A clinical data analysis system for intelligent patient monitoring and prediction

Abstract

The present invention discloses a clinical data analysis system for intelligent patient monitoring and prediction, belonging to the field of medical analysis, including a real-time monitoring module, a data acquisition module, and a knowledge base construction module; the present invention can improve the quality of patient care while effectively saving medical resources, not only enhancing the processing ability of complex medical data, but also improving the scientificity and rationality of clinical decision-making, providing personalized and timely health management solutions for patients, enabling hospitals to be more forward-looking and proactive in disease management, thereby enhancing the overall medical service level, being able to effectively address the uncertainties in clinical data and the differences in patient individual needs, providing more diverse and flexible scheme suggestions for medical staff in decision-making, making the entire system not only have real-time performance, but also achieve long-term effect optimization across time dimensions, helping to improve the treatment effect and experience of patients, and providing a scientific basis for the reasonable allocation of medical resources.

Description

Technical Field

[0001] The present invention relates to the field of medical analysis, and particularly to a clinical data analysis system for intelligent patient monitoring and prediction. Background Art

[0002] In recent years, with the aggravation of global population aging and the increase in the incidence of chronic diseases, the medical system is facing increasing pressure. With the development of medical big data and the progress of intelligent technologies, traditional patient monitoring and health risk management are gradually turning towards the direction of intelligence and data-driven. In modern medical systems, a vast amount of clinical data, including patients' vital signs, diagnosis and treatment records, laboratory test results, etc., contains rich health information. These data not only have important value for the judgment of the current condition, but also can be used to predict the health risks and development trends of patients. However, due to the complexity, high dimensionality, and strong heterogeneity of medical data, it is difficult for traditional data analysis methods to fully extract the characteristic information therein and make accurate predictions. At the same time, aiming at the personalized needs of different patients, how to provide personalized treatment plans for doctors through effective algorithms has also become an urgent problem to be solved, and has become one of the hotspots in the research and application of the medical field.

[0003] Existing clinical data analysis systems cannot improve the quality of patient care while saving medical resources, and have low processing capabilities for complex medical data, reducing the scientificity and rationality of clinical decisions; in addition, existing clinical data analysis systems cannot cope with the uncertainties in clinical data and the differences in patients' individual needs, and cannot achieve long-term effect optimization across time dimensions, which is not conducive to improving the treatment effects and experiences of patients. Therefore, we propose a clinical data analysis system for intelligent patient monitoring and prediction. Summary of the Invention

[0004] The purpose of the present invention is to solve the defects existing in the prior art, and to propose a clinical data analysis system for intelligent patient monitoring and prediction.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions:

[0006] A clinical data analysis system for intelligent patient monitoring and prediction, including a real-time monitoring module, a data acquisition module, a knowledge base construction module, an integration and extraction module, a health prediction module, a simulation and verification module, a visualization display module, a clinical decision support module, an abnormal alarm module, an optimization application module, and a decision optimization module;

[0007] The real-time monitoring module is used for continuously monitoring the physiological indicators of patients and comparing them with historical data;

[0008] The data acquisition module is used for automatically collecting patient data from various medical devices and external electronic medical record systems;

[0009] The knowledge base building module is used to associate the patient's clinical data with medical knowledge to construct a clinical medicine map;

[0010] The integration and extraction module is used to clean and integrate patient data from different sources, analyze data features, and extract key clinical feature information;

[0011] The health prediction module is used to analyze the extracted features and predict the health risks and development trends of the patients;

[0012] The simulation verification module is used to evaluate the performance of the health prediction module in different scenarios;

[0013] The visualization display module is used to visualize the prediction results and analysis information;

[0014] The clinical decision support module is used to provide medical staff with targeted treatment plans based on the prediction results;

[0015] The abnormal alarm module is used to automatically issue an alarm when abnormalities occur in the monitoring data;

[0016] The optimization application module is used to optimize the parameter information of each module of the system;

[0017] The decision optimization module is used to recommend the most suitable personalized treatment plan based on the patient's real-time health data and dynamically adjust the treatment plan.

[0018] As a further solution of the present invention, the integrated extraction module extracts key clinical feature information in the following specific steps:

[0019] S1.1: Detect the patient data of each patient in the record, and compare the characteristic values of each data record to determine whether there are exactly the same records, and delete the duplicate patient data, leaving only one copy, and detect the missing values in each group of patient data. If the missing field of the patient data is less than the preset threshold, the field is directly deleted. If the missing numerical data in the patient data is less than the preset filling value, the missing data is filled by mean filling;

[0020] S1.2: Arrange the patient data in ascending order, calculate the first quartile Q1, the second quartile Q2, and the third quartile Q3 of the arranged patient data, and calculate the interquartile range IQR by the formula IQR=Q3-Q1, and then determine the outlier range [Q1-1.5×IQR, Q3+1.5×IQR] by the quartile rule. If there is patient data that exceeds the outlier range, mark the patient data as outlier data and remove it;

