A method for predicting oxygenation index based on continuous non-invasive parameters

Abstract

The present invention discloses a method for predicting the oxygenation index based on continuous non-invasive parameters, which relates to the field of oxygenation index prediction and includes the following steps: constructing a multi-center critical care medical information database; designing an adaptively adjustable time sliding window, combining the different resolutions of each parameter in the database, obtaining non-invasive parameter data of the patient in a continuous time window related to the oxygenation index, and preprocessing the data; using a multivariate fusion filtering feature selection algorithm to obtain the correlation between the continuous non-invasive parameters and the oxygenation index; designing an oxygenation index prediction model to predict the patient's oxygenation index; optimizing the oxygenation index prediction model, and using a multi-center critical care medical information database to verify the prediction results, thereby completing the prediction of the oxygenation index based on continuous non-invasive parameters. The present invention solves the problems of the existing technology in that the parameters used are relatively simple, the model does not consider the influence of continuous physiological parameters and time on the disease, the sample size is limited, and the prediction accuracy is low.

Description

Technical Field

[0001] The present invention relates to the field of oxygenation index prediction, and in particular to an oxygenation index prediction method based on continuous non-invasive parameters. Background Art

[0002] The oxygenation index (P / F) is a goal in respiratory therapy and an important index that ensures that organs and tissues receive sufficient oxygen for oxygenation and energy production. The oxygenation index is of great significance for the diagnosis and treatment of diseases. On the one hand, the oxygenation index is the only quantitative diagnostic indicator in the diagnostic basis, providing important diagnostic evidence for medical personnel during the diagnosis process; on the other hand, the oxygenation index provides an important reference for supportive treatment. In clinical work, the patient's oxygenation index should be monitored in real time to determine the patient's oxygenation level and differentiate the severity. Appropriate treatment measures should be taken for patients with different disease severity levels, and personalized treatment plans should be formulated.

[0003] Existing oxygenation index prediction methods mainly use statistical learning methods and traditional machine learning algorithms, and the parameters used are relatively simple. The model does not consider the impact of continuous physiological parameters and time on the disease. The sample size is limited and the prediction accuracy is low. Summary of the Invention

[0004] In response to the above-mentioned deficiencies in the prior art, the present invention provides an oxygenation index prediction method based on continuous non-invasive parameters, which solves the problems of the prior art using relatively single parameters, the model not considering the impact of continuous physiological parameters and time on the disease, the limited sample size, and the low prediction accuracy.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: a method for predicting oxygenation index based on continuous non-invasive parameters, comprising the following steps:

[0006] S1: Use the localized critical care medicine information database to build a multi-center critical care medicine information database;

[0007] S2: Based on a multi-center critical care medicine information database, an adaptive time sliding window is designed. Utilizing a sliding window matching mechanism and combining the different resolutions of various parameters in the database, non-invasive parameter data related to the patient's oxygenation index within a continuous time window is obtained and preprocessed.

[0008] S3: Based on the preprocessed data, the correlation between continuous noninvasive parameters and oxygenation index was obtained using a multivariate fusion filtering feature selection algorithm;

[0009] S4: Based on the correlation between continuous non-invasive parameters and oxygenation index, an oxygenation index prediction model is designed to predict the patient's oxygenation index;

[0010] S5: Adaptively adjust the time window range according to the prediction results, adjust the enrollment data by increasing or decreasing the time window length, optimize the oxygenation index prediction model, and use the multi-center critical care medicine information database to verify the prediction results to complete the oxygenation index prediction based on continuous non-invasive parameters.

[0011] The beneficial effect of the above scheme is: the present invention obtains the correlation between multiple continuous non-invasive physiological parameters and the oxygenation index (P / F value) within a time interval, and designs an oxygenation index prediction model by introducing a time variable, while adopting a sliding window matching mechanism for data mining. The present invention adopts this sliding window matching method to select different window lengths according to different sampling frequencies to match the patient's physiological parameters, ensure data integrity, and realize continuous monitoring of the oxygenation index. This solves the problems of the existing technology that the parameters used are relatively single, the model does not consider the impact of time on the disease, the sample size is limited, and the prediction accuracy is low.

