anesthesia systems in terms of safety and accuracy.

[0007] Second, there is insufficient modeling capability for complex nonlinear physiological systems. Traditional control methods such as PID and MPC are mostly based on linear or quasi-linear control assumptions, which limit their effectiveness in handling complex nonlinear dynamics, physiological feedback mechanisms, and individual differences. Patients' pharmacokinetic and pharmacodynamic (PK / PD) parameters vary significantly among individuals and change dynamically over time during surgery. Traditional controllers that rely on fixed parameters or simplified models struggle to capture these changes, potentially leading to degraded control performance. Furthermore, MPC requires accurate state-space models when modeling complex systems, which is difficult and costly to implement in practice.

[0008] Third, reinforcement learning methods are heavily reliant on online interaction and lack safety and controllability. While recent methods such as RL, deep RL, and hierarchical RL have significantly improved control performance, most rely on simulated environments for training or online updates for policy improvement. These methods are highly dependent on the accuracy of the environment model, and online exploration introduces uncontrollable risks, making direct clinical application difficult. Furthermore, simulators used in existing RL studies often oversimplify real physiological processes, resulting in weak transferability of policies to real patients and a "model-reality gap." While offline RL can reduce risks, current methods still require large amounts of high-quality data and do not fully utilize pre-operative information or external auxiliary information available during surgery.

[0009] Fourth, existing closed-loop systems lack the integration and utilization of surgical task characteristics. Most existing studies only focus on controlling the depth of anesthesia itself, without explicitly incorporating key surgical characteristics such as surgical stage, future stimulus intensity, surgical procedure type, and patient position changes into the control system. This results in the system's inability to automatically adjust control strategies according to different surgical stages, and the inability to establish a proactive early warning mechanism based on surgical events. For example, at points where significant nociceptive stimulation may occur, existing systems often rely solely on post-operative BIS or hemodynamic responses for retrospective regulation, lacking proactive control capabilities.

[0010] Fifth, there is a lack of holistic perioperative management capabilities. Most existing automated anesthesia technologies only address intraoperative anesthesia depth control, failing to systematically cover preoperative prediction, intraoperative dynamic control, and postoperative recovery assessment. In practical clinical applications, this may prevent the system from forming a complete anesthesia decision-making chain, reducing the consistency and stability of perioperative management.

[0011] Based on the aforementioned shortcomings, existing automated anesthesia technologies still have significant room for improvement in terms of accuracy, safety, real-time performance, and intelligence. Therefore, it is necessary to propose a novel perioperative intelligent anesthesia control system that integrates future prediction capabilities, nonlinear modeling capabilities, offline reinforcement learning strategies, and surgical feature perception capabilities to overcome the deficiencies of existing technologies and further enhance their clinical value. Summary of the Invention

[0012] The purpose of this invention is to address the shortcomings of existing technologies by providing a perioperative anesthesia method based on prospective early warning and control, which can effectively solve the aforementioned problems.

[0013] To achieve the above requirements, the technical solution adopted by the present invention is: to provide a perioperative anesthesia method based on prospective early warning and control, which includes the following steps:

[0014] S1: Steps for data acquisition and preprocessing;

[0015] S2: Proceed with the steps of introducing a forward-looking early warning model;

[0016] S3: The steps to build an intelligent agent based on a forward-looking early warning model;

[0017] S4: Steps for training based on the reinforcement learning paradigm;

[0018] S5: Steps for individualized anesthesia control.

[0019] The advantages of this perioperative anesthesia method based on prospective early warning and control are as follows:

[0020] The perioperative automated anesthesia system of this invention combines prospective prediction with intelligent drug regulation to achieve precise control of the depth of anesthesia; at the same time, it uses reinforcement learning strategies to optimize dosage decisions in different patient groups and reduce drug dosage; and the online learning module can achieve personalized adjustments to adapt to individual differences in elderly, obese and high-risk patients; the system can provide early warning of intraoperative risks to ensure surgical safety and postoperative recovery quality. Attached Figure Description

[0021] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, use the same reference numerals to denote the same or similar parts. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0022] Figure 1 A schematic diagram of a perioperative anesthesia method based on prospective early warning and control according to an embodiment of this application is shown. Detailed Implementation

[0023] To make the objectives, technical solutions and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and specific embodiments.

