Postoperative care decision support method and system

Abstract

The invention provides a postoperative care decision support method and system, and the method comprises the steps: fusing the physiological time sequence data of a patient, a static medical record and a real-time inspection result, constructing a multi-modal state vector St, and inputting a deep Q network; the network includes a shared coding layer, a state value assessment stream, a dual-channel dominant function stream, and a clinical event prediction stream. Calculating an action Q value Q (St, a) through a state value and a dominant function, then calculating a Shannon entropy H of the action Q value, mapping the Shannon entropy H into an exploration rate epsilon, selecting a current action At by adopting an epsilon-greedy strategy, calculating an award Rt according to patient response and a clinical event after execution, forming a state transition tetrad (St, At, Rt, St + 1), and storing the state transition tetrad into an experience pool; and calculating the priority P in combination with the time difference error delta and the clinical risk C during sampling, carrying out preferential sampling, and updating network parameters through weighted summation of Q value loss and clinical event prediction loss.

Description

Technical Field

[0001] This application belongs to the field of decision support, and in particular relates to a method and system for postoperative nursing decision support. Background Technology

[0002] Postoperatively, clinicians must comprehensively analyze a patient's dynamic physiological signs, static medical records, and various laboratory test results within a limited timeframe to make precise treatment or nursing decisions. This process heavily relies on physicians' professional knowledge and clinical experience, facing risks of information overload, cognitive burden, and potential decision delays or errors. Traditional clinical decision support systems are mostly based on expert knowledge bases and rule engines. While they offer some interpretability, rule development is time-consuming and labor-intensive, struggles to cover all clinical scenarios, and is difficult to update and maintain. Supervised learning-based machine learning models are used to predict specific risks such as postoperative complications or mortality, but these models typically perform static predictions and cannot provide direct guidance for sequential and dynamic decision-making throughout the entire postoperative care process.

[0003] Deep reinforcement learning (DRL) models clinical decision-making as an interaction between an agent and its environment, learning optimal decision-making strategies from historical data to maximize long-term cumulative rewards, such as patient health benefits. However, standard DQN models struggle to effectively integrate temporal, static, and other multimodal information when dealing with heterogeneous medical data, resulting in insufficient representation of patient states. The nature of actions in clinical decision-making varies greatly; for example, the risks and impacts of proactive intervention versus observation and waiting are drastically different. Existing network architectures, such as Dueling DQN, typically evaluate the advantage function of all actions indiscriminately, failing to reflect these inherent differences. Fixed or simply decaying exploration strategies, such as ε-greedy strategies, cannot adaptively adjust to decision uncertainty, leading to low exploration efficiency and neglecting the learning value of key samples such as high-risk states or rare adverse events, affecting the model's convergence speed and performance. This not only limits the depth of the model's state representation learning but also reduces its reliability and practicality in clinical practice. Summary of the Invention

[0004] To address the issues of existing postoperative nursing decision-making processes failing to reflect differences and having low reliability, this application provides a postoperative nursing decision support method and system.

[0005] Firstly, this application proposes a postoperative care decision support method, comprising the following steps: We acquire postoperative physiological signs time-series data, static medical record data, and real-time test data from patients, and fuse them to construct a multimodal state vector at the current time t. ; the multimodal state vector The input is fed into a deep Q-network, which extracts feature vectors through a shared coding layer and then inputs them in parallel into a state value assessment stream, a separate dual-channel dominance function stream, and a clinical event prospective prediction stream. The separate dual-channel dominance function stream includes an intervention action dominance channel and an observation / waiting dominance channel. The Q value of each alternative action a is calculated by combining the output of the state value assessment stream with the output of the dominance channel corresponding to each action. Shannon entropy is calculated based on the set of Q-values ​​of all candidate actions in the current state, and the Shannon entropy is mapped to the exploration rate ε according to a preset nonlinear function; an ε-greedy strategy is used to select and execute the current action based on the Q-values. ; Perform actions Then, the reward value is calculated based on changes in the patient's physiological state and whether a pre-set adverse clinical event occurs. And obtain the state at the next moment. Transition state quadruple Store in the experience replay pool; When sampling data from the experience replay pool to update network parameters, the time difference error based on each quadruple is used. State during summation and transition Associated pre-defined clinical risk levels Calculate the overall priority Priority sampling is performed; the temporal difference loss of the Q value and the adverse outcome prediction loss of the prospective prediction stream output of the clinical event are weighted and summed to form a joint loss function to update the deep Q network.