[0012] Furthermore, S1 established a localized critical care medicine information database through the clinical critical care medicine database MIMIC-IV provided by PhysioNet for model training. At the same time, it used the local critical care medicine department to build a local critical care medicine database to achieve multi-center data set verification.

[0013] The beneficial effect of the above further scheme is: by establishing a localized, multi-center critical care medicine information database through the above method, the problem of insufficient sample size in traditional research methods and the fact that most of them are single-center clinical retrospective studies is solved.

[0014] Furthermore, when S2 obtains the non-invasive parameter data of the patient in a continuous time window related to the oxygenation index, the following criteria are included:

[0015] (1) Patients aged 16 years or older;

[0016] (2) the patient's ICU stay was greater than or equal to 48 hours;

[0017] (3) the patient had undergone mechanical ventilation;

[0018] (4) patients had blood gas analysis data;

[0019] (5) The patient's minimum oxygenation index in the previous seven days is less than or equal to 300.

[0020] The beneficial effect of the above further scheme is that during the patient screening process, by setting the above standards, the accuracy of the oxygenation index prediction of the enrolled patients is ensured, while the accuracy of the prediction model is improved.

[0021] Furthermore, the data is preprocessed in S2, including the following steps:

[0022] S2-1: For the critical care medicine information database, normal ranges of various physiological parameters were set in combination with clinical knowledge during the data extraction process, and abnormal values ​​outside the normal range were excluded;

[0023] S2-2: Based on the data excluding outliers, the K-NN model is used to interpolate the randomly missing data, and the Random Forest algorithm is used to fill the missing data of strongly correlated parameters;

[0024] S2-3: Use the adaptive synthetic oversampling ADASYN algorithm to oversample the completed data set to keep the data samples balanced and complete data preprocessing.

[0025] The beneficial effect of the above further scheme is: since there are many problems with directly extracted data, in order to establish the subsequent model, the data is preprocessed, abnormal data is excluded, missing data is interpolated, and unbalanced data is oversampled to ensure that the data samples remain balanced.

[0026] Furthermore, S3 includes the following sub-steps:

[0027] S3-1: Based on the preprocessed data, the importance of feature values ​​is evaluated using the multivariate fusion filtering feature selection algorithm;

[0028] S3-2: Based on the importance of eigenvalues, a primary-secondary accumulation approach was adopted, using the logistic regression algorithm, neural network algorithm, AdaBoost algorithm, and XGBoost algorithm, combined with the cross-validation method, to select the optimal feature subset from a large number of eigenvalues, and then select the minimum feature subset from the optimal feature subset to obtain the correlation between continuous noninvasive parameters and oxygenation index.

[0029] The beneficial effect of the above further scheme is that after preprocessing, the data set may include a large number of eigenvalues, and the importance of the eigenvalues ​​relative to the outcome variables may vary greatly. The feature selection algorithm is used to conduct preliminary feature importance ranking work. The traditional filtering feature selection algorithm usually uses only a single algorithm to evaluate the correlation between the eigenvalue and the outcome variable. The evaluation method is too single, and only one indicator among the distance vector, dependency, and mutual information is considered to give the importance score, which is prone to errors. In order to improve the shortcomings of the original algorithm, the present invention uses a multi-fusion filtering feature selection algorithm to evaluate the importance of the eigenvalue. On the basis of the traditional filtering algorithm, from the three dimensions of distance vector, dependency, and mutual information, three corresponding evaluation algorithms are selected: Relief-F, Chi-squared coefficient, and mutual information. Finally, the results are weighted to calculate the final eigenvalue importance score. This method can comprehensively consider the multi-dimensional information between features and between features and outcome variables, and comprehensively reflect the importance of eigenvalues.

[0030] Furthermore, the importance calculation formula of the eigenvalue in S3-1 is:

[0031]

[0032] Among them, R j is the importance score of the feature value, n is the number of feature selection methods, F i,j F is the importance score of the jth eigenvalue under the i-th feature selection method, i,j,max is the maximum value of the importance score of the j-th eigenvalue under the i-th feature selection method, F i,j,min It is the minimum importance score of the j-th eigenvalue under the i-th feature selection method.

