An Adaptive Intelligent Control Method for Ventilator Parameters

Abstract

This application discloses an adaptive intelligent control method for ventilator parameters, comprising: constructing a disease perception model and inputting a sequence of state basis vectors into it, outputting a probability vector representing the probability of the patient being in the acute, stable, and recovery phases; determining a dynamic weight vector for state modulation through a weighted generative neural network and combining it with the state basis vectors to determine a dynamic state vector; designing a hierarchical reward function, including basic safety sub-item rewards, iatrogenic risk sub-item rewards, and long-term prognosis sub-item rewards, determining a dynamic weight vector for reward synthesis based on the disease phase probability vector, and weighted summing it with the reward values ​​of each sub-item to obtain a total reward value; based on the dynamic state vector, determining a ventilator parameter control scheme through a reinforcement learning agent, performing safety verification, and if the verification fails, selecting a safety fallback scheme as the control scheme for parameter control, and feeding the verification results back to the agent for strategy optimization. This application can improve the accuracy and safety of parameter control.

Description

Technical Field

[0001] This application relates to the field of intelligent control technology for medical devices, and relates to, but is not limited to, an adaptive intelligent adjustment method for ventilator parameters. Background Technology

[0002] Mechanical ventilation is a critical life support measure for patients with respiratory failure in the intensive care unit. Setting ventilation parameters such as FiO2, positive end-expiratory pressure (PEEP), and tidal volume requires striking a balance between improving oxygenation, implementing lung-protective ventilation strategies, and maintaining hemodynamic stability. This involves high clinical decision-making complexity and significant individual variability. Current clinical practice relies heavily on physician experience and static treatment guidelines, making it difficult to respond in real-time to the dynamic changes in a patient's condition, and easily leading to complications such as inadequate ventilation and ventilator-associated lung injury.

[0003] In existing technologies, ventilator parameters are controlled through closed-loop ventilation systems or intelligent parameter tuning algorithms. For example, closed-loop controllers based on fixed rules can automatically adjust parameters according to preset goals. However, this method lacks the ability to perceive the overall stage of the patient's condition and cannot dynamically adjust treatment goals according to the disease progression. Furthermore, some studies use machine learning models to predict disease progression or weaning timing, but this method is usually based on single-modal data and does not form a decision-making closed loop with the treatment action generation process. In recent years, reinforcement learning has been applied to this field, but this method cannot dynamically perceive the patient's condition and struggles to balance multiple goals such as immediate safety, iatrogenic risk, and long-term prognosis. It typically relies on simple post-event alarms and lacks a mechanism for proactively intercepting hidden risks that may arise from combined parameters, resulting in insufficient safety assurance. Therefore, a more comprehensive ventilator parameter control method is urgently needed, capable of integrating multimodal physiological information, perceiving the disease stage in real time, and adaptively adjusting treatment goals and safety strategies at different stages. This would address many of the problems existing in current technologies and ultimately significantly improve the precision, safety, and personalization of treatment. Summary of the Invention

[0004] This application provides an adaptive intelligent control method for ventilator parameters.

[0005] The technical solution of this application embodiment is implemented as follows: In a first aspect, embodiments of this application provide an adaptive intelligent control method for ventilator parameters. The method includes: collecting a patient's respiratory mechanics data, hemodynamic data, and gas exchange data, performing standardized preprocessing, and constructing a state basis vector; constructing a disease perception model based on a deep learning network, inputting the state basis vector sequence into the trained disease perception model, and outputting a disease stage probability vector representing the probability that the patient is in the acute, stable, and recovery phases; determining a state modulation dynamic weight vector based on the disease stage probability vector using a weighted generative neural network, and multiplying the state basis vector and the state modulation dynamic weight vector element-wise to obtain a dynamic state vector; designing a hierarchical reward function, which includes a basic safety sub-item reward, a medical risk sub-item reward, and a long-term prognosis sub-item reward, based on... The reward synthesis dynamic weight vector is determined based on the disease stage probability vector. The reward value of each sub-item is calculated in real time and weighted and summed according to the reward synthesis dynamic weight vector to obtain the total reward value. Based on the dynamic state vector, a reinforcement learning agent based on a dual-delay deep deterministic policy gradient algorithm makes decisions to obtain a ventilator parameter control scheme. Based on the ventilator parameter control scheme, the state basis vector, and the disease stage probability vector, safety verification based on clinical rules and probabilistic risk assessment based on a risk prediction model are performed sequentially. If the verification is successful, the ventilator parameter control scheme is adopted as the final control scheme. If it fails, a safety fallback scheme is determined based on the patient's clinical characteristics. Ventilator parameter control is executed according to the final control scheme, and the verification results are fed back to the reinforcement learning agent for strategy optimization.

[0006] The technical solution provided in this application collects patients' respiratory mechanics data, hemodynamic data, and gas exchange data, performs standardized preprocessing, and constructs state basis vectors to provide a reliable and high-quality data foundation for determining subsequent parameter control schemes. A disease perception model is constructed based on a deep learning network. The state basis vector sequence is input into the trained disease perception model, which outputs a disease stage probability vector representing the probability that the patient is in the acute, stable, or recovery phase. This achieves real-time quantitative perception of the patient's dynamic clinical stage and provides crucial stage-specific contextual information. Based on the disease stage probability vector, a state modulation dynamic weight vector is determined through a weighted generative neural network. The state basis vector and the state modulation dynamic weight vector are multiplied element-wise to obtain a dynamic state vector closely related to the current disease stage. This allows the reinforcement learning agent to focus on key physiological signals, improving the decision-making relevance of state information. A hierarchical reward function is designed, including basic safety sub-item rewards, iatrogenic risk sub-item rewards, and long-term prognosis sub-item rewards, to be dynamically coupled with the disease stage. The reward synthesis dynamic weight vector is determined based on the disease stage probability vector, and the reward values ​​for each sub-item are calculated in real time and dynamically adjusted according to the reward synthesis. The weighted summation of the weight vectors yields the total reward value, enabling the agent's optimization objective to adaptively change with clinical priorities, fundamentally resolving the problems of multi-objective conflict and stage objective mismatch. Based on the dynamic state vector, a reinforcement learning agent using a dual-delay deep deterministic policy gradient algorithm makes decisions, resulting in a precise ventilator parameter control scheme. This achieves a shift from static clinical rules or fixed patterns to dynamic, personalized intelligent decision-making. Based on the ventilator parameter control scheme, state basis vector, and disease stage probability vector, safety verification based on clinical rules and probabilistic risk assessment based on a risk prediction model are performed sequentially. If the verification passes, the ventilator parameter control scheme is adopted as the final control scheme; otherwise, a safety fallback scheme is determined based on the patient's clinical characteristics. Ventilator parameter control is executed according to the final control scheme, and the verification results are fed back to the reinforcement learning agent for policy optimization. Through dual safety verification, high-risk operations are proactively intercepted before action execution, and an intelligent fallback scheme is activated to ensure the safety of the scheme execution. The verification results are fed back to optimize the agent, thus constructing a safe closed loop of decision-making-verification-learning, achieving a unity of safety and personalized parameter tuning capabilities.

