Method, system, and storage medium for weaning strategy generation for invasive mechanical ventilation

Abstract

The application discloses a weaning strategy generation method and system for invasive mechanical ventilation and a storage medium, and can be widely applied to the technical field of medical data processing. The method determines patient subtype types and characteristics of each subtype type by first time sequence average data meeting second preset conditions, dynamically predicts patient state transition conditions in the next period by second time sequence data meeting third preset conditions, and then constructs a Markov decision model according to the patient subtype types and the characteristics of each subtype type, so as to predict optimal weaning strategies of each subtype by taking the patient state transition conditions as input data of the Markov decision model, thereby providing a weaning reference for medical staff for the invasive mechanical ventilation, and improving the accuracy of weaning decisions for multiple-source heterogeneous critical patients.

Description

Technical Field

[0001] This invention relates to the field of medical data processing technology, and in particular to a method, system, and storage medium for generating invasive mechanical ventilation termination strategies. Background Technology

[0002] In related technologies, precise weaning from invasive mechanical ventilation is crucial for improving the prognosis of critically ill patients. Premature weaning endangers life, while delayed weaning can easily lead to ventilator dependence, infection, pneumonia, and waste of medical resources. Currently, some ventilator parameters or screening criteria are used to assist doctors in making weaning decisions. The universally applicable shallow and rapid breathing index has a high accuracy rate, but low specificity. Most existing machine learning prediction models have poor performance or over-screen the enrolled population, making them unsuitable for the diverse and heterogeneous critically ill patients, thus rendering the models unusable in clinical practice. Summary of the Invention

[0003] This invention aims to address at least one of the technical problems existing in the prior art. To this end, this invention proposes a method, system, and storage medium for generating invasive mechanical ventilation cessation strategies, which can improve the accuracy of weaning decisions for multi-source heterogeneous critically ill patients.

[0004] On one hand, embodiments of the present invention provide a method for generating a termination strategy for invasive mechanical ventilation, comprising the following steps:

[0005] Acquire data on invasively mechanically ventilated patients that meet the first preset requirements;

[0006] First time series average data are formed by extracting patient data that meets the second preset conditions from the patient data;

[0007] The patient subtype and the characteristics of each subtype are determined based on the first time series average data;

[0008] The second time series data is composed of patient data that meets the third preset condition obtained from the patient data;

[0009] Based on the second time series data, dynamically predict the patient's condition transition in the next period;

[0010] A Markov decision model is constructed based on the patient subtype and the characteristics of each subtype;

[0011] The patient's state transition information is input into the Markov decision model to predict the optimal stopping strategy for each subtype.

[0012] In some embodiments, acquiring invasive mechanical ventilation patient data that meets the first preset requirements includes:

[0013] Obtain data from several patients undergoing invasive ventilation;

[0014] The first dataset is formed by extracting patient data from the aforementioned data of patients undergoing invasive ventilation whose invasive ventilation duration is greater than or equal to a first preset duration, whose age is greater than or equal to a first preset age, and whose weaning time is greater than the time of death.

[0015] A second dataset is created by extracting patient data from the first dataset that represents patients undergoing invasive mechanical ventilation for the first time.

[0016] The patient data within the second dataset is classified.

[0017] In some embodiments, classifying the patient data within the second dataset includes:

[0018] When the patient data in the second dataset meets the first preset weaning status, the current patient data is classified into the successful data group; the first preset weaning status is that the patient does not need to restart invasive ventilation within 24 hours of ventilator weaning.

[0019] When the patient data in the second dataset meets the second preset weaning status, the current patient data is classified as the failed data group; the second preset weaning status is that the patient needs to restart invasive ventilation or dies within 24 hours after being weaned from the ventilator.

[0020] In some embodiments, the invasive mechanical ventilation patient data includes the time point at which the patient underwent invasive ventilation, the patient's demographic characteristics, basic patient characteristics, laboratory indicators, scores, respiratory parameters, or outcome indicators.

[0021] In some embodiments, obtaining patient data that meets a second preset condition from the patient data to form a first time series average data includes:

[0022] Extract the patient data belonging to the 24 hours after the start of ventilation, and calculate the average value of each indicator in the patient data belonging to the 24 hours after the start of ventilation.