[0024] In the following description, references to "an embodiment," "an embodiment," "an example," "example," etc., indicate that the described embodiment or example may include a particular feature, structure, characteristic, property, element, or limitation, but not every embodiment or example necessarily includes that particular feature, structure, characteristic, property, element, or limitation. Furthermore, the repeated use of the phrase "an embodiment according to this application," while possibly referring to the same embodiment, does not necessarily refer to the same embodiment.

[0025] For simplicity, certain technical features known to those skilled in the art are omitted in the following description.

[0026] According to one embodiment of this application, a perioperative anesthesia method based on prospective early warning and control is provided, such as... Figure 1 As shown, it includes the following steps:

[0027] Step 1: Data Acquisition and Preprocessing

[0028] First, the system employs various high-precision physiological signal acquisition devices (such as a 12-lead electrocardiogram (ECG) machine, a finger-clip pulse oximeter, etc.). The system uses devices such as photoplethysmography (PPG) sensors and gas flow sensors to monitor patients in real time. These devices can provide data with millisecond-level resolution, including but not limited to basic physiological information (such as resting heart rate and baseline blood oxygen levels), key physiological parameters during surgery (such as intraoperative mean arterial pressure (MAP), blood oxygen saturation fluctuation range, and depth of anesthesia index (BIS)), and the patient's historical anesthesia records (such as previous dosages of anesthetic drugs and adverse reactions).

[0029] To ensure data accuracy and usability, all collected data underwent rigorous preprocessing. For example, ECG signals were bandpass filtered (0.5-45 Hz) to remove baseline drift and power frequency interference, blood oxygen saturation data were subjected to Kalman filtering to reduce measurement noise, and heart rate and respiratory rate signals were denoised using wavelet transform. Furthermore, all data were standardized (Z-score normalization) to eliminate the impact of individual differences on model training, thereby improving the stability and prediction accuracy of the machine learning model.

[0030] Step 2: Introduce a forward-looking early warning model

[0031] One of the core elements of this patent is the prospective early warning model (or human transfer model P), which aims to predict future physiological trends based on the patient's real-time physiological data and surgery-related parameters. The model's prediction scope covers dynamic changes in anesthesia depth, adjustments in drug requirements, and fluctuations in key physiological indicators, thereby providing a scientific basis for precision anesthesia and improving intraoperative safety and stability.

[0032] However, while traditional pharmacokinetic / pharmacodynamic (PK / PD) models are widely used in clinical anesthesia, they still have many limitations. First, PK / PD models are primarily built based on population-average parameters, making it difficult to accurately model individual patient physiological characteristics. For example, key factors affecting drug metabolism—age, weight, liver and kidney function, and genetic metabolic traits—are often not fully considered within the traditional PK / PD framework, leading to biases in individualized predictions. For instance, Remifentanil has a lower clearance rate in elderly patients, but PK / PD models often underestimate its duration of effect, thus affecting the accurate prediction of anesthesia depth. Furthermore, PK / PD models have limited adjustment capabilities, typically using first- or second-order differential equations to describe changes in drug concentration over time, making it difficult to dynamically adapt to fluctuations in individual physiological states. In the event of intraoperative emergencies (such as hypotension, painful stimuli, or abnormal metabolism), the response of traditional models exhibits significant lag, making it difficult to provide real-time adjustment strategies and leading to instability in drug dosage recommendations.

[0033] Furthermore, traditional PK / PD models have limitations in terms of data input dimensions, primarily relying on drug concentration-response relationships and failing to fully integrate multimodal physiological signals. For example, the BIS index (bispectral index) is an important parameter for assessing the depth of anesthesia, but PK / PD models struggle to integrate it with EEG signals, thus affecting prediction accuracy. Simultaneously, because PK / PD models mostly use linear or exponential functions to fit drug metabolism curves, they struggle to accurately describe complex nonlinear drug-physiological interactions. For instance, the synergistic effect of propofol and remifentanil is not a simple linear additive relationship but is dynamically regulated by multiple physiological factors, and traditional PK / PD models have limited ability to model this.