[0007] Optionally, the disease perception model includes a bidirectional long short-term memory network layer, a multi-head self-attention layer, a feature aggregation and mapping layer, and a probability output layer connected in sequence. The step of inputting the state basis vector sequence into the trained disease perception model and outputting a disease stage probability vector representing the probability that the patient is in the acute, stable, or recovery phase includes: receiving the state basis vector sequence through the bidirectional long short-term memory network layer and performing context encoding on the state basis vector sequence to output a feature matrix; receiving the feature matrix through the multi-head self-attention layer and mapping the feature matrix into a query matrix, a key matrix, and a value matrix; uniformly dividing each matrix along the feature dimension into multiple attention heads and calculating scaled dot product annotations. The system obtains the outputs of each attention head and concatenates them to obtain a merged matrix. This merged matrix is ​​then linearly projected to obtain an attention-enhanced feature matrix. The attention-enhanced feature matrix is ​​received through a global feature aggregation and mapping layer, and global average pooling is performed on it to obtain a global feature vector. This global feature vector is then mapped to a disease stage score vector through a fully connected layer. The disease stage score vector includes the original scores for the acute, stable, and recovery phases. Finally, the disease stage score vector is received through a probability output layer, and normalized using a Softmax function with a temperature parameter to obtain a disease stage probability vector.

[0008] Optionally, the disease perception model is trained through the following process: constructing a training sample set, which includes historical state basis vector sequences and actual disease stage labels, and dividing the training sample set into a training set and a validation set; training the disease perception model using the training set, outputting the predicted disease stage probability, and calculating the cross-entropy loss value between the predicted disease stage probability and the actual disease stage label; optimizing the parameters of the disease perception model using the cross-entropy loss value as the optimization objective through a backpropagation algorithm, and monitoring the performance on the validation set; stopping training when the validation set loss no longer continues to decrease, thus obtaining a preliminary trained model; fixing the parameters of the preliminary trained model, and optimizing the temperature parameters in the probability output layer using the validation set until the calibration error of the predicted probability is minimized, thus obtaining the optimal temperature parameters; and embedding the optimal temperature parameters into the probability output layer, which, together with the parameters of the preliminary trained model, constitute the trained disease perception model.

[0009] Optionally, the step of determining a state-modulated dynamic weight vector based on the disease stage probability vector using a weighted generative neural network, and multiplying the state basis vector and the state-modulated dynamic weight vector element-wise to obtain a dynamic state vector includes: determining an original weight vector based on the disease stage probability vector using a weighted generative neural network; normalizing the original weight vector using a Softmax function with a temperature parameter to obtain a state-modulated dynamic weight vector, wherein each dynamic weight is within a preset numerical range and the sum of all dynamic weights is 1; and multiplying the state basis vector and the state-modulated dynamic weight vector element-wise to obtain a dynamic state vector, wherein the dynamic state vector has the same dimension as the state basis vector.

[0010] Optionally, the step of determining the reward synthesis dynamic weight vector based on the disease stage probability vector, calculating the reward value of each sub-item in real time, and performing a weighted summation based on the reward synthesis dynamic weight vector to obtain the total reward value includes: determining the reward synthesis dynamic weight vector based on the disease stage probability vector by performing a weighted summation on preset acute phase, stable phase, and recovery phase baseline weight vectors; determining the basic safety sub-item reward value based on the patient's immediate vital sign data, wherein the basic safety sub-item reward value is a weighted sum of oxygenation safety reward and ventilation safety reward; and determining the iatrogenic risk sub-item based on the patient's ventilator parameters and physiological data. The reward value includes rewards for barotrauma risk, hypoventilation risk, and patient-ventilator asynchrony risk. Based on the patient's physiological trends across multiple time scales, a long-term prognosis reward value is determined, calculated using the results of oxygenation trend assessment, perfusion and organ function trend assessment, and respiratory effort and weaning preparation assessment. The basic safety reward value, iatrogenic risk reward value, and long-term prognosis reward value are each normalized. The normalized reward values ​​are then weighted and summed with the dynamic weights corresponding to the reward composite dynamic weight vector to obtain the total reward value.

[0011] Optionally, the step of obtaining a ventilator parameter control scheme by making decisions based on the dynamic state vector and using a reinforcement learning agent trained on a dual-delay deep deterministic policy gradient algorithm includes: constructing a state sequence based on the dynamic state vector; processing the state sequence through a policy network of a reinforcement learning agent trained on a dual-delay deep deterministic policy gradient algorithm to determine a normalized parameter adjustment action vector; determining the dynamic safety adjustment range of each ventilator parameter by boundary prediction based on the patient's current physiological state; and mapping the normalized parameter adjustment action vector to the dynamic safety adjustment range to obtain the ventilator parameter control scheme.