[0023] In some embodiments, determining the patient subtype and the characteristics of each subtype based on the first time-series average data includes:

[0024] The first time series average data is preprocessed, and the preprocessing includes data cleaning, splitting the dataset, data imputation, data standardization, and data dimensionality reduction.

[0025] Cluster analysis was performed on the preprocessed patient data to obtain patient subtypes and the characteristics of each subtype.

[0026] In some embodiments, constructing a Markov decision model based on the patient subtype and the features of each subtype includes:

[0027] A Markov decision model is constructed based on the patient subtype and the characteristics of each subtype, the Markov decision model including state, action, reward and probability transition;

[0028] The Markov decision model is trained using the heterogeneous policy Q-learning method.

[0029] On the other hand, embodiments of the present invention provide a system for generating a termination strategy for invasive mechanical ventilation, comprising:

[0030] The acquisition module is used to acquire data of invasive mechanical ventilation patients that meet the first preset requirements;

[0031] The first screening module is used to obtain patient data that meets the second preset conditions from the patient data to form a first time series average data.

[0032] The determination module is used to determine the patient subtype and the characteristics of each subtype based on the first time series average data;

[0033] The second screening module is used to obtain patient data that meets the third preset conditions from the patient data to form a second time series data;

[0034] The first prediction module is used to dynamically predict the patient's state transition in the next period based on the second time series data;

[0035] The module is used to construct a Markov decision model based on the patient subtype and the features of each subtype.

[0036] The second prediction module is used to input the patient's state transition information into the Markov decision model to predict the optimal stopping strategy for each subtype.

[0037] On the other hand, embodiments of the present invention provide a system for generating a termination strategy for invasive mechanical ventilation, comprising:

[0038] At least one memory for storing programs;

[0039] At least one processor is configured to load the program to execute the method for generating the invasive mechanical ventilation termination strategy.

[0040] On the other hand, embodiments of the present invention provide a storage medium storing a computer-executable program, which, when executed by a processor, is used to implement the method for generating the invasive mechanical ventilation termination strategy.

[0041] The present invention provides a method for generating a termination strategy for invasive mechanical ventilation, which has the following beneficial effects:

[0042] This invention determines patient subtypes and characteristics of each subtype using average time-series data that meets a second preset condition, and dynamically predicts patient status transitions in the next time period using second time-series data that meets a third preset condition. Then, a Markov decision model is constructed based on the patient subtypes and characteristics of each subtype, using patient status transitions as input data to predict the optimal cessation strategy for each subtype. This provides medical staff with a reference for weaning from invasive mechanical ventilation, improving the accuracy of weaning decisions for multi-source heterogeneous critically ill patients.

[0043] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0044] The present invention will be further described below with reference to the accompanying drawings and embodiments, wherein:

[0045] Figure 1 This is a flowchart of a method for generating a termination strategy for invasive mechanical ventilation according to an embodiment of the present invention;

[0046] Figure 2 This is a flowchart illustrating the screening process for invasive mechanical ventilation patient data according to an embodiment of the present invention.

[0047] Figure 3 This is a flowchart illustrating the identification of subtypes of critically ill patients requiring invasive ventilation, as described in an embodiment of the present invention.

[0048] Figure 4 This is a flowchart illustrating the construction of the optimal stopping strategy model according to an embodiment of the present invention;

[0049] Figure 5 This is a schematic diagram of the decision-making time points in an embodiment of the present invention. Detailed Implementation

[0050] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0051] In the description of this invention, "several" means one or more, "multiple" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0052] In the description of this invention, unless otherwise explicitly defined, terms such as "set up," "install," and "connect" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this invention in conjunction with the specific content of the technical solution.

[0053] In the description of this invention, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0054] Reference Figure 1 This invention provides a method for generating a termination strategy for invasive mechanical ventilation. This method can be applied to the backend processor of cloud servers and medical platforms. Specifically, in the application process... Figure 1 The method shown includes, but is not limited to, the following steps:

[0055] Step 110: Obtain data of invasive mechanical ventilation patients that meet the first preset requirements.