[0034] To address the aforementioned issues, this patent's human transfer model proposes to employ a deep learning-based temporal model (including but not limited to: Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), Transformers, etc.). Any simple modifications, equivalent substitutions, and improvements made to the above embodiments within the spirit and principles of this invention are still within the protection scope of this invention. The input data X_t includes multidimensional physiological signals at the current moment, and the output predicted value is the distribution P(Y_{t+\Delta t}|X_t) of the future state. The model can generate real-time trends in the patient's physiological state and output early warning signals, providing guidance for subsequent regulation.

[0035] The model optimizes its predictions by minimizing the following loss function: ;

[0036] Where Y{t+\Delta t}^{(i)} is the true value, \hat{Y}{t+\Delta t}^{(i)} is the predicted value, \theta is the model parameter, \lambda is the regularization coefficient, and N is the number of samples.

[0037] The state space S_t of this invention comprises three key components:

[0038] 1. Current physiological data X_t:

[0039] Current patient physiological signals (such as ECG, blood oxygen saturation, heart rate, respiratory rate, etc.) provide real-time information on anesthesia depth, patient response, and other physiological indicators. This information reflects the immediate status of anesthesia and forms the basis for the system's real-time decision-making.

[0040] 2. Forward-looking forecast data: \hat{Y}_{t+\Delta t}

[0041] Through a prospective predictive model, the system can predict the trend of changes in anesthesia depth, fluctuations in drug demand, and possible changes in physiological status of patients within a certain time step (t + Delta t). This prediction not only improves the response speed of anesthesia control but also helps the system better handle the dependence on time series data when making decisions, avoiding the limitations of relying solely on current data. Prospective data enables the system to anticipate the needs for anesthesia depth and drug infusion in advance, thereby achieving proactive regulation.

[0042] 3. Historical drug infusion records D_{history}:

[0043] Combining these three pieces of information, the state space S_t of this invention is: ;

[0044] The system incorporates historical drug infusion quantities, rates, and variations into its state space to capture cumulative drug effects, changes in drug metabolism rates, and their relationship to patient physiological responses. This historical information helps the system understand the long-term cumulative effects of drug efficacy, thereby avoiding the adverse impact of short-term decisions on long-term anesthetic outcomes.

[0045] The core advantage of this model lies in its ability to predict the changing trends of patients during anesthesia based on a combination of current physiological state and historical data, and to output key early warning signals that may occur in the future, such as excessive or insufficient depth of anesthesia.

[0046] Step 3: Intelligent Agent Based on Forward-Looking Early Warning Model

[0047] It is divided into the following sub-steps:

[0048] 1. Traditional dose prediction:

[0049] First, the system still relies on the traditional agent to predict the initial dose in the historical state space x_t, generating the initial dose value \bar{a}_t: ;

[0050] This step preserves the stability of traditional methods, ensuring the system has a reliable predictive foundation in the initial stage. F will be modeled using a Transformer model. The model's input includes the current multidimensional physiological signal $x_t$, the initial dose $\bar{a}_t$, and the future state $\hat{s}_{t+1}$ predicted by the dynamic human environment transition model. These input data are embedded into the model's input vector, typically processed through positional encoding to preserve temporal information. ;

[0051] Next, the Transformer uses self-attention to process these inputs. Each input vector is mapped to a query vector Q, a key vector K, and a value vector V, and a self-attention score is calculated. The formula for calculating self-attention is: ;

[0052] Here, d_k is the dimension of the key vector. This step allows the model to focus on the relationships between different time steps in the input sequence, thereby learning the complex temporal dependencies between physiological states and drug dosage.

[0053] Then, the Transformer uses a multi-head attention mechanism to compute multiple different attention heads in parallel to capture different temporal patterns, and then concatenates and linearly transforms the results: ;

[0054] Here, each head_i is obtained through the self-attention formula described above, and W^O is the output weight matrix. Through this process, the Transformer can learn multiple different temporal relationships and combine them as the model's output.

[0055] Next, the Transformer performs a nonlinear transformation using a feedforward network to enhance the model's expressive power. The feedforward network typically consists of two linear layers and an activation function (such as ReLU): ;

[0056] Finally, after multiple attention and feedforward layers, the Transformer outputs a new representation h_t, which contains information about the current physiological state, the initial dose, and the predicted future state. The model then transforms h_t into the base dose value \bar{a_t} through a linear layer. ;

[0057] Where W_h is the weight matrix of the linear layer, and b_h is the bias term.