[0012] Optionally, the reinforcement learning agent is trained through a process based on a dual-delay deep deterministic policy gradient algorithm: an initial policy is obtained by training a policy network using a behavior cloning algorithm based on expert demonstration data; the initial policy is then trained offline using a conservative Q-learning framework based on historical patient interaction data. During this stage, an experience tuple is constructed based on the historical patient interaction data, and the current state, action, reward, and next state are sampled; the next action is generated using a target policy network based on the next state; value estimates are calculated using a dual Q-value estimation network based on the next state and the next action, and the minimum value is taken; a temporal difference target value is calculated by combining the reward and the minimum value estimate; a loss function is constructed based on the difference between the temporal difference target value and the predicted value of the current Q-value estimation network; the loss is calculated, and the parameters of the current Q-value estimation network are updated. The parameters of the policy network are updated based on the output of the current Q-value estimation network, wherein the update frequency of the policy network is lower than that of the double Q-value estimation network. Using a soft update mechanism, the parameters of the policy network and the double Q-value estimation network are synchronized to the corresponding target network to obtain an optimized policy. Based on real-time interactive data, the optimized policy is fine-tuned online in shadow mode. During this stage, empirical tuples are constructed based on real-time interactive data, where each empirical tuple includes the current state, the parameter adjustment action vector of the executed policy output, the actual executed action, the next state, the reward, and the safety verification flag. In the background, the policy gradient is corrected using an importance sampling method based on the empirical tuples, and the parameters of the policy network and the double Q-value estimation network are updated. At fixed intervals, offline performance and safety assessments are performed on the background updated policy. When the background updated policy is superior to the current policy and meets the safety threshold, a policy update proposal is generated. After clinical review, the policy update proposal is deployed as the execution policy.

[0013] Secondly, embodiments of this application provide an electronic device, including a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the program, it implements the steps in the aforementioned adaptive intelligent control method for ventilator parameters.

[0014] Thirdly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in the aforementioned adaptive intelligent control method for ventilator parameters.

[0015] The beneficial effects of the technical solutions provided in this application include at least the following: An adaptive intelligent control method for ventilator parameters is provided, which collects the patient's respiratory mechanics data, hemodynamic data, and gas exchange data and performs standardized preprocessing to construct a state basis vector, providing a reliable and high-quality data foundation for determining subsequent parameter control schemes; a disease perception model is constructed based on a deep learning network, and the state basis vector sequence is input into the trained disease perception model, outputting a disease stage probability vector representing the probability that the patient is in the acute, stable, and recovery phases, realizing real-time quantitative perception of the patient's dynamic clinical stage and providing key stage-specific contextual information; based on the disease stage probability vector, a state modulation dynamic weight vector is determined through a weighted generative neural network, and the state basis vector and the state modulation dynamic weight vector are multiplied element-wise to obtain a dynamic state vector closely related to the current disease stage, enabling the reinforcement learning agent to focus on key physiological signals and improving the decision-making relevance of state information; a hierarchical reward function is designed, including basic safety sub-item rewards, iatrogenic risk sub-item rewards, and long-term prognosis sub-item rewards, to be dynamically coupled with the disease stage, and the reward synthesis dynamic weight vector is determined based on the disease stage probability vector, realizing… The reward value for each sub-item is calculated and weighted summed based on the dynamic weight vector synthesized from the rewards to obtain the total reward value. This allows the agent's optimization objective to adaptively change with clinical priorities, fundamentally solving the problems of multi-objective conflict and stage objective mismatch. Based on the dynamic state vector, a reinforcement learning agent using a dual-delay deep deterministic policy gradient algorithm makes decisions to obtain a precise ventilator parameter control scheme, thereby realizing the transformation from static clinical rules or fixed patterns to dynamic and personalized intelligent decision-making. Based on the ventilator parameter control scheme, state basis vector, and disease stage probability vector, clinical rule-based... Safety verification and probabilistic risk assessment based on a risk prediction model are performed. If the verification passes, the ventilator parameter adjustment plan is adopted as the final adjustment plan. If it fails, a safety fallback plan is determined based on the patient's clinical characteristics. The ventilator parameter adjustment is executed according to the final adjustment plan, and the verification results are fed back to the reinforcement learning agent for strategy optimization. Through dual safety verification, high-risk operations are proactively intercepted before the action is executed, and the intelligent fallback plan is activated to ensure the safety of the plan execution. The verification results are fed back to optimize the agent, thereby constructing a safety closed loop of decision-making-verification-learning, achieving a unity of safety and individualized parameter tuning capabilities. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 A flowchart illustrating an adaptive intelligent control method for ventilator parameters provided in this application embodiment; Figure 2 This is a schematic diagram of the hardware entity of an electronic device provided in an embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In the following description, references are made to “some embodiments,” which describe a subset of all possible embodiments. However, it is understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0019] It should be noted that the terms "first, second, and third" used in the embodiments of this application are merely to distinguish similar objects and do not represent a specific ordering of objects. It is understood that "first, second, and third" can be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0020] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of this application pertain. It should also be understood that terms such as those defined in general dictionaries should be understood to have a meaning consistent with their meaning in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0021] The embodiments of this application will be further described below with reference to the accompanying drawings.

[0022] In view of the current problems in the research on ventilator parameter regulation in the field of intelligent control technology of medical equipment, this application provides an adaptive intelligent regulation method for ventilator parameters.

[0023] The technical solution of this application is described below, starting with the method embodiments.

[0024] Please refer to Figure 1 It illustrates a flowchart of an adaptive intelligent control method for ventilator parameters provided in an embodiment of this application, such as... Figure 1 As shown, the method includes at least the following steps S110 to S160.

[0025] Step S110: Collect the patient's respiratory mechanics data, hemodynamic data, and gas exchange data, perform standardized preprocessing, and construct the state basis vector.

[0026] In this embodiment, three types of physiological data—respiratory mechanics data, hemodynamic data, and gas exchange data—are collected from bedside monitors, ventilators, blood gas analyzers, and laboratory information systems in the intensive care unit. Further, the collected physiological data undergoes standardized preprocessing to construct a state basis vector. Specifically, the respiratory mechanics, hemodynamics, and gas exchange data are time-aligned based on the central monitoring time. For continuous data, a fixed-duration time window is used for sliding, and statistical characteristics within the window are calculated. For discrete data, each measurement is assigned a time decay weight, causing its influence to decrease over time. When the timestamp deviation between different data exceeds a preset threshold (e.g., 30 seconds), soft alignment is performed using a dynamic time warping algorithm to ensure all physiological data are aligned on the same time axis. For time-aligned physiological data, outliers are first removed using the Grubbs test and physiological range elimination. For missing values, interpolation methods and autoregressive models are used to impute missing values ​​based on their duration and type. For the physiological data after handling for anomalies and missing data, time-domain features, frequency-domain features, and variability features are extracted from the physiological data within each 5-minute time window. A recursive feature elimination method is then used to select the 20 features with the highest discriminative power from all extracted features. Finally, the features are standardized using the Z-score method based on the patient's individual baseline to construct a state basis vector. This state basis vector has 20 dimensions, and one vector is output every 5 minutes. This vector contains the standardized instantaneous value of the selected physiological feature, its trend within the recent time window, and its anomaly score relative to the patient's historical baseline, thus providing a reliable data foundation for subsequent processing.

