Machine learning-based hospitalization resource allocation method under weekend effect

Abstract

The invention provides a machine learning-based hospitalization resource allocation method under a weekend effect, and the method employs a machine learning method to construct a hospitalization patient leaving mode prediction framework under the weekend effect, thereby completing the allocation of hospitalization resources. Firstly, a random forest algorithm is utilized to identify key factors influencing a hospital leaving mode; exploring a causal relationship among the key factors by using a Bayesian causal network and a causal forest model; analyzing the time sequence data by utilizing a vector autoregression model, capturing a dynamic relationship among variables, measuring the similarity of admission and operation processes of different patients by applying a dynamic time warping technology, and generating time sequence similarity features; the key features, the causal relationship features and the time sequence features are combined and input into the long short-term memory network model, the leaving-hospital mode of the patient is output, the long short-term memory network model can process complex time sequence data, the prediction accuracy is improved, and a basis is provided for completing hospitalization resource allocation.

Description

Technical Field

[0001] The present invention relates to machine learning technology, and in particular to inpatient resource allocation technology based on machine learning. Background Art

[0002] The weekend effect is caused by differences in the severity of patients' illness, treatment plans, and allocation of medical resources during non-working hours. At present, the weekend effect is widespread due to medical system policies at home and abroad, but there are still many unexplored areas in existing research, especially the specific impact of the weekend effect on specific diseases, patient groups, or medical services. Although traditional epidemiological statistical analysis methods and meta-analysis have been used to study the weekend effect, they are still insufficient in clarifying the causes and the utilization efficiency of medical resources is also low. With the in-depth understanding of the weekend effect and its potential causes, traditional statistical analysis methods may not be able to fully capture the dynamic changes and causal relationships behind this complex phenomenon. The weekend effect involves a variety of complex factors, including changes in patients' illness, resource allocation in hospitals, and the working status of medical staff, which are significantly heterogeneous in time and space. In order to more comprehensively understand and predict the impact of the weekend effect on the allocation of inpatient resources, more advanced analysis methods are needed.

[0003] Machine learning and deep learning technologies have shown great potential in the field of medical data analysis. These technologies can not only process large amounts of complex medical data, but also automatically capture potential patterns and causal relationships in the data through model training. However, existing studies have mostly focused on the impact of specific diseases or single factors, and lack systematic research on the comprehensive issue of the weekend effect. Using machine learning technology, combined with causal inference and time series analysis methods, we can more accurately identify the key factors that affect the clinical outcomes of hospitalized patients and their interactions, thereby providing more accurate predictions and more instructive support for the allocation of inpatient resources. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method for improving the allocation of inpatient resources by predicting the patient's discharge pattern, in view of the unclear impact of the weekend effect on the clinical outcomes of inpatients.

[0005] The technical solution adopted by the present invention to solve the above technical problems is a method for allocating hospitalization resources under the weekend effect based on machine learning, comprising the steps of:

[0006] Data collection and preprocessing steps: Inpatient data were collected from the medical service database in date order, and the date variables in the inpatient data were marked as working days and non-working days; the inpatient data included the patient's personal information, data during admission, data during surgery, and data during discharge;

[0007] Key feature selection step: Use the random forest algorithm to construct a decision tree to evaluate the importance of each variable in the inpatient data, and identify the key variables that affect the discharge mode as key features;

[0008] Causal feature selection step: Based on the selected key feature - discharge mode, a Bayesian causal network is constructed to explore the causal relationship between the key features including the weekend effect and the discharge mode. The nonlinear relationship and interaction between the key features of the causal forest model based on the decision tree are used to obtain the causal relationship features for predicting the discharge mode of inpatients under the weekend effect.

[0009] Steps for generating time series features: Use the time series data consisting of the data of the hospitalized patients during the admission period, the data of the surgery period, and the data of the discharge period to train the vector autoregression model. The vector autoregression model learns to capture the dynamic change relationship of each variable at each time point during the admission period, the surgery period, and the discharge period. The dynamic time warping technology is used to measure the similarity of the admission and surgery processes of different patients. The trained vector autoregression model is used to receive the input time series data of the hospitalized patients during the admission period and the surgery period, calculate the similarity of each hospitalized patient in the time series, and output a set of time series features of the admission-surgery-discharge mode corresponding to each patient;

[0010] Discharge mode prediction steps: merge key features, causal relationship features and time series features, preprocess the merged features, and then input the preprocessed features into the long short-term memory network. The long short-term memory network outputs the time series features of each time step to the fully connected layer; the fully connected layer outputs the predicted discharge mode for each time step based on the input time series features;

[0011] Inpatient resource allocation steps: Inpatient resources are allocated based on predicted discharge patterns.