[0058] 2. Dynamic Human-Environment Transfer Model:

[0059] Next, the system introduces the dynamic human environment transition model P trained in step 2 to predict the patient's new state \hat{s}_{t+1} after receiving the initial dose \bar{a}_t: ;

[0060] This model simulates the dynamic response of the human body to anesthesia dose, capturing individual differences and real-time physiological changes, and providing a scientific basis for subsequent dose adjustments.

[0061] 3. Prospective dose adjustment:

[0062] Finally, based on the initial dose \bar{a}_t and the predicted new state \hat{s}_{t+1}, the system calculates the dose adjustment value \delta a_t using the prospective early warning model F, and generates the final dose prediction value \hat{a}_t: ;

[0063] F will reuse the same model architecture as F_{old}, and the processing is similar, so it will not be described in detail here.

[0064] This step employs a proactive early warning mechanism to dynamically adjust dose predictions, ensuring the accuracy and safety of anesthetic dosage. By introducing a dynamic human environment transfer model and a proactive early warning mechanism, the accuracy and adaptability of anesthetic dosage prediction are significantly improved, effectively addressing individual differences and real-time physiological changes, reducing anesthetic risks, and enhancing surgical safety. Furthermore, the modular design of this invention facilitates integration into existing anesthesia systems, offering broad application prospects.

[0065] Step 4: Training based on reinforcement learning paradigm

[0066] Based on steps 2 and 3, we define the state space and agent policy, respectively. The state space of the reinforcement learning model includes the current physiological data X_t, the predicted data y_t + Δt, and historical drug infusion records. The agent policy, benefiting from prospective information, forms a more granular dose prediction model.

[0067] Based on this information, the system evaluates the effectiveness of the drug infusion strategy and uses a reward function to drive the system to learn a better control strategy. The reward function is designed as follows: ;

[0068] Where ΔA is the change in drug infusion rate, D{target} is the desired depth of anesthesia, D{predicted} is the currently predicted depth of anesthesia, and \alpha and \beta are adjustment coefficients.

[0069] To address complex, high-dimensional control problems, the system employs a Deep Q-Network (DQN) as a reinforcement learning framework. Q-learning achieves the optimal policy by iteratively updating the Q-value function, with the goal of maximizing future cumulative rewards. ;

[0070] Here, γ is the discount factor. By combining real-time monitoring data with prospective prediction results, the system dynamically adjusts the drug infusion rate to achieve precise control of the depth of anesthesia.

[0071] Step 5: Personalized Anesthesia Control

[0072] Traditional reinforcement learning models typically have fixed strategies after training, but the patient's physiological state during anesthesia can change significantly due to individual differences and dynamic changes during the surgical procedure. This non-static nature places higher demands on the adaptability of the control model. This invention introduces online learning to update the drug infusion model in real time, enabling it to gradually adapt to the patient's personalized needs. This invention achieves the following functions by updating the time-series prediction model in real time through an online learning module:

[0073] 1. Real-time updates: The online learning module dynamically adjusts the parameters of the time-series prediction model based on the latest collected patient data, ensuring that it maintains high-precision prediction capabilities under different patients and surgical scenarios.

[0074] 2. Regularization Constraints: To avoid overfitting or instability during frequent model updates, the online learning module introduces the following regularization terms:

[0075] - Parameter smoothing regularization: Constrains the update magnitude to prevent drastic changes in model parameters.

[0076] - Time consistency regularization: ensures a smooth transition of prediction results at consecutive time points.

[0077] - Structural sparsity regularization: Improves the generalization ability of the model by sparsifying the parameters.

[0078] The online learning module employs a gradient optimization method to dynamically adjust model parameters based on real-time data. Its optimization objective is to minimize the following loss function: ;

[0079] The first term is the prediction error loss; the second term is parameter smoothing regularization, which controls the magnitude of parameter updates; the third term is time consistency regularization, which constrains the smoothness of prediction results; and the fourth term is sparsity regularization, which improves the model's generalization ability. \lambda_1\lambda_2\lambda_3 are regularization weight coefficients, dynamically adjusted to adapt to different patients and surgical scenarios.