[0056] In this embodiment, a first dataset can be formed by extracting patient data from a plurality of invasive ventilation patient data sets. Patient data sets must have an invasive ventilation duration greater than or equal to a first preset duration, a patient age greater than or equal to a first preset age, and a weaning time greater than a death time. Then, a second dataset can be created from the first dataset set by extracting patient data sets representing patients undergoing invasive mechanical ventilation for the first time. Finally, the patient data in the second dataset set can be categorized. The plurality of invasive ventilation patient data sets can be historical data, such as relevant data directly retrieved from a hospital platform. In this embodiment, after obtaining the second dataset set, the relationship between the patient data in the second dataset set and a first preset weaning state and a second preset weaning state can be determined. Specifically, the first preset weaning state is that the patient does not need to restart invasive ventilation within 24 hours of ventilator weaning; the second preset weaning state is that the patient needs to restart invasive ventilation within 24 hours of ventilator weaning or has died. For example, if the patient data in the second dataset set meets the first preset weaning state, the current patient data is categorized as a successful data group; if the patient data in the second dataset set meets the second preset weaning state, the current patient data is categorized as a failed data group.

[0057] In this embodiment, the occurrence of data for each patient is treated as an event. For example... Figure 2 As shown, invasive mechanical ventilation events were first acquired. Events with a duration of invasive ventilation greater than 24 hours were extracted from these events, excluding those with a ventilation duration of less than 24 hours, those weaning time greater than the time of death, and those with an age under 18 years old. Next, events representing the first use of invasive ventilation were extracted from the events with a duration of invasive ventilation greater than 24 hours, excluding other events. Then, the first-time invasive ventilation events were categorized into a successful group and a failed group. During the categorization process, if the patient needed to restart invasive mechanical ventilation within 48 hours of weaning or died, it was considered a failed weaning attempt; all others were considered successful.

[0058] In this embodiment, the patient data for invasive mechanical ventilation includes the time point at which the patient begins invasive ventilation, the patient's demographic characteristics, basic patient characteristics, laboratory indicators, scores, respiratory parameters, or outcome indicators. Specifically, the time point at which the patient begins invasive ventilation refers to all data from the start to the discontinuation of invasive mechanical ventilation for critically ill patients; the patient's demographic characteristics include age, sex, race, height, weight, and BMI; the patient's basic characteristics include heart rate, respiratory rate, blood oxygen saturation, blood pressure, and body temperature. Laboratory parameters include blood urea nitrogen, creatinine, platelets, red blood cells, hematocrit, white blood cells, hemoglobin, blood glucose, chloride, potassium, sodium, calcium, pH, bicarbonate, partial pressure of oxygen, and partial pressure of carbon dioxide. Scores include the Glasgow Coma Scale (GCS). Respiratory parameters include tidal volume, positive end-expiratory pressure, and oxygen concentration. Outcome parameters include ICU mortality rate, in-hospital mortality rate, length of ICU admission, length of hospital stay, duration of invasive mechanical ventilation, time from admission to ICU admission, time from ICU admission to the start of ventilation, and ventilator weaning success rate. ICU admission disease classification is also included. In this embodiment, all time-related variables are uniformly converted to hours, other indicators are uniformly converted to international standard units, and outliers have been removed based on their range.

[0059] Step 120: Obtain patient data that meets the second preset conditions from the patient data to form a first time series average data, and determine the patient subtype and the characteristics of each subtype based on the first time series average data.

[0060] In this embodiment, as Figure 3As shown, the first time series average data can be obtained by extracting patient data belonging to the 24 hours after the start of ventilation to form the first time series average data, and calculating the average value of each indicator in the patient data belonging to the 24 hours after the start of ventilation. Then, the obtained data can be preprocessed using Python toolkits. Preprocessing includes data cleaning, dataset splitting, data imputation, data standardization, and data dimensionality reduction. Specifically, data cleaning includes using the Pandas toolkit to read the data and remove columns and rows with a missing rate >30%; dataset splitting includes using the Sklearn toolkit to split the dataset into a modeling set (55%) and a validation set (45%); data imputation includes using the miceforest toolkit to fill in missing values ​​in the modeling set and the validation set respectively; data standardization includes using the Sklearn toolkit to perform log transformation on non-normally distributed columns in the two imputed datasets, and then performing standardization and normalization; data dimensionality reduction includes using the Sklearn toolkit to perform Principal Component Analysis (PCA) on the two datasets and find the number of principal components that can explain 80% of the variance of each dataset, thus obtaining the dimensionality-reduced dataset.