[0006] Secondly, this application provides a postoperative care decision support system, including the following modules: The Q-value calculation module is used to acquire postoperative patients' physiological signs time-series data, static medical record data, and real-time test data, and fuse them to construct a multimodal state vector at the current time t. ; the multimodal state vector The input is fed into a deep Q-network, which extracts feature vectors through a shared coding layer and then inputs them in parallel into a state value assessment stream, a separate dual-channel dominance function stream, and a clinical event prospective prediction stream. The separate dual-channel dominance function stream includes an intervention action dominance channel and an observation / waiting dominance channel. The Q value of each alternative action a is calculated by combining the output of the state value assessment stream with the output of the dominance channel corresponding to each action. The decision-making module calculates the Shannon entropy based on the set of Q-values ​​of all candidate actions in the current state, and maps the Shannon entropy to an exploration rate ε according to a preset nonlinear function; it then selects and executes the current action based on the Q-value using an ε-greedy strategy. ; Perform actions Then, the reward value is calculated based on changes in the patient's physiological state and whether a pre-set adverse clinical event occurs. And obtain the state at the next moment. Transition state quadruple Store in the experience replay pool; The update module is used to update the network parameters based on the time difference error of each quadruple when sampling data from the experience replay pool. State during summation and transition Associated pre-defined clinical risk levels Calculate the overall priority Priority sampling is performed; the temporal difference loss of the Q value and the adverse outcome prediction loss of the prospective prediction stream output of the clinical event are weighted and summed to form a joint loss function to update the deep Q network.

[0007] This invention utilizes a separate dual-channel advantage function stream to perform differentiated value assessments on two distinct clinical behaviors: intervention and observation / waiting. This makes action selection more targeted and rational. A parallel prospective clinical event prediction stream serves as an auxiliary task, enhancing the network's ability to extract deep state features and giving the model foresight to predict future adverse events. Through joint loss function optimization, it ensures that the final decision-making strategy pursues long-term benefits while mitigating key risks, thus improving decision-making safety. Furthermore, using the Shannon entropy of the Q-value set to set the exploration rate correlates the intensity of exploration with decision uncertainty, improving learning efficiency. In the experience replay stage, clinical risk level and temporal difference error are introduced to jointly determine sampling priority, enabling the model to focus on learning key experiences with significant prognostic impact, accelerating the model's convergence speed and performance in handling critical situations. Attached Figure Description

[0008] Figure 1 A flowchart of a specific embodiment; Figure 2 A schematic diagram of multimodal state vector generation; Figure 3 This is a schematic diagram of the adaptive exploration rate based on Shannon entropy; Figure 4 This is a schematic diagram of the comprehensive priority calculation. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. 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. It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with relevant laws, regulations, and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0010] In the following embodiments, only the embodiments of the method class are used as examples for illustration. The embodiments of the system class are the same and will not be described again below.

[0011] A specific embodiment of a postoperative nursing decision support method, such as... Figure 1 As shown, it includes the following steps: Step 1: Obtain postoperative physiological signs time-series data, static medical record data, and real-time test data of the patient, and fuse them to construct a multimodal state vector at the current time t. ; the multimodal state vector The input is fed into a deep Q-network, which extracts feature vectors through a shared coding layer and then feeds them in parallel into a state value assessment stream, a separate dual-channel dominance function stream, and a prospective clinical event prediction stream. The separate dual-channel dominance function stream includes an intervention action dominance channel and an observation / waiting dominance channel. The Q-value of each candidate action 'a' is calculated by combining the output of the state value assessment stream with the output of the dominance channel corresponding to each action. ; Postoperative patient data is collected. In one embodiment, the data includes time-series physiological signs data, static medical record data, and real-time laboratory data. The time-series physiological signs data, such as continuous monitoring values ​​of heart rate, respiratory rate, blood pressure, and blood oxygen saturation, are encoded using a Long Short-Term Memory (LSTM) network. The static medical record data, such as age, gender, surgical type, and history of underlying diseases, along with the real-time laboratory data, such as blood gas analysis and electrolyte levels, are encoded using a Multilayer Perceptron (MLP). Finally, all encoded feature vectors are concatenated to form a unified multimodal state vector. ,like Figure 2 As shown.