[0080] The specific formula for online updates is as follows: ;

[0081] The alert information greatly improves the safety of anesthesia and the quality of postoperative recovery for patients.

[0082] To further illustrate the specific implementation effects of this invention, the following detailed description of the perioperative automated anesthesia system based on prospective early warning and reinforcement learning, using three research and development test examples, is provided. Each example provides specific experimental conditions, parameter settings, and system output results to demonstrate the application effects of this invention in different patient groups.

[0083] Example 1: Ordinary adult patients

[0084] Experimental subject: Male, 35 years old, weight 70 kg, height 175 cm, ASA I.

[0085] Experimental conditions and parameters: Anesthetic drugs: propofol and remifentanil; Initial dose prediction Propofol 50 mg / h, remifentanil 0.1 μg / kg / min; prospective prediction time step Minutes; State space input X_t includes ECG, BIS, The signal has 12 channels including MAP; the reinforcement learning strategy is DQN, the discount factor γ = 0.95, and the learning rate η = 0.001.

[0086] Experimental Procedure: The system collects and preprocesses patient physiological data in real time; a prospective prediction model predicts the trends of BIS and MAP changes within the next minute; the initial dose agent F_old outputs based on the current X_t. Dynamic human environment transfer model P simulates physiological response after drug administration. Prospective dose adjustment F outputs the final dose. And execute the infusion.

[0087] Experimental results: The anesthesia depth BIS was maintained for 92% of the target range (40–60); MAP was maintained at 75–85 mmHg without significant overshoot or undershoot; drug dosage was optimized: the total amount of propofol was reduced by about 12% compared to traditional PID control, and remifentanil was reduced by about 10%; the system predicted intraoperative hypotension events 30 seconds in advance and automatically adjusted the dosage to avoid excessively low blood pressure.

[0088] Example 2: Elderly patients

[0089] Subject: Female, 72 years old, weighing 62 kg, height 160 cm, ASA II, with mild renal insufficiency.

[0090] Experimental conditions and parameters: Anesthetic drugs: propofol 40 mg / h, remifentanil 0.08 μg / kg / min (initial dose); prospective prediction time step. Minutes; the state space input X_t includes ECG, BIS, SpO2, MAP, respiratory rate, and historical drug infusion records; the reinforcement learning strategy is SAC, the discount factor γ = 0.99, and the learning rate η = 0.0005.

[0091] Experimental procedure: The system collects and standardizes physiological signals in real time; a prospective prediction model predicts the depth of anesthesia and changes in blood pressure within the next 2 minutes; an initial dose agent is generated. The human body environment transfer model P is generated. Prospective dose adjustment F outputs the final dose. Infusion is performed in real time; the online learning module fine-tunes model parameters based on real-time physiological responses to adapt to the drug metabolism characteristics of elderly patients.

[0092] Experimental results: The anesthesia depth BIS was maintained within the target range for 88% of the time, which is about 15% higher than that of traditional PID control; MAP was maintained at 70–80 mmHg, and a 45-second early warning successfully prevented a sudden drop in blood pressure; the total amount of drugs was reduced by 18% (propofol) and 15% (remifentanil) compared with the traditional method; after the system learned online, the prediction error decreased from ±6 to ±3 BIS units.

[0093] Example 3: Obese patients

[0094] Experimental subject: Male, 45 years old, weight 110 kg, height 178 cm, BMI = 34, ASA II, with a history of hypertension.

[0095] Experimental conditions and parameters:

[0096] Anesthetic drugs: propofol 60 mg / h, remifentanil 0.12 μg / kg / min; prospective prediction time step Minutes; State space input X_t includes ECG, BIS, MAP, PPG signals and historical drug infusion data; the reinforcement learning strategy was DDPG, with a discount factor γ = 0.98 and a learning rate η = 0.0007.

[0097] Experimental procedure: Real-time acquisition and standardization of multimodal physiological data; prospective prediction model to predict anesthesia depth and blood pressure trends within 1.5 minutes; initial dose prediction. Outputted from F_old, then generated by the environment transfer model. Prospective dose adjustment F outputs the final dose. It also performs infusion; the online learning module updates model parameters in real time to address the nonlinear drug metabolism characteristics of obese patients.