[0033] The beneficial effect of the above further solution is: the importance score of the eigenvalue is calculated by the above formula to complete the importance ranking of the eigenvalue.

[0034] Furthermore, the multivariate fusion filtering feature selection algorithm includes the ReliefF algorithm, the Chi-squared coefficient, and the mutual information. The formula of the ReliefF algorithm is:

[0035]

[0036] in, is the component of the correlation statistic corresponding to eigenvalue j, For the sample, For the first type of sample m The nearest neighbor sample of For the first type of sample m The nearest neighbor sample, θ l is the proportion of the second type of sample l in the data set, h is the guessed nearest neighbor, and diff(·) is the derivative;

[0037] The calculation formula of the Chi-squared coefficient is:

[0038]

[0039] Among them, X 2 is the chi-square value, A pq is the distance vector between the pth interval and the qth class, E pq A pq The expected frequency of , k is the total sample size;

[0040] The mutual information I(X,Y) calculation formula is:

[0041]

[0042] Among them, Pxy (x,y) is the joint distribution of variables x and y, P x (x) is the marginal distribution of x, P y (y) is the marginal distribution of y.

[0043] The beneficial effect of the above further scheme is that the filtering method only evaluates the importance of the feature value through its own algorithm, generally evaluating the importance of the feature through three dimensions: distance vector, dependency, and mutual information. It does not require feedback adjustment of the subsequent classifier results, and can also well reflect the importance of the feature. Its advantages are fast operation speed and small computing resource usage, which is suitable for early exploratory research. In order to comprehensively consider the multi-dimensional information between features and between features and outcome variables, and comprehensively reflect the importance of feature values, the present invention selects three filtering algorithms: Relief-F, Chi-squared coefficient, and mutual information.

[0044] Furthermore, S4 includes the following sub-steps:

[0045] S4-1: Normalize continuous noninvasive parameters according to the minimum feature subset and group the normalized datasets into chronological groups;

[0046] S4-2: Build an LSTM neural network model based on the grouping results and set the initial parameters of the LSTM neural network model;

[0047] S4-3: Divide the dataset into a training set and a test set, use the training set to train the LSTM neural network model, adjust the parameters, calculate the loss function of the LSTM neural network model using the forward feedback process, and use the Adam algorithm to determine the iterative update method of the weight parameters to complete the optimization of the LSTM neural network model;

[0048] S4-4: Input the test set into the optimized LSTM neural network model for prediction and output the prediction results;

[0049] S4-5: Based on the prediction results, the LSTM deep learning algorithm is used to construct an oxygenation index prediction model to predict the patient's oxygenation index level.

[0050] The beneficial effect of the above further scheme is: the present invention builds an oxygenation index prediction model by utilizing the advantages of LSTM deep learning algorithm in processing time series data, thereby achieving accurate prediction of the oxygenation index. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Flowchart of a method for predicting oxygenation index based on continuous noninvasive parameters.

[0052] Figure 2Schematic diagram of adaptively adjusting the time window range.

[0053] Figure 3 Schematic diagram for selecting a minimum feature subset using a machine learning algorithm.

[0054] Figure 4 This is a diagram of importance scoring. DETAILED DESCRIPTION

[0055] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0056] like Figure 1 As shown, a method for predicting oxygenation index based on continuous non-invasive parameters includes the following steps:

[0057] S1: Use the localized critical care medicine information database to build a multi-center critical care medicine information database;

[0058] S2: Based on a multi-center critical care medicine information database, an adaptive time sliding window is designed. Utilizing a sliding window matching mechanism and combining the different resolutions of various parameters in the database, non-invasive parameter data related to the patient's oxygenation index within a continuous time window is obtained and preprocessed.