[0012] The shared coding layer consists of several fully connected layers used to process the input state vector. Deep feature extraction is performed; the State Value Evaluation Stream is a fully connected network that outputs a single scalar value to compute the state value. The intervention action dominance channel and the observation / wait dominance channel are two independent fully connected networks. The former outputs the dominance function vector for all intervention actions. The latter outputs the scalar value of the advantage function for the observation-wait action. The prospective prediction stream for clinical events is preferably a fully connected network with a Sigmoid activation function, outputting the probability of an adverse event occurring within a specific future time window; in an optional embodiment, the Q-value for intervention action a is calculated using the following formula: For the observation-wait action, its Q value is .

[0013] Step 2: Calculate the Shannon entropy based on the set of Q-values ​​of all candidate actions in the current state, and map the Shannon entropy to the exploration rate ε according to a preset nonlinear function; use an ε-greedy strategy to select and execute the current action based on the Q-value. ; Perform actions Then, the reward value is calculated based on changes in the patient's physiological state and whether a pre-set adverse clinical event occurs. And obtain the state at the next moment. Transition state quadruple Store in the experience replay pool; The Softmax function is used to transform the set of Q-values ​​into a probability distribution, where the temperature parameter is used to adjust the smoothness of the distribution. Based on this probability distribution, the Shannon entropy H is calculated. The calculated entropy H is then normalized to a preset minimum and maximum exploration rate using a linear mapping function, yielding the current exploration rate ε. A random number between zero and one is generated; if this number is less than ε, a random action is selected from all candidate actions. Otherwise, choose the action with the largest Q value as... Reward Value The weighted average consists of three parts: First, a clinical score based on ordinal categorical variables, such as the change in the SOFA sequential organ failure score (positive reward for a lower score, negative reward for a higher score); second, an event penalty, where a large negative penalty is applied if a pre-set serious adverse event, such as septic shock or respiratory failure, occurs in the next moment; and third, action cost, which provides a small negative reward for each intervention to encourage simplified treatment. Next-moment state By collecting patient data at time t+1 and using... The same construction method is used to obtain the state transition quadruple. The state transition quadruple is stored in an experience replay pool implemented using a first-in-first-out double-ended queue structure.

[0014] Step 3, when sampling data from the experience replay pool to update network parameters, based on the time difference error of each quadruple. State during summation and transition Associated pre-defined clinical risk levels Calculate the overall priority Priority sampling is performed; the temporal difference loss of the Q value and the adverse outcome prediction loss of the prospective prediction stream output of the clinical event are weighted and summed to form a joint loss function to update the deep Q network.

[0015] Time difference error The absolute value of the difference between the target Q value and the current Q value; clinical risk level. It is a preset value, the magnitude of which is related to the state. The risk level is positively correlated with the patient's SOFA score; the higher the score, the higher the risk. (Comprehensive priority) Through formula The calculation shows that, among which and Here, k is a weighting coefficient, k is a small constant to avoid zero priority, and α is a hyperparameter controlling the priority. Sampling utilizes a SumTree data structure, based on the priority of each transition. Perform efficient probabilistic sampling.

[0016] The temporal difference loss for Q-values ​​is preferably calculated using the Huber loss function to enhance robustness to outliers, and the bias introduced by priority sampling weights is corrected by introducing importance sampling weights; the adverse outcome prediction loss is calculated using the binary cross-entropy loss function BCE Loss, whose label is whether an adverse event actually occurred within a predetermined time window after the state occurred; the joint loss function is expressed as follows: , where λ is the weight hyperparameter balancing the two tasks; the Adam optimizer is used to perform gradient descent on the joint loss function, thereby synchronously updating all parameters of the entire deep Q network.