[0098] Experimental results: The depth of anesthesia (BIS) was maintained at 90% within the target range, an improvement of approximately 12% compared to simple PID control; the mean arterial pressure (MAP) was maintained at 80–90 mmHg, and the system predicted blood pressure fluctuations 30–50 seconds in advance; drug dosage optimization: propofol was reduced by approximately 14%, and remifentanil by approximately 11%; the system can predict hypertensive or hypotensive events in obese patients during surgery and dynamically adjust drug dosages to achieve precise anesthesia.

[0099] In summary, the three examples above demonstrate that the perioperative automated anesthesia system of this invention combines prospective prediction with intelligent drug regulation to achieve precise control of anesthesia depth. Simultaneously, it utilizes reinforcement learning strategies to optimize dosage decisions across different patient groups, reducing drug usage. Furthermore, the online learning module enables personalized adjustments to adapt to individual differences in elderly, obese, and high-risk patients. The system can also provide early warnings of intraoperative risks, ensuring surgical safety and postoperative recovery quality.

[0100] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the claims.

Claims

1. A perioperative anesthesia method based on prospective early warning and control, characterized in that, Includes the following steps: S1: Steps for data acquisition and preprocessing; S2: Proceed with the steps of introducing a forward-looking early warning model; S3: The steps to build an intelligent agent based on a forward-looking early warning model; S4: Steps for training based on the reinforcement learning paradigm; S5: Steps for individualized anesthesia control.

2. The perioperative anesthesia method based on prospective early warning and control according to claim 1, characterized in that, Step S1 specifically includes: The device employs high-precision physiological signal acquisition, photoplethysmography (PPG) sensors, and gas flow sensors to monitor patients in real time. It provides data with millisecond-level resolution, including but not limited to basic physiological information, physiological parameters during surgery, blood oxygen saturation fluctuation range, anesthesia depth index, and the patient's historical anesthesia records. All acquired data undergoes data preprocessing and standardization.

3. The perioperative anesthesia method based on prospective early warning and control according to claim 1, characterized in that, Step S2 specifically includes: The predictive scope of the aforementioned forward-looking early warning model covers the dynamic changes in anesthesia depth, the adjustment trend of drug demand, and the fluctuations of key physiological indicators. The aforementioned forward-looking early warning model is trained using a deep learning-based time-series model. The input data X_t includes multidimensional physiological signals at the current moment, and the output predicted value is the distribution P(Y_{t+\Delta t}|X_t) of the future state. The model generates real-time trends in the patient's physiological state and outputs early warning signals to guide subsequent interventions. The model optimizes its predictions by minimizing the following loss function: ； Where Y{t+\Delta t}^{(i)} is the true value, \hat{Y}{t+\Delta t}^{(i)} is the predicted value, \theta is the model parameter, \lambda is the regularization coefficient, and N is the number of samples; The state space S_t includes the following key components: Current physiological data X_t: Current patient physiological signals provide real-time anesthesia depth, patient response, and other physiological indicators. This information reflects the immediate status of anesthesia and is the basis for the system's real-time decision-making. Prospective prediction data \hat{Y}_{t+\Delta t}: Through the prospective prediction model, the system can predict the trend of changes in the patient's anesthesia depth, fluctuations in drug demand, and possible changes in physiological state within a certain time step in the future. Prospective data enables the system to predict the anesthesia depth and drug infusion needs in advance, thereby achieving proactive control. Historical drug infusion records D_{history}: Combining these three pieces of information, the state space S_t of this invention is: ； The system incorporates the quantity, rate, and changes in historical drug infusions into the state space to capture the cumulative effect of drugs, changes in drug metabolism rate, and their relationship with the patient's physiological response.