[0059] S3: Based on the preprocessed data, the correlation between continuous noninvasive parameters and oxygenation index was obtained using a multivariate fusion filtering feature selection algorithm;

[0060] S4: Based on the correlation between continuous non-invasive parameters and oxygenation index, an oxygenation index prediction model is designed to predict the patient's oxygenation index;

[0061] S5: Adaptively adjust the time window range according to the prediction results, adjust the enrollment data by increasing or decreasing the time window length, optimize the oxygenation index prediction model, and use the multi-center critical care medicine information database to verify the prediction results to complete the oxygenation index prediction based on continuous non-invasive parameters.

[0062] In S1, a localized critical care medicine information database was established using the MIMIC-IV clinical critical care medicine database provided by PhysioNet for model training. At the same time, a local critical care medicine database was constructed using the local critical care medicine department to achieve multi-center data set verification.

[0063] In one embodiment of the present invention, the hospital's existing VMware virtualization platform is used to create a virtualized server platform, providing a localized software and hardware resource platform for the MIMIC-IV database, deploying a PostgreSQL distributed data management platform on the virtualization platform, and using a relational database scripting language to write a database creation script (including database schema, database table structure, and database index) based on the MIMIC-IV data structure, thereby improving data access speed and algorithm training efficiency. MIMIC-IV can be used to carry out data screening and model training related work for oxygenation index. Based on the ICU intensive care system of the Military Region General Hospital, according to the ICD-10 coding of ARDS patients, the data collection scope and rules are formulated, the Kettle tool is used to design a data extraction workflow, and an SQL data extraction script is written to achieve functions such as extraction, quality inspection, data conversion, and data filtering of medical data, and to build the hospital's own critical care medicine database for subsequent model verification.

[0064] When obtaining noninvasive parameter data related to the patient's oxygenation index within a continuous time window in S2, the following criteria are included:

[0065] (1) Patients aged 16 years or older;

[0066] (2) the patient's ICU stay was greater than or equal to 48 hours;

[0067] (3) the patient had undergone mechanical ventilation;

[0068] (4) patients had blood gas analysis data;

[0069] (5) The patient's minimum oxygenation index in the previous seven days is less than or equal to 300.

[0070] In S2, a sliding window matching mechanism is used for data extraction. Combined with the different resolutions of various parameters in the critical care medical information database, different time windows are formulated to obtain continuous non-invasive parameters related to the oxygenation index.

[0071] In one embodiment of the present invention, SpO2 is easy to obtain in the ICU and is a continuous monitoring parameter. Its sampling resolution in the database is generally once an hour, but sometimes there will be a time mismatch. Therefore, the time window of SpO2 is set to within two hours before blood gas analysis. The sampling frequency of parameters such as heart rate, body temperature, respiratory rate, tidal volume, etc. is generally once every 2 to 4 hours, so the time window is set to within the first four hours. Figure 2 shown.

[0072] In S2, data is preprocessed, including the following steps:

[0073] S2-1: For the critical care medicine information database, normal ranges of various physiological parameters were set in combination with clinical knowledge during the data extraction process, and abnormal values ​​outside the normal range were excluded;

[0074] S2-2: Based on the data excluding outliers, the K-NN model is used to interpolate the randomly missing data, and the Random Forest algorithm is used to fill the missing data of strongly correlated parameters;

[0075] S2-3: Use the adaptive synthetic oversampling ADASYN algorithm to oversample the completed data set to keep the data samples balanced and complete data preprocessing.

[0076] An unbalanced dataset can pose numerous challenges for model training and final results. From a model training perspective, if a class has very few samples, the dataset is unbalanced. This class with less data will provide much less relevant information, and the final results may be more biased towards the class with more data, leading to significant model bias. This paper uses the Adaptive Synthetic Oversampling (ADASYN) algorithm to oversample the dataset, ensuring balanced data samples.

[0077] S3 includes the following sub-steps:

[0078] S3-1: Based on the preprocessed data, the importance of feature values ​​is evaluated using the multivariate fusion filtering feature selection algorithm;

[0079] S3-2: Based on the importance of eigenvalues, the primary and secondary accumulation method is used, and the logistic regression algorithm, neural network algorithm, AdaBoost algorithm and XGBoost algorithm are used in combination with the cross-validation method to select the optimal feature subset from a large number of eigenvalues, and the minimum feature subset is selected from the optimal feature subset to obtain the correlation between continuous non-invasive parameters and oxygenation index, such as Figure 3 shown.