[0061] After preprocessing the patient data, cluster analysis can be performed on the preprocessed patient data to obtain patient subtypes and the characteristics of each subtype. Specifically, the clustering process includes K-means clustering and hierarchical clustering, as follows:

[0062] K-means clustering: Use the Sklearn toolkit to draw an elbow plot and find the inflection point; calculate the silhouette coefficient for different numbers of clusters and determine the optimal number of clusters based on the elbow plot; perform cluster analysis based on the number of clusters and draw two-dimensional and three-dimensional plots of the clustering results; perform statistical analysis on the clustering results, such as ANOVA, chi-square test and rank-sum test, analyze the differences of variables between each cluster, and summarize the characteristics of each cluster;

[0063] Hierarchical clustering: Use the Sklearn and SciPy toolkits to draw a dendrogram, find the optimal branching points, and determine the number of clusters; perform cluster analysis based on the number of clusters, and draw two-dimensional and three-dimensional plots of the clustering results; perform statistical analysis on the clustering results, such as ANOVA, chi-square test, and rank-sum test, to analyze the differences of variables between each cluster and summarize the characteristics of each cluster.

[0064] After obtaining the results of different cluster analysis models, the same operation is performed using the validation set data. First, the consistency between the model set and the validation set results is compared, and then the results of which clustering is more reasonable and more interpretable are compared and analyzed. Finally, one clustering result is determined as the result of the subtype analysis.

[0065] Step 130: Obtain patient data that meets the third preset condition from the patient data to form a second time series data, and dynamically predict the patient status transition in the next period based on the second time series data.

[0066] In this embodiment, the patient data under the third preset condition includes time series data of various indicators for critically ill patients from the start of invasive mechanical ventilation to its cessation. After obtaining the second event sequence data, each variable is divided every 4 hours, and the average value is taken as the data point; according to the patient's RSBI index, the patient's condition is divided into four groups from poor to good: <[105, +∞), [65, 105), [35, 65), (0, 35)>; a dynamic prediction model is established based on the selected indicator variables using algorithms such as Random Forest (RF) and XGBoost, and the patient's condition grouping for the next time period is dynamically predicted through the dynamic prediction model.

[0067] In this embodiment, the data is first cleaned, and the cleaning process includes:

[0068] The variables included in the dynamic prediction model are heart rate, respiratory rate, blood oxygen saturation, blood pressure, body temperature, blood urea nitrogen, creatinine, platelets, red blood cells, hematocrit, white blood cells, hemoglobin, blood glucose, chloride, potassium, sodium, calcium, pH, bicarbonate, partial pressure of oxygen, partial pressure of carbon dioxide, Glasgow Coma Scale score, tidal volume, positive end-expiratory pressure, and oxygen concentration. For the dynamic prediction model of patient state transitions, time-series data of various indicators are required. However, because the detection frequency of some variables is inconsistent—for example, the detection frequency of ventilator parameters is lower than that of vital signs—an average is first calculated for critically ill patients every four hours from the time they are put on a ventilator to reduce the missing data rate. If the missing data rate is >50%, the data is deleted. If a 4-hour interval is missing, iterative filling is performed using preceding and following data until no interval is missing.

[0069] Since the RSBI index has few records in the database, it needs to be calculated using formula (1):

[0070]

[0071] Generally, a lower RSBI index indicates better respiratory function in patients. Based on the critical values ​​for RSBI classification in relevant studies, this embodiment divides RSBI into the following four intervals (from Class 1 to Class 4): <[105, +∞), [65, 105), [35, 65), (0, 35)>.

[0072] In this embodiment, the data in the existing public dataset is divided into a training set (60%), an internal validation set (20%), and a test set (20%). After cleaning and preprocessing the data, the data for each time period, the RSBI status grouping for the current time period, and the patient subtype classification are input into the model. The RSBI status grouping for the next time period is used as the model's prediction target. Predictive models are built using algorithms such as logistic regression, random forest, and XGBoost. The toolkits used are Python-based Scikit-learn and XGBoost. The performance of the model is measured by the AUC (Area Under Curve) of the ROC (Receiver Operating Characteristic Curve). A higher AUC indicates better model performance. The model with the best performance is selected as the final prediction model.