[0017] In an optional embodiment, the fusion constructs the multimodal state vector at the current time t. ,include: Temporal data of heart rate, blood pressure, blood oxygen saturation, and respiratory rate are input into a long short-term memory network for encoding; Static medical record data of patient age, gender, preoperative diagnosis, surgical procedure, and comorbidities are combined with real-time test data of blood routine, electrolytes, and blood glucose, and then input into a multilayer sensor for encoding. The output vector of the Long Short-Term Memory network is concatenated with the output vector of the Multilayer Perceptron to form the multimodal state vector. .

[0018] Two parallel neural networks are used to process dynamic time-series data and static or semi-static data respectively, to achieve a comprehensive representation of the patient's condition. For example, for a patient undergoing postoperative monitoring, heart rate, blood pressure, blood oxygen saturation, and respiratory rate data for the past 60 minutes are collected, forming a 60x4 time-series matrix. This matrix is ​​input into a Long Short-Term Memory (LSTM) network with 128 hidden units, which can capture the dynamic trends and potential correlations of changes in vital signs, such as a continuous decrease in blood pressure accompanied by a compensatory increase in heart rate, and output a 128-dimensional vector to summarize this dynamic information.

[0019] Meanwhile, the patient's static information, such as age 68, male, preoperative diagnosis of lung cancer, lobectomy, and type 2 diabetes, along with his latest immediate test results, such as hemoglobin 10.5 g / dL, potassium 3.8 mmol / L, and blood glucose 180 mg / dL, is integrated into a feature vector. This vector is input into a multilayer perceptron with two hidden layers, each with 256 and 128 neurons respectively, for nonlinear transformation and feature extraction, generating a 64-dimensional encoding vector. The 128-dimensional dynamic feature vector output from the Long Short-Term Memory network is concatenated with the 64-dimensional static feature vector output from the multilayer perceptron to form a final 192-dimensional multimodal state vector. .

[0020] In an optional embodiment, the Q-value of each candidate action a is calculated by combining the output of the state value evaluation stream with the output of the dominant channel corresponding to each action. ,include: Intervention actions The formula for calculating the Q value is: ; Observe the waiting action The formula for calculating the Q value is as follows: ; in, The output of the state value assessment stream; For intervention of the dominant channel targeting the movement The output; To observe the output of the waiting dominant channel; M is the total number of intervention actions; This involves summing the dominance values ​​for all M intervention actions.

[0021] To improve the stability and efficiency of learning, a competitive deep Q-network architecture is employed. This architecture decomposes the value of an action into the value of the state itself and the relative advantage of taking a specific action in that state. (State value) Evaluated by a separate network flow, which represents the current state. Below this, the baseline value of the expected return that can be obtained regardless of what subsequent actions are taken. For example, for a patient in a stable vital signs state, its It might be a high positive value, such as 0.9, but for a patient on the verge of shock, its... It might be a negative value, such as -0.5. The advantage value is calculated by a separate channel, measuring the quality of a specific action compared to other actions. For intervention actions, the advantage function can be centered by subtracting the average of the advantage values ​​of all intervention actions, reducing the variance of the Q-value, and ensuring the state value. This allows for a more accurate learning of the true value of a state. For example, suppose there are three intervention actions with advantage values ​​of 0.4, 0.2, and -0.1, with an average of 0.167. The final Q-value calculation for the first action would be the original advantage value of 0.4 minus the average value of 0.167, plus the state value V(S_t). This enables the model to identify the relatively optimal choice even when all available interventions are ineffective, thus allowing for more refined decision-making.

[0022] In an optional embodiment, the mapping formula for mapping the Shannon entropy to the exploration rate ε according to a preset nonlinear function is: ; in, and The maximum and minimum exploration rates are preset; H is the Shannon entropy of the set of Q values ​​in the current state; is the entropy activation threshold; k is the kurtosis coefficient of the function.