[0021] S1.3: Pass Calculate the Z-score of the remaining patient data that cannot be sorted in ascending order, where represents the standardized score of the patient data, represents the patient data value, represents the mean of the patient data, represents the standard deviation of the patient data. If the absolute Z-score of the patient data is greater than 3, the patient is regarded as abnormal and excluded;

[0022] S1.4: Check the same patient data in different data sources, unify its naming and unit, adopt the timestamp alignment strategy, integrate the records in different data sources by time, and unify the data format to obtain the integrated patient dataset. Then, standardize each group of data in the patient dataset;

[0023] S1.5: Calculate the linear correlation between each group of patient characteristics and the target variable set by medical staff through the Pearson correlation coefficient, screen the characteristics with a correlation higher than the preset correlation threshold with the target, use PCA to reduce the feature dimension, then calculate the information gain of each group of patient characteristics, screen the characteristics with an information gain higher than the preset gain threshold, analyze the influence of each group of patient characteristics on the variance of the target variable, and extract the characteristics with an influence on the target variance higher than the preset influence value.

[0024] As a further solution of the present invention, the specific steps for the health prediction module to predict the health risks and development trends of patients are as follows:

[0025] S2.1: Arrange the extracted groups of patient characteristics in chronological order, and normalize each group of patient characteristics to make the data range of each feature consistent. Then, convert the processed patient characteristics into a three-dimensional format, i.e., [number of samples, number of time steps, number of features], and build a bidirectional GRU network architecture including an input layer, a Bi-GRU layer, and an output layer to construct the required prediction model;

[0026] S2.2: Divide the processed patient characteristics into a training set and a validation set, and transmit the patient characteristics in the training set to the prediction model. The input layer of the prediction model receives the preprocessed patient characteristics and performs forward GRU unit calculation and backward GRU unit calculation on the patient characteristics through the Bi-GRU layer;

[0027] S2.3: The output layer receives the output data of the Bi-GRU layer, combines the hidden states of the forward GRU unit and the backward GRU unit to form the final output of the bidirectional GRU, calculates the error between the predicted value and the true label through the mean square error loss function, calculates the gradient of the loss function with respect to the network parameters of each layer of the prediction model through the backpropagation algorithm, and updates the parameters through the Adam optimizer based on the calculated gradient;

[0028] S2.4: Input the patient characteristics in the validation set into the trained prediction model, calculate the loss value and accuracy performance metrics of the model on the validation set, evaluate its prediction effect on the patient's health risks and development trends. If the performance metrics do not reach the preset target values, retrain and validate the prediction model until the model performance metrics reach the preset target values;

[0029] S2.5: Deploy the trained model in the clinical system, input the latest patient data into the prediction model for forward propagation calculation, input the forward propagation result into the output layer of the model, and the output layer performs classification calculation on the received data;

[0030] S2.6: If the target is binary classification, use the sigmoid activation function to output the risk probability. If the target is multi-classification or regression, use the softmax activation function or linear activation function to output the corresponding predicted values respectively, and interpret each group of prediction results, determine the risk score and make a judgment;

[0031] 2.7: Re-collect data at fixed intervals and update the model prediction. When the prediction result shows that the risk value is higher than the set safety threshold, the alarm module automatically sends an alarm to the medical staff and records the corresponding patient data for further analysis.

[0032] As a further solution of the present invention, the specific calculation formula of the forward GRU unit described in S2.2 is as follows:

[0033]

[0034] In the formula, represents the update gate, which is used to control the update amount of the previous hidden state; represents the Sigmoid activation function, which is used to map the input value between 0 and 1; represents the weight matrix of the update gate; represents the hidden state at the th time step; represents the input at the current th time step; represents the bias term of the update gate; represents the reset gate, which is used to control the amount of forgetting; represents the weight matrix of the reset gate; represents the bias term of the reset gate; represents the candidate hidden state; represents the weight matrix of the candidate hidden state; represents the bias term of the candidate hidden state; represents the hidden state at the

[0035] The specific calculation formula of the mean square error loss function described in S2.3 is as follows:

[0036]

[0037] In the formula, represents the mean square error value; represents the number of patient data; represents the index of the patient data; represents the actual value of the th patient data; represents the predicted value of the

[0038] As a further solution of the present invention, the specific steps for the simulation verification module to evaluate the performance of the health prediction module in different scenarios are as follows:

[0039] S3.1: Collect the status of the current patient data and the prediction results corresponding to the prediction model, and use them as the initial nodes. Then, use the status after the subsequent single prediction model simulation as the child nodes of the initial nodes, and calculate the UCB values of the generated groups of child nodes;

[0040] S3.2: Use the upper confidence bound strategy to gradually select the child node with the highest UCB value until an uncompletely expanded node, i.e., a leaf node, is reached. Expand the selected node, add one or more new prediction scenarios for the current node status, and generate one or more child nodes, where the prediction scenario is specifically the performance of the prediction model under new patient data, different feature combinations, or different input data;

[0041] S3.3: Starting from the expanded new node, simulate the behavior of the prediction model, perform a complete simulation starting from the current node, use random sampling or model prediction to obtain the results on the simulation path, record the performance of the model on this simulation path, and calculate the corresponding reward value for the model performance according to the simulation results;

[0042] S3.4: Gradually backtrack from the termination node to the root node, update the visit times and cumulative rewards of each node, repeat the steps of selection, expansion, simulation, and backtracking until the preset iteration time is reached, and gradually select the child node with the highest reward value from the root node until the termination node is reached;

[0043] S3.5: Use the selected path as the best verification path of the model, analyze the performance of the model under the best path, evaluate the accuracy, stability of the model in different scenarios and its prediction ability for the health risks of patients, identify the advantages and disadvantages of the model under each feature combination, time period, or input data, and improve the prediction model based on the analysis results.