[0073] Step 140: Construct a Markov decision model based on the patient subtype and the characteristics of each subtype, and input the patient state transition information into the Markov decision model to predict the optimal stopping strategy for each subtype.

[0074] In this embodiment, a Markov decision model can be constructed based on the patient subtype and the features of each subtype. The Markov decision model includes prediction logic such as state, action, reward, and probability transition. Then, the Markov decision model is trained using a heterogeneous policy Q-learning method. Specifically, as... Figure 4 As shown, this embodiment can be understood as follows:

[0075] First, based on the constructed subtypes of critically ill patients on ventilators, Markov decision processes are constructed for each subtype. Each Markov decision process includes state, action, reward, and transition probability. Details are as follows:

[0076] Time range: such as Figure 5 As shown, this embodiment defines a time period t as every 4 hours, where T is the upper limit of a pre-determined time range. Based on the fact that the ventilation time of the included cases was all over 24 hours, and that over 90% of patients had mechanical ventilation durations within 200 hours, this embodiment will start from the 24th hour of patient ventilation (t=0) and divide the time into 44 time periods (T=44). Figure 5 As shown, all patients will be weaned off the ventilator during the final period.

[0077] States: Based on existing standardized data and clinicians' experience, this embodiment selects the following six variables as patient states: respiratory rate (RR), heart rate (HR), pH value, positive end-expiratory pressure (PEEP), inhaled oxygen concentration (FiO2), and blood oxygen concentration (SpO2). By analyzing the distribution of each indicator in each subtype and dividing each indicator into several intervals, the size of the state space in the Markov decision process can be appropriately reduced, accelerating model operation and convergence.

[0078] Because this embodiment divides patients into four state groups according to the RSBI index and incorporates predictions of the patient's state in the next time period, decision-makers can not only observe the patient's current state group c t It can also know the next predicted state c. t+1 Therefore, the patient's state transition binary. <c t ,c t+1 The state will also be incorporated into the Markov decision process. The state space S is finite, as shown in equation (2), and the state consists of a vector with 8 variable dimensions:

[0079] s t =[RR,HR,pH,PEEP,FiO2,SpO2,c t ,c t+1 ] Formula (2)

[0080] Actions: For each time interval t, the action space A contains only two discontinuous actions a. t :a t =1 Stop mechanical ventilation and a t =0 Continue mechanical ventilation. At the final time T, a T =1.

[0081] Rewards: The reward function consists of three parts: The degree to which the action affects the stability of the patient's vital signs is measured; the greater the change, the smaller the return. The system measures the patient's condition after weaning from the ventilator. If the patient is successfully weaned from the ventilator, the result is a positive number; otherwise, it is a negative number. The duration of mechanical ventilation is measured, and the reward is deducted by 1 for each additional period, thereby shortening the patient's mechanical ventilation time. Therefore, for each state transition... t ,a t ,s t+1 >Returns t+1 As shown in formula (3):

[0082] ​

[0083] The return function of formula (3) can reflect both the degree of change in the patient's condition and the impact of ventilator withdrawal failure and prolonged ventilation time on the patient.

[0084] Transition Probability: Transition probability P(s) t+1 |s t ,a t The definition is that the patient is in state s. t After action a t Then transition to state s t+1 The probability of this is unknown in the settings of this embodiment.

[0085] As shown in Equation (4), the goal of reinforcement learning is to learn a policy π from a large amount of data to maximize the cumulative expected reward over time T:

[0086]

[0087] Here, γ is the discount factor, representing the relative weight of immediate and long-term returns, and its value is generally set between 0 and 1; in this embodiment, it will be set to 0.99. This embodiment will construct a Markov decision process for each patient subtype to find a personalized optimal ventilator weaning strategy that maximizes the expected value of the return function for each patient subtype.