[0027] Step S120: Construct a disease perception model based on a deep learning network, input the state basis vector sequence into the trained disease perception model, and output a disease stage probability vector representing the probability that the patient is in the acute phase, stable phase, and recovery phase.

[0028] In this embodiment, the disease perception model includes a bidirectional long short-term memory (LSTM) network layer, a multi-head self-attention layer, a feature aggregation and mapping layer, and a probability output layer connected in sequence. The state basis vector sequence is input into the trained disease perception model, and the model outputs a disease stage probability vector. Specifically, the bidirectional LTM network layer receives the state basis vector sequence and performs context encoding on it. This bidirectional LTM network layer includes a forward LTM network and a backward LTM network, which process the state basis vector sequence along the forward and backward time axes, respectively, to capture the forward and backward dependencies of physiological indicators. The hidden layer dimension of each LTM network is set to 128. By concatenating the forward and backward hidden states at each time step, the context-encoded feature matrix is ​​obtained. Furthermore, the feature matrix is ​​received through a multi-head self-attention layer, and mapped into a query matrix, a key matrix, and a value matrix through a learnable linear transformation. The query matrix, key matrix, and value matrix are then uniformly divided into multiple attention heads along the feature dimension, and the scaled dot product attention is calculated for each attention head to obtain the output of each attention head. Finally, the outputs of each attention head are concatenated along the feature dimension to obtain a merged matrix. This merged matrix is ​​then linearly projected through a learnable linear projection matrix to obtain the attention-enhanced feature matrix.

[0029] Furthermore, the attention-enhanced feature matrix is ​​received through a global feature aggregation and mapping layer, and global average pooling is performed on the attention-enhanced feature matrix along the time dimension to obtain a global feature vector. This global feature vector is processed through a two-layer fully connected network. The first fully connected layer (e.g., a 64-dimensional neuron) uses the ReLU activation function, while the second layer (e.g., a 3-dimensional neuron) has no activation function. This global feature vector is mapped into a three-dimensional disease stage score vector, which corresponds to the original scores of the acute phase, stable phase, and recovery phase, respectively. The disease stage score vector is received through a probability output layer, and the disease stage score vector is normalized and calibrated using a Softmax function with a temperature parameter to obtain a disease stage probability vector representing the probability that the patient is in the acute phase, stable phase, and recovery phase. ,in, This represents the score vector indicating the stage of the illness. Let represent the probability that the patient is in each stage of their condition at time t, and satisfy . ,in Let represent the probabilities that the patient is in the acute phase, stable phase, and recovery phase at time t, respectively. The temperature parameter is a hyperparameter that is optimized and fixed on the validation set after model training, with the goal of minimizing the calibration error of the predicted probability. For example, 0.8. This temperature parameter can make the probability values ​​of each disease stage output by the model more reliable in terms of confidence.

[0030] In this embodiment, the training process of the disease perception model is based on a dataset containing a large amount of historical patient data and their corresponding disease stage labels. First, a training sample set is constructed, including historical state basis vector sequences and actual disease stage labels. These historical state basis vector sequences are obtained from historical patient data. Clinical experts independently double-blindly label the historical state basis vectors of each time series according to standards such as the Berlin ARDS definition, and after arbitration, a gold standard is formed to determine the actual disease stage labels. The training sample set is divided into a training set and a validation set. The disease perception model is trained using the training set, outputting the predicted disease stage probability. The cross-entropy loss value between the predicted disease stage probability and the actual disease stage label is calculated. With minimizing the cross-entropy loss value as the optimization objective, the parameters of the bidirectional long short-term memory network layer, multi-head self-attention layer, feature aggregation and mapping layer, and probability output layer in the disease perception model are optimized using the backpropagation algorithm. Performance is monitored on the validation set. Training stops when the validation set loss no longer decreases continuously, resulting in a preliminary trained model. Furthermore, the parameters of the initially trained model are fixed, and the temperature parameters in the probability output layer are optimized using the validation set until the calibration error of the predicted probability is minimized, obtaining the optimal temperature parameters. These optimal temperature parameters are then fixed into the probability output layer, forming the final trained disease perception model together with the parameters of each network layer in the initially trained model. During actual operation, the trained disease perception model runs as a real-time inference module, receiving the latest state basis vector sequence every 5 minutes. This sequence is processed through each network layer to output a real-time disease stage probability vector, providing a core basis for subsequent adaptive scheduling.

[0031] Step S130: Based on the disease stage probability vector, determine the state modulation dynamic weight vector through a weighted generating neural network, and multiply the state basis vector and the state modulation dynamic weight vector element by element to obtain the dynamic state vector.

[0032] In this embodiment, a state-modulated dynamic weight vector is determined by a weight generation neural network, and the dynamic state vector is obtained by element-wise multiplication with the state basis vector. Specifically, the weight generation neural network adopts a lightweight multilayer perceptron architecture. It receives the disease stage probability vector through the input layer and selectively concatenates some statistical features (such as mean and variance) of the state basis vector as contextual information. This weight generation neural network includes at least one hidden layer, processes the input data using the ReLU activation function, and generates the original weight vector through the Sigmoid activation function in the output layer. Furthermore, to ensure that the generated weights have good numerical characteristics and clear physical meaning, the original weight vector is normalized and calibrated. First, the original weight vector is normalized using a Softmax function with a temperature parameter to enhance weight sparsity, resulting in a dynamic weight vector where each dynamic weight is within a preset numerical range, and the sum of all dynamic weights is 1. Furthermore, the state basis vector and the state-modulated dynamic weight vector are multiplied element-wise to obtain the dynamic state vector. The dynamic state vector has the same dimension as the state basis vector, but each physiological feature value is modulated by its corresponding dynamic weight. The larger the weight, the more important the corresponding physiological feature value is to the decision at the current stage of the disease, and its contribution to the reinforcement learning state representation is enhanced accordingly. Conversely, the smaller the weight, the more the contribution is suppressed. Thus, the dynamic state vector becomes a state representation whose content adaptively evolves with the stage of the disease, and can be directly used as the input of the subsequent reinforcement learning agent. This realizes the transformation of the state space from a static list of physiological indicators to a dynamic highlighting of the disease, providing high-quality input for subsequent personalized decision-making.