[0044] As a further solution of the present invention, the specific steps for the clinical decision support module to provide targeted clinical suggestions to doctors are as follows:

[0045] S4.1: Collect different health states of the patient during the treatment process and construct the corresponding state space , and take the treatment decisions or drug adjustment plans that medical staff can take as actions in the MDP to construct the action space ;

[0046] S4.2: According to historical data, calculate the probability that the patient transfers from the current health state to the next health state after the treatment plan , where represents the health state of the patient within time , represents the health state of the patient within time , , and , and is within the range of [1, m];

[0047] S4.3: According to the standard that improving the patient's health state by the treatment plan gives a positive reward, and causing side effects or deterioration gives a negative reward, construct the corresponding reward function to evaluate the improvement of the patient's health after taking each treatment plan. In the state space , assign a random treatment plan to each state to initialize the policy, and denote it as ;

[0048] S4.4: Through the Bellman equation, update the value function of each state according to the actions executed by the current policy , and use the recursive calculation method to gradually iterate and update until the change in the value function of all states is less than the preset threshold. Then, in each state, select the action that maximizes the value function , and use it as the action of the new policy and update the original policy;

[0049] S4.5: Compare the policies before and after the update. If they are the same in all states, it is considered that the policy has converged; otherwise, continue to execute the policy evaluation and improvement steps using the new policy until the policy no longer changes in all states, and then provide the treatment plan in the optimal policy to the medical staff.

[0050] As a further solution of the present invention, the specific calculation formula of the Bellman equation described in S4.4 is as follows:

[0051]

[0052] In the formula, represents the state value under the current policy; represents the next health state; represents that the patient starts from state through the policy action transfers to state probability; represents the reward when taking the current policy action ; represents the discount factor, which is used to balance current and future rewards; represents the state value under the current policy.

[0053] As a further solution of the present invention, the specific steps for the decision optimization module to dynamically adjust the treatment plan are as follows:

[0054] S5.1: The decision optimization module collects in real time the health state of the patient during treatment under the treatment plan of the optimal policy, initializes the pheromone and heuristic function of each treatment path in the treatment plan, then randomly generates multiple groups of individuals, and uses the current health state of the patient as the initial node;

[0055] S5.2: According to the pheromone and heuristic function of each treatment path, calculate the selection probability of all treatment paths in the current state, and randomly select the next action according to each selection probability;

[0056] S5.3: After each group of individuals selects a path, simulate the treatment result of the treatment path, record the effect of each path according to the simulated treatment result, and update the pheromone on the passed path after each group of individuals completes its path;

[0057] S5.4: Repeat path selection and pheromone update. At the same time, in each round of iteration, all individuals will re-select paths, update pheromones, and adjust the priority of path selection according to the simulation results until the maximum number of iterations is reached or the change in pheromone concentration converges within a preset range;

[0058] S5.5: After all iterations are completed, evaluate the pheromone concentration of all treatment paths, select the path with the highest pheromone concentration, record the corresponding treatment means, sort out each treatment plan in the best path, and adjust the treatment plan generated by the clinical decision support module based on the generated groups of treatment plans. At the same time, feedback each group of plans to the medical staff for viewing and selection through the visualization display module.

[0059] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0060] 1. The present invention converts the extracted patient features into a three-dimensional format and constructs a bidirectional GRU network architecture including an input layer, a Bi-GRU layer, and an output layer to build the required prediction model. The processed patient features are divided into a training set and a validation set, and the patient features in the training set are transmitted to the prediction model. The input layer of the prediction model receives the preprocessed patient features, and the Bi-GRU layer performs forward GRU unit calculations and backward GRU unit calculations on the patient features. The output layer receives the output data of the Bi-GRU layer and combines the hidden states of the forward GRU unit and the backward GRU unit to form the final output of the bidirectional GRU, calculates the error between the predicted value and the true value, and adjusts the model parameters through backpropagation. Then, the model is verified, and the training and verification are repeated until the model reaches the preset target value. The state of the current patient data and the corresponding prediction results of the prediction model are collected and used as the initial node. Then, the state after the subsequent simulation of the prediction model is used as the child node of the initial node, and the steps of selection, expansion, simulation, and backtracking are repeated until the preset iteration time is reached. The child node with the highest reward value is gradually selected downward from the root node until the termination node is reached, and the selected path is used as the best verification path of the model. The performance of the model under the best path is analyzed, the accuracy, stability, and its prediction ability for the patient's health risk in different scenarios are evaluated, the advantages and disadvantages of the model in various feature combinations, time periods, or input data are identified, and the prediction model is improved based on the analysis results. The trained model is deployed in the clinical system, and the latest patient data is input into the prediction model for forward propagation calculation. The forward propagation result is input into the output layer of the model, and the output layer performs classification calculation on the received data and outputs the risk probability or predicted value. It can improve the quality of patient care while effectively saving medical resources, not only enhance the processing ability of complex medical data, but also improve the scientificity and rationality of clinical decision-making, provide personalized and timely health management solutions for patients, enable the hospital to be more forward-looking and proactive in disease management, and thus improve the overall medical service level.