[0023] Shannon entropy H is used to quantify the uncertainty of the model regarding the optimal action choice in the current state. When the model's Q-value predictions for all candidate actions are very close, for example, Q-values ​​for four actions are 0.81, 0.83, 0.79, and 0.80 respectively, the Q-value distribution is relatively uniform, and the calculated Shannon entropy H will be relatively high. This indicates that the model is not entirely confident about which action is the best, and is in a state of high uncertainty. A high entropy value H, for example, H equals 1.5, if it exceeds the preset entropy activation threshold... If the result is 0.8, then the power of the exponent is positive, making the denominator close to 1, and the calculated exploration rate ε will be close to its maximum value. For example, 0.9, Figure 3 Let Shannon entropy H and exploration rate be... The diagram illustrates that the exploration rate increases with increasing Shannon entropy. Under high uncertainty, the model will randomly select an action with a high probability, rather than sticking to what is currently considered the optimal action, thus enhancing the exploration of the environment. Conversely, if the model predicts a much higher Q-value for a particular action than for other actions (e.g., 1.5, 0.2, 0.1, 0.3), the entropy H will be very low, for example, 0.4, below the threshold. In this case, the exploration rate ε will approach its minimum. For example, a value of 0.05 will result in a high probability of selecting the action with the highest current Q value, i.e., taking advantage of it. This allows the agent to boldly explore when confused and decisively take advantage when confident, achieving a dynamic balance between exploration and exploitation.

[0024] In an optional embodiment, the calculation of the reward value... ,include: If a patient experiences a pre-set serious adverse event after the action is performed, a pre-set large negative reward value will be given. If one or more core physiological indicators improve, a preset positive reward value is given; If the observation and waiting action is performed and the patient's core physiological indicators remain within the preset safe range, a slight negative reward is given as a time cost. If the core physiological indicators fail to improve within the predetermined time after the intervention is performed, a corresponding negative reward will be given as punishment for resource consumption or ineffective intervention.

[0025] The role of a reward function is to transform complex clinical goals into explicit numerical signals. For example, if a patient develops acute kidney injury within the next 4 hours after an intervention, which is a pre-defined serious adverse event, a large negative reward would be given, for example, a value of -100, thus prompting the model to strive to avoid decision sequences that lead to this outcome in the future.

[0026] On the positive side, if a patient's mean arterial pressure rises from 55 mmHg to 68 mmHg within 15 minutes after administering an increased dose of vasoactive medication, exceeding the preset improvement threshold, a positive reward of +20 is awarded to encourage such effective interventions. For the observation and waiting action, if the patient's core indicators, such as heart rate and oxygen saturation, remain within the target safe range after the intervention (e.g., heart rate consistently between 60 and 100 beats per minute), a slight negative reward of -0.5 is awarded, representing the cost of time wasted and prompting the model to take action when necessary rather than waiting indefinitely. If an intervention, such as an infusion of 250 ml of crystalloid solution, fails to alleviate hypotension after 30 minutes, a moderate negative reward of -5 is awarded to penalize ineffective resource consumption and delayed treatment.

[0027] In an optional embodiment, the calculation of the overall priority... The calculation formula is as follows: ; in, This represents the absolute value of the sample time difference error; For sample state Corresponding clinical risk level score; is the normalization function used to map the absolute value of the time difference error to the clinical risk level score to a uniform scale; w is the preset weighting coefficient.

[0028] This embodiment improves upon the traditional priority experience replay mechanism based solely on time difference error by introducing clinical risk as a second priority dimension. (Time difference error) This reflects the gap between the model's predicted Q-value and the actual observed reward. A larger error indicates that the model's learning of that state transition is less thorough, thus requiring more relearning. For example, a sample with a TD error absolute value of 0.9, significantly higher than other samples, indicates a substantial surprise in the model's prediction at that point, suggesting high learning value. Figure 4 As shown.

[0029] However, for certain clinically extremely dangerous conditions, even if the model can predict the consequences well (i.e., the TD error is small), repeated training is necessary to ensure absolute accuracy. For example, a sample of early-stage septic shock, with its corresponding clinical risk score... The score could be as high as 9, while another stable-state sample with a slightly faster heart rate scored only 2. By normalizing and weighting the TD error and clinical risk score, samples representing clinically critical situations can be given higher sampling priority, even if their TD error is small. The weighting coefficient w, for example, set to 0.7, can be used to adjust the importance of these two factors, allowing for adjustments to correct cognitive biases in the model or enhance its ability to respond to critical clinical scenarios, as needed.