[0033] Step S140: Design a hierarchical reward function, which includes basic safety sub-item reward, medical risk sub-item reward and long-term prognosis sub-item reward. Determine the reward synthesis dynamic weight vector according to the disease stage probability vector, calculate the reward value of each sub-item in real time and perform weighted summation according to the reward synthesis dynamic weight vector to obtain the total reward value.

[0034] In this embodiment, a dynamic weight vector for reward synthesis is determined based on the probability vector of disease stage, and the total reward value is obtained by weighted summation with the reward values ​​of each sub-item. Specifically, based on the probability vector of disease stage, a baseline weight vector for the acute phase, a baseline weight vector for the stable phase, and a baseline weight vector for the recovery phase are preset. Each baseline weight vector is a three-dimensional non-negative vector, and the sum of each component is 1, for example [0.6, 0.3, 0.1]. The dynamic weight vector for reward synthesis is determined by weighted summation of the preset baseline weight vectors for the acute phase, the stable phase, and the recovery phase. This allows the weight to transition smoothly with the disease stage, emphasizing basic safety in the acute phase and long-term prognosis in the recovery phase, thereby aligning the reward with clinical goals. Furthermore, the reward values ​​for each sub-item are determined. The basic safety sub-item reward value is determined based on the patient's immediate vital signs data to reflect life-threatening immediate risks. Specifically, it is the sum of the oxygenation safety penalty and the ventilation safety penalty, both of which use continuous functions. The iatrogenic risk sub-item reward value is determined based on the patient's ventilator parameters and physiological data to quantify the risk of ventilator-related injuries. Specifically, it is the sum of the barotrauma risk penalty, the hypoventilation risk penalty, and the patient-ventilator asynchrony risk penalty. Among these, the barotrauma risk penalty is determined based on plateau pressure, driving pressure, and FiO2; the hypoventilation risk penalty is determined based on minute ventilation and its rate of change; and the patient-ventilator asynchrony risk penalty is derived from respiratory waveform analysis. The asynchronous index was determined; the long-term prognostic sub-item reward value was determined based on the patient's physiological trends across multiple time scales to assess the potential impact of treatment on the patient's medium- to long-term recovery. This long-term prognostic sub-item reward value was determined based on the results of oxygenation trend assessment, perfusion and organ function trend assessment, and respiratory effort and weaning preparation assessment. Specifically, the oxygenation trend assessment result was determined by calculating the weighted linear regression slope of the oxygenation index over the past hour; the perfusion and organ function trend assessment result was determined based on the trends of lactate clearance and central venous oxygen saturation over the past 6 hours; and the respiratory effort and weaning preparation assessment result was primarily used during the recovery period and was determined based on the shallow and rapid breathing index trend and inspiratory effort index. When the oxygenation trend assessment result was greater than +10... If the oxygenation trend assessment result is greater than +5 mmHg / h and the perfusion assessment result improves, the long-term prognosis sub-item reward value is +2; if the oxygenation trend assessment result is greater than +5 mmHg / h or the weaning preparation assessment result improves, the long-term prognosis sub-item reward value is +1; if the oxygenation trend assessment result is less than -10 mmHg / h or the perfusion assessment result worsens, the long-term prognosis sub-item reward value is -1; otherwise, it is 0.

[0035] Furthermore, after the three sub-item reward values ​​are determined, the numerical range of each sub-item reward value is first normalized, such as by mapping it to the [-1, 1] interval using the tanh function, to obtain the normalized basic safety sub-item reward value, iatrogenic risk sub-item reward value, and long-term prognosis sub-item reward value. The normalized sub-item reward values ​​are then weighted and summed with the dynamic weights corresponding to the reward synthesis dynamic weight vector to obtain the total reward value. This total reward value is calculated every 5 minutes, dynamically evolves over time, and is output in real time to the subsequent reinforcement learning agent. It serves as the core basis for evaluating the quality of actions, calculating policy gradients, and setting optimization goals for its policy network and Q-value estimation network during training, guiding it to learn ventilation strategies that conform to the clinical stage priority. The technical solution provided in this application embodiment, through a dynamic reward mechanism deeply coupled with the disease stage, enables the agent to prioritize life safety in the acute phase, learn to balance multiple objectives in the stable phase, and actively prepare for weaning in the recovery phase, thereby achieving safe, effective, and personalized stage adaptive optimization.

[0036] Step S150: Based on the dynamic state vector, a reinforcement learning agent based on a dual-delay deep deterministic policy gradient algorithm makes a decision to obtain a ventilator parameter control scheme.

[0037] In this embodiment, to improve the smoothness and trend perception of decision-making, a buffer containing the dynamic state vectors of the most recent K time steps is maintained, and a state sequence is constructed based on this buffer as the input of the agent. Based on this state sequence, a reinforcement learning agent trained on the dual-delay deep deterministic policy gradient algorithm makes a decision to obtain a ventilator parameter adjustment scheme. Specifically, the state sequence is processed by the policy network of the reinforcement learning agent trained on the dual-delay deep deterministic policy gradient algorithm. After performing forward computation, a normalized original parameter adjustment action vector a(t)∈[-1, 1]^M is output, where M is the number of ventilator parameters to be adjusted, such as M=4, corresponding to FiO2, PEEP, tidal volume, and respiratory rate, respectively. In addition, to ensure the absolute clinical safety of the output actions, a dynamic action constraint module is set up. Based on the patient's current real-time physiological state, such as lung compliance and oxygenation index, a lightweight boundary prediction network or predefined rules are used to calculate the dynamic safety adjustment range of each ventilator parameter at the current moment. The normalized parameter adjustment action vector output by the strategy network is linearly mapped to this dynamic safety adjustment range, thus obtaining a ventilator parameter control scheme that meets the safety boundary.