[0061] 2. The present invention constructs a corresponding state space and action space by collecting different health states of patients during the treatment process and treatment decisions or drug adjustment plans that medical staff can adopt. Then, based on historical data, it calculates the probability that a patient transfers from the current health state to the next health state after a treatment plan, evaluates the improvement of the patient's health after adopting each treatment plan, assigns a random treatment plan to each state to initialize the strategy, and through the Bellman equation, updates the value function of each state according to the actions executed by the current strategy, and uses a recursive calculation method to gradually iterate and update until the change in the value function of all states is less than a preset threshold. Then, in each state, it selects the action that maximizes the value function, takes it as the action of the new strategy, and updates the original strategy. It compares the strategies before and after the update. If the two are the same in all states, it is considered that the strategy has converged; otherwise, it continues to execute the strategy evaluation and improvement steps using the new strategy until the strategy no longer changes in all states. Then, it provides the treatment plan in the optimal strategy to the medical staff, collects the health state of the patient during the treatment under the treatment plan in the optimal strategy in real time, and then randomly generates multiple groups of individuals. Taking the health state of the current patient as the initial node, it calculates the selection probability of all treatment paths in the current state according to the pheromone and heuristic function of each treatment path, and selects the next action by random selection. After each group of individuals selects a path, it simulates the treatment result of the treatment path, records the effect of each path, and updates the pheromone on the passed path. It repeats the path selection and pheromone update until the maximum number of iterations is reached or the change in pheromone concentration converges to a preset range. After completing all iterations, it evaluates the pheromone concentration of all treatment paths, selects the path with the highest pheromone concentration, records the corresponding treatment means, sorts out each treatment plan in the best path, and adjusts the treatment plan generated by the clinical decision support module based on the generated groups of treatment plans. At the same time, it feeds back each group of plans to the medical staff for viewing and selection through the visualization display module, which can effectively cope with the uncertainty in clinical data and the differences in patient individual needs, provide more diverse and flexible plan suggestions for medical staff when making decisions, make the entire system not only have real-time performance, but also achieve long-term effect optimization across time dimensions, help improve the treatment effect and experience of patients, and provide a scientific basis for the reasonable allocation of medical resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, and do not constitute a limitation to the present invention.

[0063] Figure 1 It is a system block diagram of a clinical data analysis system for intelligent patient monitoring and prediction proposed by the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0064] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0065] Example 1

[0066] Reference Figure 1 , a clinical data analysis system for intelligent patient monitoring and prediction, includes a real-time monitoring module, a data acquisition module, a knowledge base building module, an integration and extraction module, a health prediction module, a simulation verification module, a visualization display module, a clinical decision support module, an abnormal alarm module, an optimization application module and a decision optimization module.

[0067] The real-time monitoring module is used to continuously monitor the patient's physiological indicators and compare them with historical data; the data acquisition module is used to automatically collect patient data from various medical devices and external electronic medical record systems; the knowledge base building module is used to associate the patient's clinical data with medical knowledge and construct a clinical medicine map.

[0068] The integration and extraction module is used to clean and integrate patient data from different sources, analyze data features, and extract key clinical feature information.

[0069] Specifically, the patient data of each patient is detected, and the characteristic values of each data record are compared to determine whether there are exactly the same records, and duplicate patient data are deleted, leaving only one copy. The missing values in each group of patient data are detected. If the missing field of the patient data is less than the preset threshold, the field is directly deleted. If the missing numerical data in the patient data is less than the preset filling value, the missing data is filled by mean filling, and the patient data is arranged in ascending order. Then, the first quartile Q1, the second quartile Q2, and the third quartile Q3 of the arranged patient data are calculated, and the quartile range IQR is calculated by the formula IQR=Q3-Q1, and then the outlier range [Q1-1.5×IQR, Q3+1.5×IQR] is determined by the quartile rule. If there is patient data exceeding the outlier range, the patient data is marked as abnormal data and removed. Calculate the Z score of the remaining patient data that cannot be sorted in ascending order, where represents the standardized score of the patient data, represents the patient data value, represents the mean of the patient data, Represents the standard deviation of the patient data. If the absolute Z-score of the patient data is greater than 3, the patient is considered abnormal and excluded. Check the same patient data in different data sources, unify its naming and units, adopt a timestamp alignment strategy, integrate the records in different data sources by time, and unify the data format to obtain the integrated patient dataset. Then, standardize each group of data in the patient dataset, calculate the linear correlation between each group of patient characteristics and the target variable set by medical staff through the Pearson correlation coefficient, and screen the characteristics with a correlation higher than the preset correlation threshold with the target. Use PCA to reduce the feature dimension, then calculate the information gain of each group of patient characteristics, screen the characteristics with an information gain higher than the preset gain threshold, analyze the impact of each group of patient characteristics on the variance of the target variable, and extract the characteristics with a target variance impact higher than the preset impact value.