[0080] In order to ensure the stability of feature selection results, this paper designs a multivariate fusion method based on the filtering feature selection algorithm, uses Relief-F, chi-square test, and MIFS algorithm to calculate the importance score of feature values ​​relative to the outcome variable, normalizes the score results and sums them up to give the final importance score, such as Figure 4 shown.

[0081] To address the issues presented in this paper, appropriate hyperparameter settings are selected to improve the speed and accuracy of the subset search algorithm. Different machine learning algorithms require specific approaches for hyperparameter optimization. Logistic regression hyperparameter adjustment primarily involves regularization selection, solver function selection, and iteration number selection. Neural network hyperparameter adjustment involves activation function and solver function selection. AdaBoost hyperparameter adjustment primarily involves algorithm implementation and learning rate selection. XGBoost hyperparameter adjustment primarily involves learning rate selection and tree depth selection.

[0082] The formula for calculating the importance of eigenvalues ​​in S3-1 is:

[0083]

[0084] Among them, R j is the importance score of the feature value, n is the number of feature selection methods, F i,j F is the importance score of the jth eigenvalue under the i-th feature selection method, i,j,max is the maximum value of the importance score of the j-th eigenvalue under the i-th feature selection method, F i,j,min It is the minimum importance score of the j-th eigenvalue under the i-th feature selection method.

[0085] The multivariate fusion filtering feature selection algorithm includes the ReliefF algorithm, the Chi-squared coefficient, and mutual information. The formula of the ReliefF algorithm is:

[0086]

[0087] in, is the component of the correlation statistic corresponding to eigenvalue j, For the sample, For the first type of sample m The nearest neighbor sample of For the first type of sample m The nearest neighbor sample, θ l is the proportion of the second type of sample l in the data set, h is the guessed nearest neighbor, and diff(·) is the derivative;

[0088] The calculation formula of the Chi-squared coefficient is:

[0089]

[0090] Among them, X 2 is the chi-square value, A pq is the distance vector between the pth interval and the qth class, E pq A pq The expected frequency of , k is the total sample size;

[0091] The mutual information I(X,Y) calculation formula is:

[0092]

[0093] Among them, P xy (x,y) is the joint distribution of variables x and y, P x (x) is the marginal distribution of x, P y (y) is the marginal distribution of y.

[0094] S4 includes the following sub-steps:

[0095] S4-1: Normalize the continuous non-invasive parameters according to the minimum feature subset, and group the normalized data set in chronological order. In this embodiment, the data set is divided into four groups: 2h, 4h, 6h, and 12h;

[0096] S4-2: Build an LSTM neural network model based on the grouping results and set the initial parameters of the LSTM neural network model;

[0097] S4-3: Divide the dataset into a training set and a test set, use the training set to train the LSTM neural network model, adjust the parameters, calculate the loss function of the LSTM neural network model using the forward feedback process, and use the Adam algorithm to determine the iterative update method of the weight parameters to complete the optimization of the LSTM neural network model;

[0098] S4-4: Input the test set into the optimized LSTM neural network model for prediction and output the prediction results;

[0099] S4-5: Based on the prediction results, the LSTM deep learning algorithm is used to construct an oxygenation index prediction model to predict the patient's oxygenation index level.

[0100] In one embodiment of the present invention, to comprehensively validate the oxygenation index prediction method designed in the present invention and compare the advantages and disadvantages of different methods, a comprehensive comparison of the performance of traditional machine learning algorithms, ensemble learning algorithms, and LSTM deep learning algorithms in predicting oxygenation index was conducted using performance evaluation indicators such as sensitivity, specificity, accuracy, negative predictive value, positive predictive value, AUC, receiver operating characteristic (ROC) curve, and balanced error rate. The model was also validated using different MIMIC-IV datasets and a critical care dataset from a military general hospital to study the impact of different datasets on model performance.