[0038] In this embodiment, the reinforcement learning agent used for decision-making is trained using a process based on a dual-delay deep deterministic policy gradient algorithm. Specifically, the training process consists of three steps: First, based on expert demonstration data, a policy network is trained using a behavior cloning algorithm to obtain an initial policy. Second, based on historical patient interaction data, the initial policy is trained offline using a conservative Q-learning framework. In this stage, an experience tuple is constructed based on the historical patient interaction data, and the current state, action, reward, and next state are sampled. Based on the next state, the next action is generated through a target policy network. Based on the next state and the next action, a value estimate is calculated using a dual Q-value estimation network, and the minimum value is taken. The temporal difference target value is calculated by combining the reward and the minimum value estimate. A loss function is constructed based on the difference between the temporal difference target value and the predicted value of the current Q-value estimation network. The loss is calculated, and the parameters of the current Q-value estimation network are updated. The parameters of the policy network are updated based on the output of the current Q-value estimation network. The policy network updates less frequently than the dual Q-value estimation network. Using a soft update mechanism, the parameters of the policy network and the dual Q-value estimation network are synchronized to the corresponding target network, thus obtaining the optimized policy. Finally, based on real-time interactive data, the optimized policy is fine-tuned online in shadow mode. During this stage, empirical tuples are constructed based on the real-time interactive data. Each empirical tuple contains the current state, the parameter adjustment action vector output by the executed policy, the actual executed action, the next state, the reward, and the safety verification flag. In the background, the policy gradient is corrected using an importance sampling method based on the empirical tuples, and the parameters of the policy network and the dual Q-value estimation network are updated. At fixed intervals, offline performance and safety assessments are performed on the background-updated policy. When the background-updated policy is superior to the current policy and meets the safety threshold, a policy update proposal is generated. After clinical review, the policy update proposal is deployed as the execution policy, resulting in a trained reinforcement learning agent. This provides a core guarantee for determining a safe and stable parameter control scheme.

[0039] Step S160: Based on the ventilator parameter control scheme, the state basis vector, and the disease stage probability vector, perform safety verification based on clinical rules and probabilistic risk assessment based on risk prediction model in sequence; if the verification is successful, the ventilator parameter control scheme is adopted as the final control scheme; if it fails, a safety fallback scheme is determined based on the patient's clinical characteristics; ventilator parameter control is performed according to the final control scheme, and the verification results are fed back to the reinforcement learning agent for strategy optimization.

[0040] In this embodiment, to ensure the safety of the ventilator parameter control scheme, independent and rigorous two-level safety verification is performed based on the ventilator parameter control scheme, state basis vector, and disease stage probability vector. Specifically, firstly, absolute safety verification based on clinical rules is performed. The ventilator parameter control scheme is superimposed with the current ventilator operating parameters to generate candidate parameter combinations. Based on these candidate parameter combinations, verification is performed according to predefined hard rules in clinical guidelines and expert consensus, including parameter range checks (e.g., whether the tidal volume is between 4-8 mL / kg ideal body weight), parameter logical consistency checks (e.g., whether there is an unreasonable combination of high FiO2 and low PEEP), and parameter change rate checks (e.g., whether the single adjustment amplitude exceeds the 50% safety threshold). If any safety verification fails, a safety interception is immediately triggered. Furthermore, if absolute safety verification based on clinical rules is passed, a probabilistic risk assessment based on a risk prediction model is performed. Candidate parameter combinations, state basis vectors, and patient basic information are input into the barotrauma risk prediction model, which outputs a probability estimate and uncertainty estimate of barotrauma occurring within the next 6 hours. This model is based on XGBoost ensemble learning, and the interception decision uses an individualized dynamic threshold. First, a basic threshold is determined based on the probability vector of the disease stage, such as 0.20 for the acute phase, 0.15 for the stable phase, and 0.10 for the recovery phase. Then, an intermediate threshold is obtained by adjusting according to patient risk characteristics such as COPD. Finally, the final threshold is determined by combining the uncertainty estimate. If the probability estimate of barotrauma occurring within the next 6 hours is greater than or equal to the final threshold, it is judged as high risk and interception is triggered.

[0041] Furthermore, if the safety verification passes, the ventilator parameter adjustment plan will be used as the final adjustment plan. If an interception is triggered, i.e., the safety verification fails, a safety fallback plan will be intelligently selected from the pre-set fallback plan library as an alternative plan based on the patient's clinical characteristics. The specific selection logic is based on a decision tree. For ARDS patients, the corresponding parameters in the ARDSnet low PEEP / FiO2 table are selected. For non-ARDS patients but whose doctors have preset a safety range, the median value of the range is selected. In other cases, the current parameters are maintained or the average of recent safety parameters is used. Similarly, the safety fallback plan also needs to be verified through safety verification. Finally, the validated ventilator parameter control scheme and safety fallback scheme are packaged into a final control scheme and pushed to the doctor's interface for confirmation and execution. After obtaining explicit authorization, the final control scheme is directly sent to the ventilator for parameter control through a standard medical interface (such as HL7). The complete verification results, including the original control scheme, whether interception occurred, the reason for interception, the final control scheme executed, and the risk probability, are recorded and fed back to the experience replay buffer of the reinforcement learning agent for policy optimization. Among them, experience tuples marked as high risk will be given higher priority in the training of the agent and will be mainly used for the safety value learning of the Critic network. This enables the agent to learn from the mistakes and continuously optimize the safety boundary of its policy, forming a safety-learning feedback loop.