[0070] The health prediction module is used to analyze the extracted features and predict the health risks and development trends of patients.

[0071] Specifically, the extracted groups of patient characteristics are arranged in chronological order, and the groups of patient characteristics are normalized to make the data ranges of each characteristic consistent. Then, the processed patient characteristics are converted into a three-dimensional format, i.e., [number of samples, number of time steps, number of features]. A bidirectional GRU network architecture including an input layer, a Bi-GRU layer, and an output layer is built to construct the required prediction model. The processed patient characteristics are divided into a training set and a validation set, and the patient characteristics in the training set are transmitted to the prediction model. The input layer of the prediction model receives the preprocessed patient characteristics, and the Bi-GRU layer performs forward GRU cell calculations and backward GRU cell calculations on the patient characteristics. The output layer receives the output data of the Bi-GRU layer and combines the hidden states of the forward GRU cell and the backward GRU cell to form the final output of the bidirectional GRU. The error between the predicted value and the true label is calculated through the mean square error loss function. The gradients of the loss function with respect to the network parameters of each layer of the prediction model are calculated through the backpropagation algorithm. Based on the calculated gradients, the parameters are updated through the Adam optimizer. The patient characteristics in the validation set are input into the trained prediction model, and the loss value and accuracy performance indicators of the model on the validation set are calculated to evaluate its prediction effect on the patient's health risks and development trends. If the performance indicators do not reach the preset target values, the prediction model is retrained and validated until the model performance indicators reach the preset target values. The trained model is deployed in the clinical system, and the latest patient data is input into the prediction model for forward propagation calculation. The forward propagation result is input into the output layer of the model, and the output layer performs classification calculations on the received data. If the target is binary classification, the sigmoid activation function is used to output the risk probability. If the target is multi-classification or regression, the softmax activation function or the linear activation function is used to output the corresponding predicted values, and the prediction results of each group are interpreted to determine the risk score and make a judgment. Data is collected again at fixed intervals, and the model prediction is updated. When the prediction result shows that the risk value is higher than the set safety threshold, the alarm module automatically sends an alarm to the medical staff and records the corresponding patient data for further analysis.

[0072] In this embodiment, the specific calculation formula for the forward GRU cell calculation is as follows:

[0073]

[0074] In the formula, represents the update gate, which is used to control the update amount of the previous hidden state; represents the Sigmoid activation function, which is used to map the input value to between 0 and 1; represents the weight matrix of the update gate; represents the th hidden state at the current Input for one time step; Bias term representing the update gate; Represents the reset gate, used to control the amount of forgetting; Weight matrix representing the reset gate; Bias term representing the reset gate; Represents the candidate hidden state; Weight matrix representing the candidate hidden state; Bias term representing the candidate hidden state; Represents the hidden state at the

[0075] The specific calculation formula of the mean squared error loss function is as follows:

[0076]

[0077] In the formula, Represents the mean squared error value; Represents the number of patient data; Represents the index of the patient data; Represents the actual value of the

[0078] The simulation verification module is used to evaluate the performance of the health prediction module in different scenarios.

[0079] ​​Specifically, collect the status of current patient data and the prediction results corresponding to the prediction model, and use them as the initial nodes. Then, use the status after the subsequent prediction model simulation as the child nodes of the initial nodes, calculate the UCB values of the generated groups of child nodes, and use the upper confidence bound strategy to gradually select the child node with the highest UCB value until an unexpanded node, that is, a leaf node, is reached. Expand the selected node, add one or more new prediction scenarios for the current node status, generate one or more child nodes, where the prediction scenario specifically refers to the performance of the prediction model under new patient data, different feature combinations, or different input data. Starting from the expanded new node, simulate the behavior of the prediction model, perform a complete simulation starting from the current node, use random sampling or model prediction to obtain the results on the simulation path, record the performance of the model on this simulation path, and calculate the corresponding reward value for the model performance according to the simulation results. Trace back from the termination node to the root node step by step, update the visit times and cumulative rewards of each node, repeat the steps of selection, expansion, simulation, and backtracking until the preset iteration time is reached, select the child node with the highest reward value from the root node downward step by step until the termination node is reached, use the selected path as the best verification path of the model, analyze the performance of the model under the best path, evaluate the accuracy, stability of the model in different scenarios, and its prediction ability for patient health risks, identify the advantages and disadvantages of the model under each feature combination, time period, or input data, and improve the prediction model based on the analysis results.