4. The perioperative anesthesia method based on prospective early warning and control according to claim 1, characterized in that... Step 3 specifically includes: S31: Traditional Dose Prediction: The system performs initial dose prediction based on the historical state space x_t using a traditional intelligent agent, generating an initial dose value \bar{a}_t: ； To maintain the stability of traditional methods and ensure a reliable predictive foundation for the system in the initial stage, F will be modeled using a Transformer model. The input data is embedded into the model's input vector and processed through positional encoding to preserve temporal information. ； The Transformer uses a self-attention mechanism to process these inputs. Each input vector is mapped to a query vector Q, a key vector K, and a value vector V. A self-attention score is calculated using the following formula: ； Where d_k is the dimension of the key vector, it allows the model to focus on the relationship between each time step in the input sequence, thereby learning the complex temporal dependency between physiological state and drug dosage; Transformer uses a multi-head self-attention mechanism to compute multiple different attention heads in parallel to capture different temporal patterns, and then concatenates and linearly transforms the results: ； Here, each head_i is obtained through the above self-attention formula, W^O is the output weight matrix, the Transformer learns multiple different temporal relationships and combines them as the output of the model; The Transformer performs nonlinear transformations through a feedforward neural network, enhancing the model's expressive power. A feedforward network typically consists of two linear layers and an activation function. ； After multiple attention and feedforward layers, the Transformer outputs a new representation h_t, which contains information about the current physiological state, the initial dose, and the predicted future state. A linear layer then transforms h_t into the baseline dose value \bar{a_t}. ； Where W_h is the weight matrix of the linear layer, and b_h is the bias term; S32: Dynamic Human Environment Transfer Model. The system incorporates the prospective early warning model trained in step S2 to predict the patient's new state \hat{s}_{t+1} after receiving the initial dose \bar{a}_t. ； By simulating the dynamic response of the human body to anesthetic doses, individual differences and real-time physiological changes are captured, providing a scientific basis for subsequent dose adjustments; S33: Proactive dose adjustment. Based on the initial dose \bar{a}_t and the predicted new state \hat{s}_{t+1}, the system calculates the dose adjustment value \delta a_t using the proactive early warning model F, and generates the final dose prediction value \hat{a}_t. ； In this case, F will reuse the same model architecture as F_{old}.

5. The perioperative anesthesia method based on prospective early warning and control according to claim 1, characterized in that, Step S4 specifically includes: The state space and agent policy are defined separately. The state space of the reinforcement learning model includes the current physiological data X_t, the predicted data \hat{Y}_{t+\Delta t}, and the historical drug infusion records. The agent policy benefits from prospective information, forming a more granular dose prediction model. Based on this information, the system evaluates the effectiveness of the drug infusion strategy and uses a reward function to drive the system to learn a better control strategy. The reward function is: ； Where ΔA is the change in drug infusion rate, D{target} is the desired depth of anesthesia, D{predicted} is the currently predicted depth of anesthesia, and \alpha and \beta are adjustment coefficients; To address complex high-dimensional control problems, deep Q-networks are used as a reinforcement learning framework. Q-learning achieves the optimal policy by iteratively updating the Q-value function, with the goal of maximizing future cumulative rewards. ； Here, γ is a discount factor. By combining real-time monitoring data with prospective prediction results, the system dynamically adjusts the drug infusion rate to achieve precise control of the depth of anesthesia.

6. The perioperative anesthesia method based on prospective early warning and control according to claim 1, characterized in that, Step S5 specifically includes: The online learning module dynamically adjusts the parameters of the time-series prediction model based on the latest collected patient data, making it maintain high-precision prediction capabilities under different patients and surgical scenarios. To prevent overfitting or instability during frequent model updates, the online learning module introduces the following regularization terms: Parameter smoothing regularization: Constrains the update magnitude to prevent drastic changes in model parameters; Temporal consistency regularization: ensures a smooth transition of prediction results across consecutive time points; Sparse regularization: Improves the generalization ability of the model by sparsifying the parameters; The online learning module employs a gradient optimization method to dynamically adjust model parameters based on real-time data. Its optimization objective is to minimize the following loss function: ； The first term is prediction error loss; the second term is parameter smoothing regularization, which controls the magnitude of parameter updates; the third term is time consistency regularization, which constrains the smoothness of prediction results; the fourth term is sparsity regularization, which improves the generalization ability of the model; and \lambda_1\lambda_2\lambda_3 are regularization weight coefficients, which are dynamically adjusted to adapt to different patients and surgical scenarios. The specific formula for online updates is as follows: 。 7. A perioperative anesthesia system based on prospective early warning and control, characterized in that, Use any of the perioperative anesthesia methods based on prospective early warning and control as described in claims 1 to 6.

Images