[0042] In summary, the adaptive intelligent control method for ventilator parameters provided in this application collects and standardizes the patient's respiratory mechanics, hemodynamics, and gas exchange data to construct a state basis vector, providing a reliable and high-quality data foundation for determining subsequent parameter control schemes. A disease perception model is constructed based on a deep learning network. The state basis vector sequence is input into the trained disease perception model, which outputs a disease stage probability vector representing the probability of the patient being in the acute, stable, or recovery phase. This achieves real-time quantitative perception of the patient's dynamic clinical stage and provides crucial stage-specific contextual information. Based on the disease stage probability vector, a state modulation dynamic weight vector is determined through a weighted generative neural network. The state basis vector and the state modulation dynamic weight vector are multiplied element-wise to obtain a dynamic state vector closely related to the current disease stage. This allows the reinforcement learning agent to focus on key physiological signals, improving the decision-making relevance of state information. A hierarchical reward function is designed, including basic safety sub-item rewards, iatrogenic risk sub-item rewards, and long-term prognosis sub-item rewards, to be dynamically coupled with the disease stage. The reward synthesis dynamic weight vector is determined based on the disease stage probability vector, and the rewards for each sub-item are calculated in real time. The reward value is obtained by weighted summation of the dynamic weight vector synthesized based on the reward, enabling the agent's optimization objective to adapt to changes in clinical priority, fundamentally solving the problems of multi-objective conflict and stage objective mismatch. Based on the dynamic state vector, a reinforcement learning agent based on a dual-delay deep deterministic policy gradient algorithm makes decisions to obtain a precise ventilator parameter control scheme, thereby realizing the transformation from static clinical rules or fixed patterns to dynamic personalized intelligent decision-making. Based on the ventilator parameter control scheme, state basis vector, and disease stage probability vector, safety verification based on clinical rules and probabilistic risk assessment based on risk prediction models are performed sequentially. If the verification is successful, the ventilator parameter control scheme is used as the final control scheme; if it fails, a safety fallback scheme is determined based on the patient's clinical characteristics. Ventilator parameter control is executed according to the final control scheme, and the verification results are fed back to the reinforcement learning agent for policy optimization. Through dual safety verification, high-risk operations are proactively intercepted before action execution, and the intelligent fallback scheme is activated to ensure the safety of the scheme execution. The verification results are fed back to optimize the agent, thereby constructing a safety closed loop of decision-making-verification-learning, achieving a unity of safety and personalized parameter tuning capabilities.

[0043] It should be noted that, in the embodiments of this application, if the above-mentioned adaptive intelligent control method for ventilator parameters is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to the related technology, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, mobile hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of this application are not limited to any specific hardware and software combination.

[0044] Correspondingly, embodiments of this application provide a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the steps of any of the adaptive intelligent control methods for ventilator parameters described in the above embodiments. Correspondingly, embodiments of this application also provide a computer program product, which, when executed by a processor of an electronic device, is used to implement the steps of any of the adaptive intelligent control methods for ventilator parameters described in the above embodiments.

[0045] Based on the same technical concept, this application provides an electronic device for implementing an adaptive intelligent control method for ventilator parameters as described in the above method embodiments. Figure 2 This is a hardware entity diagram of an electronic device provided in an embodiment of this application, such as... Figure 2 As shown, the electronic device 200 includes a memory 210 and a processor 220. The memory 210 stores a computer program that can run on the processor 220. When the processor 220 executes the program, it implements the steps in any of the adaptive intelligent control methods for ventilator parameters described in the embodiments of this application.

[0046] The memory 210 is configured to store instructions and applications executable by the processor 220, and can also cache data to be processed or already processed by the processor 220 and various modules in the electronic device (e.g., image data, audio data, voice communication data and video communication data), which can be implemented by flash memory or random access memory (RAM).

[0047] The processor 220 executes a program to implement any of the steps of an adaptive intelligent control method for ventilator parameters as described above. The processor 220 typically controls the overall operation of the electronic equipment 200.

[0048] The aforementioned processor can be at least one of the following: Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), Central Processing Unit (CPU), Controller, Microcontroller, and Microprocessor. It is understood that other electronic devices can also implement the functions of the aforementioned processor, and this application does not specifically limit the specific implementation.

[0049] The aforementioned computer storage media / memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), magnetic random access memory (FRAM), flash memory, magnetic surface memory, optical disc, or compact disc read-only memory (CD-ROM), etc.; or it can be various electronic devices that include one or any combination of the above-mentioned memories, such as mobile phones, computers, tablet devices, personal digital assistants, etc.

[0050] It should be noted that the descriptions of the storage medium and device embodiments above are similar to those of the method embodiments above, and have similar beneficial effects. For technical details not disclosed in the storage medium and device embodiments of this application, please refer to the descriptions of the method embodiments of this application for understanding.

[0051] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the above-described embodiments are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0052] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0053] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0054] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0055] In addition, each functional unit in the various embodiments of this application can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0056] Alternatively, if the integrated units described above are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this application, or the parts that contribute to related technologies, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause the device automatic test line to execute all or part of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.

[0057] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0058] The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method or device embodiments. The above descriptions are merely embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An adaptive intelligent control method for ventilator parameters, characterized in that, The method includes: Patient respiratory mechanics data, hemodynamic data, and gas exchange data were collected and standardized preprocessed to construct state basis vectors; A disease perception model is constructed based on a deep learning network. The state basis vector sequence is input into the trained disease perception model, and the output is a disease stage probability vector representing the probability that the patient is in the acute phase, stable phase, and recovery phase. Based on the disease stage probability vector, a state modulation dynamic weight vector is determined by a weighted generative neural network. The state basis vector and the state modulation dynamic weight vector are multiplied element by element to obtain the dynamic state vector. Design a tiered reward function, which includes basic safety sub-item reward, medical risk sub-item reward, and long-term prognosis sub-item reward. Determine the reward synthesis dynamic weight vector based on the disease stage probability vector, calculate the reward value of each sub-item in real time, and perform weighted summation based on the reward synthesis dynamic weight vector to obtain the total reward value. Based on the dynamic state vector, a reinforcement learning agent based on a dual-delay deep deterministic policy gradient algorithm makes decisions to obtain a ventilator parameter control scheme. Based on the ventilator parameter control scheme, the state basis vector, and the disease stage probability vector, safety verification based on clinical rules and probabilistic risk assessment based on risk prediction models are performed sequentially. If the verification passes, the ventilator parameter control scheme is adopted as the final control scheme; if it fails, a safety fallback scheme is determined based on the patient's clinical characteristics. Ventilator parameter control is executed according to the final control scheme, and the verification results are fed back to the reinforcement learning agent for strategy optimization.