[0080] Embodiment 2

[0081] Refer to Figure 1 A clinical data analysis system for intelligent patient monitoring and prediction includes a real-time monitoring module, a data acquisition module, a knowledge base construction module, an integration and extraction module, a health prediction module, a simulation and verification module, a visualization display module, a clinical decision support module, an abnormal alarm module, an optimization application module, and a decision optimization module.

[0082] The visualization display module is used to visualize the prediction results and analysis information; the clinical decision support module is used to provide targeted treatment plans for medical staff according to the prediction results.

[0083] Specifically, collect different health states of patients during the treatment process and construct the corresponding state space Take the treatment decisions or drug adjustment plans that medical staff can take as actions in the MDP to construct the action space According to historical data, calculate the probability that the patient transfers from the current health state after the treatment plan to the next health state where represents patient time ​ The health status within represents the patient time The health status within , and Within the range of [1, m], according to the treatment plan to improve the patient's health status, a positive reward is given; if it causes side effects or deterioration, a negative reward is given, and a corresponding reward function is constructed , to evaluate the improvement of the patient's health after adopting each treatment plan. In the state space , a random treatment plan is assigned to each state to initialize the policy, and it is denoted as , through the Bellman equation, according to the actions executed by the current policy, update the value function of each state , and use the recursive calculation method to gradually iterate and update , until the change in the value function of all states is less than the preset threshold. Then, in each state, select the action that maximizes the value function , take it as the action of the new policy, and update the original policy. Compare the policies before and after the update. If the two are the same in all states, it is considered that the policy has converged; otherwise, continue to execute the policy evaluation and improvement steps using the new policy until the policy no longer changes in all states, and then provide the treatment plan in the optimal policy to the medical staff.

[0084] In this embodiment, the specific calculation formula of the Bellman equation is as follows:

[0085]

[0086] In the formula, represents the value of the state under the current policy; represents the next health status; represents the patient from the state through the policy action transfer to the state probability; represents taking the current policy action when the return; represents the discount factor, which is used to balance the current and future rewards; represents the state under the current policy value.

[0087] The abnormal alarm module is used to automatically send an alarm when the monitoring data is abnormal; the optimization application module is used to optimize the parameter information of each module of the system; the decision optimization module is used to recommend the most suitable personalized treatment plan according to the patient's real-time health data and dynamically adjust the treatment plan.

[0088] Specifically, the decision optimization module collects the health status of the patient during treatment under the optimal treatment plan in real time, initializes the pheromone and heuristic function of each treatment path in the treatment plan, then randomly generates multiple groups of individuals, and uses the health status of the current patient as the initial node. According to the pheromone and heuristic function of each treatment path, calculates the selection probability of all treatment paths in the current state, and based on each selection probability, selects the next action by random selection. After each group of individuals selects a path, simulates the treatment results of that treatment path, records the effects of each path according to the simulated treatment results, and updates the pheromone on the passed path after each group of individuals completes its path. Repeat the path selection and pheromone update. At the same time, in each iteration, all individuals will re-select paths, update pheromones, and adjust the priority of path selection according to the simulation results until the maximum number of iterations is reached or the change in pheromone concentration converges within the preset range. After all iterations are completed, evaluate the pheromone concentration of all treatment paths, select the path with the highest pheromone concentration, record the corresponding treatment means, sort out each treatment plan in the best path, and adjust the treatment plan generated by the clinical decision support module based on the generated groups of treatment plans. At the same time, feedback each group of plans to the medical staff through the visualization display module for viewing and selection.

Claims

1. A clinical data analysis system for intelligent patient monitoring and prediction, characterized in that It includes real-time monitoring module, data collection module, knowledge base building module, integration and extraction module, health prediction module, simulation verification module, visualization display module, clinical decision support module, abnormal alarm module, optimization application module and decision optimization module; The real-time monitoring module is used to continuously monitor the patient's physiological indicators and compare them with historical data; The data acquisition module is used to automatically collect patient data from various medical devices and external electronic medical record systems; The knowledge base building module is used to associate the patient's clinical data with medical knowledge to construct a clinical medicine map; The integration and extraction module is used to clean and integrate patient data from different sources, analyze data features, and extract key clinical feature information; The specific steps are as follows: S1.1: Detect the patient data of each patient in the record, and compare the characteristic values of each data record to determine whether there are exactly the same records, and delete the duplicate patient data, leaving only one copy, and detect the missing values in each group of patient data. If the missing field of the patient data is less than the preset threshold, the field is directly deleted. If the missing numerical data in the patient data is less than the preset filling value, the missing data is filled by mean filling; S1.2: Arrange the patient data in ascending order, calculate the first quartile Q1, the second quartile Q2 and the third quartile Q3 of the arranged patient data, and calculate the quartile range IQR by the formula IQR=Q3-Q1, and then determine the outlier range [Q1-1.5×IQR, Q3+1.5×IQR] by the quartile rule. If there is patient data that exceeds the outlier range, mark the patient data as outlier data and remove it; S1.3: By calculating the Z-score of the remaining patient data that cannot be sorted in ascending order, where Z represents the standardized score of the patient data, x represents the patient data value, μ represents the mean of the patient data, and σ represents the standard deviation of the patient data. If the absolute Z-score of the patient data is greater than 3, the patient is regarded as abnormal and excluded; S1.4: Check the same patient data in different data sources, unify their names and units, use timestamp alignment strategy, integrate records in different data sources by time, and unify the data format to obtain the integrated patient data set, and then standardize each group of data in the patient data set; S1.5: Calculate the linear correlation between each group of patient characteristics and the target variable set by medical staff through the Pearson correlation coefficient, and select features whose correlation with the target is higher than the preset correlation threshold. Use PCA to reduce the feature dimension, calculate the information gain of each group of patient characteristics, select features whose information gain is higher than the preset gain threshold, analyze the impact of each group of patient characteristics on the variance of the target variable, and extract features whose impact on the target variance is higher than the preset impact value; The health prediction module is used to analyze the extracted features and predict the health risks and development trends of the patients; The simulation verification module is used to evaluate the performance of the health prediction module in different scenarios; The visualization display module is used to visualize the prediction results and analysis information; The clinical decision support module is used to provide medical staff with targeted treatment plans based on the prediction results; The abnormal alarm module is used to automatically issue an alarm when abnormalities occur in the monitoring data; The optimization application module is used to optimize the parameter information of each module of the system; The decision optimization module is used to recommend the most suitable personalized treatment plan based on the patient's real-time health data and dynamically adjust the treatment plan.