[0101] This invention uses continuous noninvasive patient physiological parameters (basic patient information, ventilator measurement parameters, ventilator setting parameters, and monitor-related measurement parameters). These parameters are basic and easily accessible. This solution no longer relies on the patient's physiological parameters at a specific time point to identify the P / F level at the current time point. Instead, it uses a sliding time window to predict the P / F level using all these parameters within that time window. With traditional noninvasive physiological parameters, for example, if the physiological parameters at 12:00 PM are obtained and a P / F value is measured at that time, model training can be performed. However, with this continuous noninvasive physiological parameter approach, if the P / F value at 12:00 PM is known and the physiological parameters for the two-hour period from 10:00 AM to 12:00 PM are also obtained, the LSTM algorithm can be used to predict the P / F value at 12:00 PM. Because temporal variations in the patient's physiological parameters can affect model accuracy, this solution introduces a time variable to improve prediction accuracy.

[0102] This solution can continuously monitor the patient's oxygenation index level, achieving non-invasive and continuous monitoring of the oxygenation index. The oxygenation index can assist medical personnel in quickly diagnosing and treating diseases. For example, for ARDS, when the predicted result is an oxygenation index <= 100 (severe ARDS), 100 < oxygenation index <= 200 (moderate ARDS), and 200 < oxygenation index <= 300 (mild ARDS), the present invention uses a sliding window matching mechanism for data mining, focusing on other physiological parameters corresponding to the time of blood gas analysis. Since the sampling frequencies of various physiological parameters of patients in the ICU vary greatly, it is impossible to ensure that other physiological parameters are recorded at the time of blood gas analysis. Therefore, the present invention adopts this sliding window matching method, selecting different window lengths according to different sampling frequencies to match the patient's physiological parameters and ensure data integrity. At the same time, using the patient's continuous non-invasive physiological parameters, the time variable is introduced into the model establishment, and the LSTM algorithm in the recurrent neural network is used to construct a prediction model, improving the model accuracy and enhancing the model generalization ability.

[0103] The present invention solves the problems of the existing technology, such as the relatively simple parameters used, the model not considering the impact of continuous physiological parameters and time on the disease, the limited sample size, and the low prediction accuracy. It realizes non-invasive and continuous monitoring of oxygenation index levels, and assists medical personnel in quickly diagnosing and treating ARDS.

[0104] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and it should be understood that the scope of protection of the present invention is not limited to such specific descriptions and embodiments. Those skilled in the art can make various other specific variations and combinations based on the technical teachings disclosed in the present invention without departing from the essence of the present invention, and such variations and combinations are still within the scope of protection of the invention.

Claims

1. A method for predicting oxygenation index based on continuous non-invasive parameters, characterized in that: The following steps are involved: S1: Use the localized critical care medicine information database to build a multi-center critical care medicine information database; In S1, a localized critical care medicine information database is established using the MIMIC-IV clinical critical care medicine database provided by PhysioNet for model training. At the same time, a local critical care medicine database is constructed using the local critical care medicine department to achieve multi-center data set verification. When establishing a localized critical care medicine information database using the MIMIC-IV clinical critical care medicine database provided by PhysioNet, the server was built using the VMware virtualization platform, the PostgreSQL distributed data management system was deployed, and a targeted database creation script was written, including the database schema, database table structure, and database index; S2: Based on a multi-center critical care medicine information database, an adaptive time sliding window is designed. Utilizing a sliding window matching mechanism and combining the different resolutions of various parameters in the database, non-invasive parameter data related to the patient's oxygenation index within a continuous time window is obtained, and the data is preprocessed. The sliding window matching mechanism is used in combination with the different resolutions of various parameters in the database to obtain non-invasive parameter data of the patient within a continuous time window related to the oxygenation index, specifically including: SpO2 sampling resolution of once every hour, and the time window is set to within two hours before blood gas analysis; heart rate, body temperature, respiratory rate and tidal volume sampling resolution of once every 2-4 hours, and the time window is set to within four hours before blood gas analysis; S3: Based on the preprocessed data, the correlation between continuous noninvasive parameters and oxygenation index was obtained using a multivariate fusion filtering feature selection algorithm; S4: Based on the correlation between continuous non-invasive parameters and oxygenation index, an oxygenation index prediction model is designed to predict the patient's oxygenation index; The S4 includes the following sub-steps: S4-1: Normalize continuous noninvasive parameters according to the minimum feature subset and group the normalized datasets into chronological groups; S4-2: Build an LSTM neural network model based on the grouping results and set the initial parameters of the LSTM neural network model; S4-3: Divide the dataset into a training set and a test set, use the training set to train the LSTM neural network model, adjust the parameters, calculate the loss function of the LSTM neural network model using the forward feedback process, and use the Adam algorithm to determine the iterative update method of the weight parameters to complete the optimization of the LSTM neural network model; S4-4: Input the test set into the optimized LSTM neural network model for prediction and output the prediction results; S4-5: Based on the prediction results, the LSTM deep learning algorithm is used to build an oxygenation index prediction model to predict the patient's oxygenation index level; S5: Adaptively adjust the time window range according to the prediction results, adjust the enrollment data by increasing or decreasing the time window length, optimize the oxygenation index prediction model, and use the multi-center critical care medicine information database to verify the prediction results to complete the oxygenation index prediction based on continuous non-invasive parameters.