[0030] In an optional embodiment, the joint loss function The calculation formula is: ; in, The Q-value loss is calculated based on the time-series difference error; α is the prediction loss between the prediction results of the prospective prediction stream output of the clinical event and the actual clinical outcome label; α is the weight hyperparameter that balances the two losses.

[0031] A multi-task learning framework is adopted, adding an auxiliary supervised learning task in addition to the main task, namely reinforcement learning decision-making. The loss of the main task... It is the standard Q-learning loss, typically the mean squared error between the predicted Q-value and the target Q-value, which enables the model to learn the optimal long-term decision strategy. For example, for a sample, the model predicts a Q-value of 0.7, while the target Q-value calculated based on subsequent states and rewards is 0.9. Then the TD loss is calculated based on the difference of 0.2 between these two values.

[0032] The auxiliary task utilizes features extracted from the intermediate layers of the deep Q-network to predict whether a specific adverse clinical event, such as acute respiratory distress syndrome, will occur within a fixed time window, such as the next 6 hours. The loss function for this task... The optimal cross-entropy loss function is used to measure the consistency between the model's predicted probability and the true clinical label, i.e., whether the event occurred or not. For example, the model predicts an occurrence probability of 0.3, but the event actually occurred. This will generate a corresponding loss value. A joint loss function is formed by weighting these two losses with a weight α, such as 0.8. When updating network parameters through backpropagation, both tasks must be optimized simultaneously. This ensures that the state representation learned by the network is not only effective for predicting long-term returns but also discriminative for identifying short-term clinical risks.

[0033] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of this application.

[0034] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple, and relevant parts can be referred to the descriptions in the method embodiments. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0035] The method and electronic device for providing product object information provided in this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and its core ideas. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A postoperative nursing decision support method, characterized in that, Includes the following steps: We acquire postoperative physiological signs time-series data, static medical record data, and real-time test data from patients, and fuse them to construct a multimodal state vector at the current time t. ; The multimodal state vector The input is fed into a deep Q-network, which extracts feature vectors through a shared coding layer and then inputs them in parallel into a state value assessment stream, a separate dual-channel dominance function stream, and a clinical event prospective prediction stream. The separate dual-channel dominance function stream includes an intervention action dominance channel and an observation / waiting dominance channel. The Q value of each alternative action a is calculated by combining the output of the state value assessment stream with the output of the dominance channel corresponding to each action. Shannon entropy is calculated based on the set of Q-values ​​of all candidate actions in the current state, and the Shannon entropy is mapped to the exploration rate ε according to a preset nonlinear function; an ε-greedy strategy is used to select and execute the current action based on the Q-values. ; Perform actions Then, the reward value is calculated based on changes in the patient's physiological state and whether a pre-set adverse clinical event occurs. And obtain the state at the next moment. Transition state quadruple Store in the experience replay pool; When sampling data from the experience replay pool to update network parameters, the time difference error based on each quadruple is used. State during summation and transition Associated pre-defined clinical risk levels Calculate the overall priority Priority sampling is performed; the temporal difference loss of the Q value and the adverse outcome prediction loss of the prospective prediction stream output of the clinical event are weighted and summed to form a joint loss function to update the deep Q network.

2. The method according to claim 1, characterized in that, The fusion constructs the multimodal state vector at the current time t. ,include: Temporal data of heart rate, blood pressure, blood oxygen saturation, and respiratory rate are input into a long short-term memory network for encoding; Static medical record data of patient age, gender, preoperative diagnosis, surgical procedure, and comorbidities are combined with real-time test data of blood routine, electrolytes, and blood glucose, and then input into a multilayer sensor for encoding. The output vector of the Long Short-Term Memory network is concatenated with the output vector of the Multilayer Perceptron to form the multimodal state vector. .

3. The method according to claim 1, characterized in that, The calculation of the Q-value of each candidate action a by combining the output of the state value assessment stream with the output of the corresponding advantage channel of each action includes: Intervention actions The formula for calculating the Q value is: ； Observe the waiting action The formula for calculating the Q value is as follows: ; in, The output of the state value assessment stream; For intervention of the dominant channel targeting the movement The output; To observe the output of the waiting dominant channel; M is the total number of intervention actions; This involves summing the dominance values ​​for all M intervention actions.