2. The clinical data analysis system for intelligent patient monitoring and prediction according to claim 1, characterized in that, The specific steps for the health prediction module to predict the health risks and development trends of patients are as follows: S2.1: Arrange the extracted groups of patient characteristics in chronological order, and normalize each group of patient characteristics to make the data range of each feature consistent. Then, convert the processed patient characteristics into a three-dimensional format, namely [number of samples, number of time steps, number of features]. Build a bidirectional GRU network architecture including an input layer, a Bi-GRU layer, and an output layer to construct the required prediction model. S2.2: Divide the processed patient characteristics into a training set and a validation set, and transmit the patient characteristics in the training set to the prediction model. The input layer of the prediction model receives the preprocessed patient characteristics and performs forward GRU unit calculations and backward GRU unit calculations on the patient characteristics through the Bi-GRU layer. S2.3: The output layer receives the output data of the Bi-GRU layer, combines the hidden states of the forward GRU unit and the backward GRU unit to form the final output of the bidirectional GRU, calculates the error between the predicted value and the true label through the mean squared error loss function, calculates the gradient of the loss function with respect to the network parameters of each layer of the prediction model through the backpropagation algorithm, and updates the parameters based on the calculated gradient through the Adam optimizer. S2.4: Input the patient characteristics in the validation set into the trained prediction model, calculate the loss value and accuracy performance indicators of the model on the validation set, evaluate its prediction effect on the patient's health risks and development trends. If the performance indicators do not reach the preset target values, retrain and validate the prediction model until the model performance indicators reach the preset target values. S2.5: Deploy the trained model in the clinical system, input the latest patient data into the prediction model for forward propagation calculation, input the forward propagation result into the output layer of the model, and the output layer performs classification calculation on the received data. S2.6: If the target is binary classification, use the sigmoid activation function to output the risk probability. If the target is multi-classification or regression, use the softmax activation function or the linear activation function to output the corresponding predicted values respectively, and interpret each group of prediction results to determine the risk score and make a judgment. 2.7: Collect data again at fixed intervals and update the model prediction. When the prediction result shows that the risk value is higher than the set safety threshold, the alarm module automatically sends an alarm to the medical staff and records the corresponding patient data for further analysis.

3. An intelligent patient monitoring and prediction clinical data analysis system according to claim 2, characterized in that The specific calculation formula for the forward GRU unit calculation described in S2.2 is as follows: z i = α(W z ·[h i-1 , y i + b z ) r i = α(W r · [h i-1 , y i + b r ) where z i represents the update gate, which is used to control the amount of update of the previous hidden state; α represents the Sigmoid activation function, which is used to map the input value between 0 and 1; W z represents the weight matrix of the update gate; h i-1 represents the hidden state at the (i - 1)-th time step; y i represents the input at the current i-th time step; b z represents the bias term of the update gate; r i represents the reset gate, which is used to control the amount of forgetting; W r represents the weight matrix of the reset gate; b r represents the bias term of the reset gate; represents the candidate hidden state; W h Weight matrix representing the candidate hidden state; b h Bias term representing the candidate hidden state; h i Represents the hidden state at the i-th time step; The specific calculation formula for the mean squared error loss function described in S2.3 is as follows: Wherein, MSE represents the mean square error value; N represents the number of patient data; j represents the index of the patient data; q j represents the actual value of the j-th patient data; represents the predicted value of the j-th patient data.