2. The method according to claim 1, characterized in that, The disease perception model comprises a bidirectional long short-term memory network layer, a multi-head self-attention layer, a feature aggregation and mapping layer, and a probability output layer connected in sequence. The process involves inputting a sequence of state basis vectors into the trained disease perception model and outputting a probability vector representing the probability of the patient being in the acute, stable, or recovery phases of the disease, including: The state basis vector sequence is received through a bidirectional long short-term memory network layer, and the state basis vector sequence is context-encoded to output a feature matrix. The feature matrix is ​​received by a multi-head self-attention layer and mapped into a query matrix, a key matrix and a value matrix. Each matrix is ​​evenly divided into multiple attention heads along the feature dimension and scaled dot product attention is calculated. The outputs of each attention head are obtained and concatenated to obtain a merged matrix. The merged matrix is ​​linearly projected to obtain an attention-enhanced feature matrix. The attention-enhanced feature matrix is ​​received through a global feature aggregation and mapping layer, and global average pooling is performed on the attention-enhanced feature matrix to obtain a global feature vector. The global feature vector is then mapped to a disease stage score vector through a fully connected layer. The disease stage score vector includes the original scores for the acute phase, stable phase, and recovery phase. The disease stage score vector is received by the probability output layer, and the disease stage score vector is normalized by the Softmax function with temperature parameter to obtain the disease stage probability vector.

3. The method according to claim 2, characterized in that, The disease perception model was trained through the following process: Construct a training sample set, which includes a historical state basis vector sequence and real disease stage labels, and divide the training sample set into a training set and a validation set; The disease perception model is trained using the training set, and the predicted disease stage probability is output. The cross-entropy loss value between the predicted disease stage probability and the actual disease stage label is calculated. The parameters of the disease perception model are optimized using the cross-entropy loss value as the optimization objective through the backpropagation algorithm. The performance is monitored on the validation set. Training is stopped when the validation set loss no longer continues to decrease, and a preliminary training model is obtained. The parameters of the initial training model are fixed, and the temperature parameters in the probability output layer are optimized through the validation set until the calibration error of the prediction probability is minimized to obtain the optimal temperature parameters. The optimal temperature parameters are then fixed into the probability output layer, and together with the parameters of the initial training model, they constitute the trained disease perception model.

4. The method according to claim 1, characterized in that, The process of determining a state-modulated dynamic weight vector based on the disease stage probability vector using a weighted generative neural network, and then multiplying the state basis vector and the state-modulated dynamic weight vector element-wise to obtain a dynamic state vector includes: Based on the disease stage probability vector, the original weight vector is determined by a weight generation neural network. The original weight vector is then normalized by a Softmax function with a temperature parameter to obtain a state-modulated dynamic weight vector. Each dynamic weight is within a preset value range, and the sum of all dynamic weights is 1. The dynamic state vector is obtained by multiplying the state basis vector and the state modulation dynamic weight vector element by element, wherein the dynamic state vector has the same dimension as the state basis vector.

5. The method according to claim 1, characterized in that, The process involves determining a dynamic weight vector for reward synthesis based on the probability vector of the disease stage, calculating the reward value of each sub-item in real time, and performing a weighted sum based on the dynamic weight vector to obtain the total reward value, including: Based on the disease stage probability vector, the reward synthesis dynamic weight vector is determined by weighted summation of the preset acute phase, stable phase, and recovery phase baseline weight vectors. Based on the patient's real-time vital signs data, a basic safety sub-item reward value is determined, which is the weighted sum of the oxygenation safety reward and the ventilation safety reward; Based on the patient's ventilator parameters and physiological data, the reward value for each iatrogenic risk sub-item is determined. The reward value for each iatrogenic risk sub-item includes barotrauma risk reward, inadequate ventilation risk reward, and patient-ventilator asynchrony risk reward. Based on the patient's physiological trends at multiple time scales, long-term prognostic sub-item reward values ​​are determined, which are based on the results of oxygenation trend assessment, perfusion and organ function trend assessment, and respiratory effort and weaning preparation assessment. The reward values ​​for the basic safety sub-item, the medical risk sub-item, and the long-term prognosis sub-item are normalized respectively. The normalized reward values ​​of each sub-item are then weighted and summed with the dynamic weights corresponding to the reward synthesis dynamic weight vector to obtain the total reward value.

6. The method according to claim 5, characterized in that, The process of obtaining a ventilator parameter control scheme based on the dynamic state vector, using a reinforcement learning agent based on a dual-delay deep deterministic policy gradient algorithm, includes: A state sequence is constructed based on the dynamic state vector. The state sequence is processed by a policy network of a reinforcement learning agent trained on a dual-delay deep deterministic policy gradient algorithm to determine a normalized parameter adjustment action vector. Based on the patient's current physiological state, the dynamic safety adjustment range of each ventilator parameter is determined through boundary prediction. The normalized parameter adjustment action vector is then mapped to the dynamic safety adjustment range to obtain the ventilator parameter control scheme.

7. The method according to claim 6, characterized in that, The reinforcement learning agent is trained through a process based on a two-delay deep deterministic policy gradient algorithm. Based on expert demonstration data, the policy network is trained using the behavior cloning algorithm to obtain the initial policy; Based on historical patient interaction data, the initial strategy is trained offline using a conservative Q-learning framework. During this stage, an experience tuple is constructed based on the historical patient interaction data, and the current state, action, reward, and next state are sampled. Based on the next state, the next action is generated through the target policy network. Based on the next state and the next action, the value estimate is calculated through the double Q-value estimation network and the minimum value is taken. The time-series difference target value is calculated by combining the reward and the minimum value of the value estimate. A loss function is constructed based on the difference between the time-series difference target value and the predicted value of the current Q-value estimation network. The loss is calculated and the parameters of the current Q-value estimation network are updated. The parameters of the policy network are updated based on the output of the current Q-value estimation network, wherein the update frequency of the policy network is lower than that of the double Q-value estimation network. According to the soft update mechanism, the parameters of the policy network and the double Q-value estimation network are synchronized to the corresponding target network to obtain the optimized policy. Based on real-time interactive data, the optimization strategy is fine-tuned online in shadow mode. During this stage, experience tuples are constructed based on the real-time interactive data. Each experience tuple includes the current state, the parameter adjustment action vector output by the executed strategy, the actual executed action, the next state, the reward, and a safety verification flag. In the background, the strategy gradient is corrected using an importance sampling method based on the experience tuples, and the parameters of the strategy network and the double Q-value estimation network are updated. At fixed intervals, offline performance and safety assessments are performed on the background updated strategy. When the background updated strategy is superior to the current strategy and meets the safety threshold, a strategy update proposal is generated. After clinical review, the strategy update proposal is deployed as the execution strategy.

8. An electronic device comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 7.

Images