[0088] After constructing the Markov decision model, this embodiment also needs to train it. Specifically, there are various existing reinforcement learning algorithms, such as Q-Learning, SARSA (State-action-reward-state-action), and Deep Q Network, each with its own advantages. Since this embodiment plans to include a large number of patient indicator variables, the number of patient states also increases, easily leading to the curse of dimensionality. Therefore, this embodiment proposes to stratify patients based on risk according to patient subtype and RSBI index, incorporating the predicted information of patient state transitions into the patient's state. This not only reduces the dimensionality of the variables, mitigating the negative impact of excessive dimensionality on the model, but also accelerates the model training speed. Since the patient's state transition probability is unknown, the agent needs to learn using a model-free value-based learning method. Because a large amount of action trajectories and empirical data already exists, this embodiment will employ an off-policy Q-learning method to learn the Q-function Q(s,a). This function, also known as the action value function, represents the expected cumulative reward obtained by taking a certain action in a given state. The goal of Q-learning is to approximate the optimal action value function Q. * (s,a)=max π Q π (s t ,a t ), searching for state s t Next, do a t The expected value that yields the best return. This embodiment will specifically apply and compare the Fitted Q Iteration algorithm and the algorithm based on the Deep Q Network (DQN) neural network to fit the Q function. In this embodiment, the data in the existing public dataset will be divided into a training set (60%), an internal validation set (20%), and a test set (20%), and a reinforcement learning model will be built using OpenAI's Gym toolkit.

[0089] During model training, the model is simultaneously validated and evaluated to determine whether it meets the requirements. Specifically, this embodiment uses 10-fold cross-validation to first perform internal validation on each model, and then performs external validation on a multi-center critical illness dataset. The expected value of the average Q function and the 95% confidence interval are calculated respectively. The higher the Q value, the better the model performance. The optimal strategy model is selected based on the Q value.

[0090] In summary, this embodiment adds steps for patient subtype analysis and prediction of disease progression based on patients' time-series data. This allows for the development of personalized ventilator weaning strategies for critically ill patients with different subtypes, assisting physicians in decision-making. It is more efficient and flexible than generalized methods. Furthermore, when the method of this embodiment is embedded into a hospital's electronic medical record system, it can obtain patient data in real time and provide suggestions on whether ventilation should be discontinued for physicians' reference, thus assisting physicians in decision-making and improving the utilization rate of medical resources, providing a new model for medical operation management.

[0091] This invention provides a system for generating a termination strategy for invasive mechanical ventilation, comprising:

[0092] The acquisition module is used to acquire data of invasive mechanical ventilation patients that meet the first preset requirements;

[0093] The first screening module is used to obtain patient data that meets the second preset conditions from the patient data to form a first time series average data.

[0094] The determination module is used to determine the patient subtype and the characteristics of each subtype based on the first time series average data;

[0095] The second screening module is used to obtain patient data that meets the third preset conditions from the patient data to form a second time series data;

[0096] The first prediction module is used to dynamically predict the patient's state transition in the next period based on the second time series data;

[0097] The module is used to construct a Markov decision model based on the patient subtype and the features of each subtype.

[0098] The second prediction module is used to input the patient's state transition information into the Markov decision model to predict the optimal stopping strategy for each subtype.

[0099] The content of the method embodiments of the present invention is applicable to the system embodiments. The specific functions implemented in the system embodiments are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0100] This invention provides a system for generating a termination strategy for invasive mechanical ventilation, comprising:

[0101] At least one memory for storing programs;

[0102] At least one processor is used to load the program for execution. Figure 1 The method for generating a termination strategy for invasive mechanical ventilation is shown.

[0103] The content of the method embodiments of the present invention is applicable to the system embodiments. The specific functions implemented in the system embodiments are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above methods.

[0104] This invention provides a storage medium storing a computer-executable program, which, when executed by a processor, is used to implement... Figure 1 The method for generating a termination strategy for invasive mechanical ventilation is shown.

[0105] This invention also provides a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to perform... Figure 1 The method for generating a termination strategy for invasive mechanical ventilation is shown.

[0106] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments, and various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention. Furthermore, the embodiments of the present invention and the features thereof can be combined with each other unless otherwise specified.