4. An intelligent patient monitoring and prediction clinical data analysis system according to claim 2, characterized in that, The specific steps for the simulation verification module to evaluate the performance of the health prediction module in different scenarios are as follows: S3.1: Collect the status of the current patient data and the corresponding prediction results of the prediction model, and use them as the initial node. Then, use the status after the subsequent simulation of the prediction model as the child node of the initial node, and calculate the UCB values of the generated groups of child nodes. S3.2: Use the upper confidence bound strategy to gradually select the child node with the highest UCB value until an unexpanded node, i.e., a leaf node, is reached. Expand the selected node, add one or more new prediction scenarios for the current node state, and generate one or more child nodes, where the prediction scenario is specifically the performance of the prediction model under new patient data, different feature combinations, or different input data. S3.3: Starting from the newly expanded node, simulate the behavior of the prediction model, perform a complete simulation starting from the current node, use random sampling or model prediction to obtain the results on the simulation path, record the performance of the model on this simulation path, and calculate the corresponding reward value for the model performance based on the simulation results. S3.4: Gradually backtrack from the termination node to the root node, update the visit count and cumulative reward of each node, repeat the steps of selection, expansion, simulation, and backtracking until the preset iteration time is reached, and then gradually select the child node with the highest reward value from the root node until the termination node is reached. S3.5: Use the selected path as the best verification path of the model, analyze the performance of the model under the best path, evaluate the accuracy, stability of the model in different scenarios and its prediction ability for the health risks of patients, identify the advantages and disadvantages of the model under each feature combination, time period, or input data, and improve the prediction model based on the analysis results.

5. An intelligent patient monitoring and prediction clinical data analysis system according to claim 2, characterized in that, The specific steps for the clinical decision support module to provide targeted clinical suggestions to doctors are as follows: S4.1: Collect different health states of the patient during the treatment process and construct the corresponding state space S = {s1, s2,..., s n}, and take the treatment decisions or drug adjustment plans that medical staff can adopt as the actions in the MDP to construct the action space A = {a1, a2,..., a m}; S4.2: Calculate, based on historical data, the probability P(s t after treatment plan a p to transfer to the next health state s t+1 ), where s t+1 |s t , a p ). Here, s t represents the health state of the patient within time t, s t+1 represents the health state of the patient within time t + 1, a p ∈A, and p is in the range of [1, m]; S4.3: Based on the criterion that improving the patient's health status according to the treatment plan gives a positive reward, and causing side effects or deterioration gives a negative reward, construct the corresponding reward function R(s t ,a p ,s t+1 ), to evaluate the improvement of the patient's health after adopting each treatment plan. In the state space S, assign a random treatment plan to each state to initialize the policy, and denote it as π(s); S4.4: Through the Bellman equation, update the value function V(s) of each state according to the actions executed by the current strategy, and use the recursive calculation method to gradually iterate and update V(s) until the change in the value function of all states is less than the preset threshold. Then, in each state, select the action that maximizes the value function V(s), use it as the action of the new strategy, and update the original strategy. S4.5: Compare the strategies before and after the update. If the two are the same in all states, it is considered that the strategy has converged; otherwise, continue to execute the strategy evaluation and improvement steps using the new strategy until the strategy no longer changes in all states, and then provide the treatment plan in the optimal strategy to the medical staff.

6. The clinical data analysis system for intelligent patient monitoring and prediction according to claim 5, characterized in that, The specific calculation formula of the Bellman equation described in S4.4 is as follows: In the formula, V(s) represents the value of state s under the current strategy; s′ represents the next health state; P(s′|s,π(s)) represents the probability that the patient transfers from state s to state s′ through the strategy action π(s); R(s,π(s),s′) represents the reward when taking the current strategy action π(s); γ represents the discount factor, which is used to balance the current and future rewards; V(s′) represents the value of state s′ under the current strategy.

7. An intelligent patient monitoring and prediction clinical data analysis system according to claim 5, characterized in that The specific steps for the decision optimization module to dynamically adjust the treatment plan are as follows: S5.1: The decision optimization module collects in real time the health states of the patient during the treatment under the treatment plan in the optimal strategy, initializes the pheromone and heuristic function of each treatment path in the treatment plan, and then randomly generates multiple groups of individuals, using the current health state of the patient as the initial node. S5.2: Calculate the selection probabilities of all treatment paths in the current state according to the pheromones and heuristic functions of each treatment path, and select the next action by random selection according to each selection probability; S5.3: After each group of individuals selects a path, simulate the treatment results of the treatment path, record the effects of each path according to the simulated treatment results, and update the pheromones on the passed paths after each group of individuals completes their paths; S5.4: Repeat path selection and pheromone update. At the same time, in each iteration, all individuals will re-select paths, update pheromones, and adjust the priority of path selection according to the simulation results until the maximum number of iterations is reached or the change in pheromone concentration converges within a preset range; S5.5: After all iterations are completed, evaluate the pheromone concentrations of all treatment paths, select the path with the highest pheromone concentration, record the corresponding treatment means, sort out each treatment plan in the optimal path, and adjust the treatment plan generated by the clinical decision support module based on the generated treatment plans of each group. At the same time, feedback the treatment plans of each group to the medical staff for viewing and selection through the visualization display module.

Images