4. The method according to claim 1, characterized in that, The mapping formula for mapping the Shannon entropy to the exploration rate ε according to a preset nonlinear function is as follows: ； in, and The maximum and minimum exploration rates are preset; H is the Shannon entropy of the set of Q values ​​in the current state; is the entropy activation threshold; k is the kurtosis coefficient of the function.

5. The method according to claim 1, characterized in that, The calculation of reward value ,include: If a patient experiences a pre-set serious adverse event after the action is performed, a pre-set large negative reward value will be given. If one or more core physiological indicators improve, a preset positive reward value is given; If the observation and waiting action is performed and the patient's core physiological indicators remain within the preset safe range, a slight negative reward is given as a time cost. If the core physiological indicators fail to improve within the predetermined time after the intervention is performed, a corresponding negative reward will be given as punishment for resource consumption or ineffective intervention.

6. The method according to claim 1, characterized in that, The calculation of overall priority The calculation formula is as follows: ； in, This represents the absolute value of the sample time difference error; For sample state Corresponding clinical risk level score; is the normalization function used to map the absolute value of the time difference error to the clinical risk level score to a uniform scale; w is the preset weighting coefficient.

7. The method according to claim 1, characterized in that, The joint loss function The calculation formula is: ; in, The Q-value loss is calculated based on the time-series difference error; α is the prediction loss between the prediction results of the prospective prediction stream output of the clinical event and the actual clinical outcome label; α is the weight hyperparameter that balances the two losses.

8. A postoperative nursing decision support system, characterized in that, Includes the following modules: The Q-value calculation module is used to acquire postoperative patients' physiological signs time-series data, static medical record data, and real-time test data, and fuse them to construct a multimodal state vector at the current time t. ; The multimodal state vector The input is fed into a deep Q-network, which extracts feature vectors through a shared coding layer and then inputs them in parallel into a state value assessment stream, a separate dual-channel dominance function stream, and a clinical event prospective prediction stream. The separate dual-channel dominance function stream includes an intervention action dominance channel and an observation / waiting dominance channel. The Q value of each alternative action a is calculated by combining the output of the state value assessment stream with the output of the dominance channel corresponding to each action. The decision-making module calculates the Shannon entropy based on the set of Q-values ​​of all candidate actions in the current state, and maps the Shannon entropy to an exploration rate ε according to a preset nonlinear function; it then selects and executes the current action based on the Q-value using an ε-greedy strategy. ; Perform actions Then, the reward value is calculated based on changes in the patient's physiological state and whether a pre-set adverse clinical event occurs. And obtain the state at the next moment. Transition state quadruple Store in the experience replay pool; The update module is used to update the network parameters based on the time difference error of each quadruple when sampling data from the experience replay pool. State during summation and transition Associated pre-defined clinical risk levels Calculate the overall priority Priority sampling is performed; the temporal difference loss of the Q value and the adverse outcome prediction loss of the prospective prediction stream output of the clinical event are weighted and summed to form a joint loss function to update the deep Q network.

9. The system according to claim 8, characterized in that, The fusion constructs the multimodal state vector at the current time t. ,include: Temporal data of heart rate, blood pressure, blood oxygen saturation, and respiratory rate are input into a long short-term memory network for encoding; Static medical record data of patient age, gender, preoperative diagnosis, surgical procedure, and comorbidities are combined with real-time test data of blood routine, electrolytes, and blood glucose, and then input into a multilayer sensor for encoding. The output vector of the Long Short-Term Memory network is concatenated with the output vector of the Multilayer Perceptron to form the multimodal state vector. .

10. The system according to claim 8, characterized in that, The calculation of the Q-value of each candidate action a by combining the output of the state value assessment stream with the output of the corresponding advantage channel of each action includes: Intervention actions The formula for calculating the Q value is: ； Observe the waiting action The formula for calculating the Q value is as follows: ; in, The output of the state value assessment stream; For intervention of the dominant channel targeting the movement The output; To observe the output of the waiting dominant channel; M is the total number of intervention actions; This involves summing the dominance values ​​for all M intervention actions.

Images