2. The method for predicting oxygenation index based on continuous non-invasive parameters according to claim 1, characterized in that: When obtaining non-invasive parameter data of the patient in a continuous time window related to the oxygenation index in S2, the following criteria are included: (1) The patient is 16 years old or older; (2) The patient's stay in the ICU is greater than or equal to 48 hours; (3) The patient had undergone mechanical ventilation; (4) The patient has blood gas analysis data; (5) The patient's minimum oxygenation index in the previous seven days is less than or equal to 300.

3. The method for predicting oxygenation index based on continuous non-invasive parameters according to claim 1, characterized in that: The data is preprocessed in S2, including the following steps: S2-1: For the critical care medicine information database, normal ranges of various physiological parameters were set in combination with clinical knowledge during the data extraction process, and abnormal values ​​outside the normal range were excluded; S2-2: Based on the data excluding outliers, the K-NN model is used to interpolate the randomly missing data, and the Random Forest algorithm is used to fill the missing data of strongly correlated parameters; S2-3: Use the adaptive synthetic oversampling ADASYN algorithm to oversample the completed data set to keep the data samples balanced and complete data preprocessing.

4. The method for predicting oxygenation index based on continuous non-invasive parameters according to claim 1, characterized in that: The S3 includes the following sub-steps: S3-1: Based on the preprocessed data, the importance of feature values ​​is evaluated using the multivariate fusion filtering feature selection algorithm; S3-2: Based on the importance of eigenvalues, a primary-secondary accumulation approach was adopted, using the logistic regression algorithm, neural network algorithm, AdaBoost algorithm, and XGBoost algorithm, combined with the cross-validation method, to select the optimal feature subset from the eigenvalues, and the minimum feature subset from the optimal feature subset to obtain the correlation between continuous noninvasive parameters and oxygenation index.

5. The method for predicting oxygenation index based on continuous non-invasive parameters according to claim 4, characterized in that: The calculation formula for the importance of the eigenvalue in S3-1 is: in, Score the importance of the eigenvalues, is the number of feature selection methods, For the The eigenvalue is Importance score under the feature selection method, For the The eigenvalue is The maximum value of the importance score under the feature selection method, For the The eigenvalue is The minimum importance score under the feature selection method.

6. The method for predicting oxygenation index based on continuous non-invasive parameters according to claim 5, characterized in that: The multivariate fusion filtering feature selection algorithm includes the ReliefF algorithm, Chi-squared coefficient, and mutual information. The formula of the ReliefF algorithm is: in, The correlation statistic corresponds to the eigenvalue The weight, For the sample, For the first type of sample middle The nearest neighbor sample of For the first type of sample Beyond The nearest neighbor sample of For the second type of sample The proportion in the dataset, To seek guidance; The calculation formula of the Chi-squared coefficient is: in, is the chi-square value, For the interval and The distance vector between the classes, for The expected frequency, is the total sample size; The mutual information The calculation formula is: in, For variables and The joint distribution of for The marginal distribution of for The marginal distribution of .

Images