Claims

1. A method of generating a weaning strategy for invasive mechanical ventilation, characterized in that, The method comprises the following steps: obtaining invasive mechanical ventilation patient data meeting first preset requirements; obtaining patient data meeting second preset conditions from the patient data to form first time series average data; determining patient subtype types and characteristics of each of the subtype types according to the first time series average data; obtaining patient data meeting third preset conditions from the patient data to form second time series data; dynamically predicting patient state transition in the next period according to the second time series data; constructing a Markov decision model according to the patient subtype types and the characteristics of each of the subtype types; inputting the patient state transition into the Markov decision model to obtain an optimal stopping strategy for each subtype; wherein the invasive mechanical ventilation patient data comprises a time point at which the patient is subjected to invasive ventilation, patient demographic characteristics, patient basic characteristics, laboratory indexes, scores, respiratory parameters and outcome indexes; the patient basic characteristics comprise heart rate, respiratory rate, blood oxygen saturation, blood pressure and body temperature; the laboratory indexes comprise blood urea nitrogen, creatinine, platelets, red blood cells, hematocrit, white blood cells, hemoglobin, blood glucose, chlorine, potassium, sodium, calcium, pH value, bicarbonate, partial pressure of oxygen and partial pressure of carbon dioxide; the scores comprise Glasgow coma score; the respiratory parameters comprise tidal volume, positive end-expiratory pressure and oxygen concentration; and the outcome indexes comprise ICU mortality, hospital mortality, ICU admission duration, hospitalization duration, invasive mechanical ventilation duration, admission to ICU duration, ICU admission to ventilation initiation duration and ventilator weaning success rate; the obtaining of the invasive mechanical ventilation patient data meeting the first preset requirements comprises: obtaining a plurality of invasive ventilation patient data; extracting patient data with invasive ventilation duration greater than or equal to a first preset duration, patient age greater than or equal to a first preset age and extubation time greater than death time from the plurality of invasive ventilation patient data to form a first data set; obtaining patient data belonging to first-time invasive mechanical ventilation from the first data set to form a second data set; classifying the patient data in the second data set; after obtaining the second data set, judging the relationship between the patient data in the second data set and a first preset extubation state and a second preset extubation state; the first preset extubation state is that the patient does not need to start invasive ventilation again within 24 hours after ventilator weaning; and the second preset extubation state is that the patient needs to start invasive ventilation again or dies within 24 hours after ventilator weaning; the third preset condition patient data comprises time series data of each index of a severe patient from starting invasive mechanical ventilation to stopping; the construction of the Markov decision model according to the patient subtype types and the characteristics of each of the subtype types comprises: constructing a Markov decision model according to the patient subtype types and the characteristics of each of the subtype types, wherein the Markov decision model comprises state, action, reward and probability transition; and the state is specifically a patient state, comprising respiratory rate, heart rate, pH value, positive end-expiratory pressure, inhaled oxygen concentration and blood oxygen concentration. training the Markov decision model according to a Q-learning method of a heterogeneous strategy; the first time series average data is composed of patient data groups meeting a second preset condition from the patient data, including: extracting patient data belonging to 24 hours after starting ventilation from the patient data, and calculating the average value of each index in the patient data belonging to 24 hours after starting ventilation.

2. The method of claim 1, wherein, the first time series average data is composed of patient data groups meeting a second preset condition from the patient data, including: the first time series average data is composed of patient data groups meeting a second preset condition from the patient data, including: the first time series average data is composed of patient data groups meeting a second preset condition from the patient data, including:

3. A weaning strategy generation system for invasive mechanical ventilation, characterized in that the system is applied to the method of any one of claims 1-2, comprising: an acquisition module for acquiring invasive mechanical ventilation patient data meeting a first preset requirement; a first screening module for obtaining first time series average data composed of patient data meeting a second preset condition from the patient data; a determination module for determining patient subtype types and characteristics of each of the subtype types according to the first time series average data; a second screening module for obtaining second time series data composed of patient data meeting a third preset condition from the patient data; a first prediction module for dynamically predicting patient state transition in the next period according to the second time series data; a construction module for constructing a Markov decision model according to the patient subtype types and characteristics of each of the subtype types; a second prediction module for inputting the patient state transition into the Markov decision model to predict an optimal stopping strategy for each subtype.

4. An apparatus for generating a weaning strategy for invasive mechanical ventilation, characterized in that including: at least one memory for storing a program; at least one processor for loading the program to execute the invasive mechanical ventilation stopping strategy generation method of any one of claims 1-2.

5. A storage medium, characterized by wherein a computer executable program is stored, and the computer executable program is executed by a processor to implement the invasive mechanical ventilation stopping strategy generation method of any one of claims 1-2.

Images