[0012] The present invention constructs a prediction scheme for discharge mode by comprehensively applying random forest algorithm, Bayesian causal network, causal forest, vector autoregression VAR model, dynamic time warping DTW technology and long short-term memory network LSTM model. The scheme performs well in feature selection, causal relationship analysis and time series feature extraction, significantly improving the accuracy, precision, recall rate and F1 score of the prediction, providing a reliable basis for the subsequent allocation of inpatient resources.

[0013] The beneficial effect of the present invention is that it shows excellent effectiveness in predicting the patient's discharge method and can effectively extract and identify key features from a large amount of clinical data. Not only does it have a more accurate prediction capability for discharge methods, but it can also quickly adapt to new data situations. It maintains high accuracy when processing unknown data in the verification phase, showing good adaptability and generalization capabilities. On a variety of inpatient data, the present invention can maintain accurate and stable prediction capabilities, effectively improve the efficiency and quality of medical services, especially in response to the weekend effect, reduce the risk of predicting patient discharge methods, support more scientific clinical decision-making and medical management, reduce the waste of medical resources, and reduce costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 Flowchart for predicting hospital discharge patterns based on the weekend effect constructed using a machine learning approach. DETAILED DESCRIPTION

[0015] The present invention uses the admission date of inpatients as the benchmark, the surgery date as the intermediate time node, and the discharge mode as the end point to construct the clinical pathway for each patient, that is, to construct the specific medical treatment items that the patient needs to receive every day. In order to implement the method of the present invention, a novel model for predicting the patient's discharge mode is proposed. The model comprehensively utilizes the random forest algorithm, Bayesian causal network, causal forest, vector autoregression model, dynamic time warping technology and long short-term memory network model, and specifically includes the following steps:

[0016] Data collection and preprocessing steps: Clinical data of inpatients were collected from the medical service database in date order, including relevant data such as admission time, surgery time and discharge mode. The date data were marked as working days and non-working days. Non-working days included weekends and holidays to provide the time characteristics of weekend effect for subsequent analysis.

[0017] Feature selection and key factor identification steps: The importance of each variable in the inpatient data is evaluated by constructing multiple decision trees using the random forest algorithm. By analyzing various variables of the patients, such as admission date, surgery date, and admission method, key variables that affect discharge methods are identified and selected as features.

[0018] Causal relationship analysis steps: Based on the selected key feature - discharge mode, a Bayesian causal network is constructed to explore the causal relationship between the features affected by the weekend effect and the discharge mode. The causal forest model based on decision trees is used to mine the nonlinear relationship and interaction between these features, and the features that have an important impact on the discharge mode under the weekend effect are determined, such as admission date, admission route and surgery date, and the causal relationship features that are combined to accurately predict the discharge mode of inpatients under the weekend effect. The features affected by the weekend effect can be variables such as admission date, admission route, admission type, admission situation, case classification, surgery date and elective surgery.

[0019] Time series feature step: Use the vector autoregression model to train the multidimensional time series data of hospitalized patients. The surgical period is used as the mediating variable between the admission period and the discharge mode to capture the dynamic relationship between the variables at each time point in the admission period, surgical period, and discharge period. Through multivariate regression analysis, the changes in each variable are predicted. Dynamic time warping technology is used to measure the similarity of the time series of different patients during admission and surgery, and a set of similarity features of the admission-surgery time series is generated for each patient to reflect the degree of similarity of different patients in the time series. The dynamic change features of the time series generated by the vector autoregression model and the similarity features of the time series generated by the dynamic time warping technology are integrated to form the final time series features.

[0020] Discharge mode prediction steps: First, merge key variable features, causal relationship features, and time series features to ensure that all feature dimensions are in consistent and complete formats. Then, align the data in each feature according to the time of admission, surgery, and discharge to ensure that the data at all time points are correctly matched. Next, standardize the feature data and unify the dimensions of the features to ensure that the data is compared and analyzed on the same scale. After that, input the processed feature data into the long short-term memory network, which can process complex time series data and capture the dependencies between time steps. The long short-term memory network outputs the time series features of each time step to the fully connected layer for further processing. The fully connected layer predicts the discharge mode of each time step based on the time series features output by the long short-term memory network. Finally, a more accurate prediction result of the discharge mode of inpatients is obtained.

[0021] Inpatient resource allocation steps: Inpatient resources are allocated based on predicted discharge patterns.

[0022] Example

[0023] The allocation of inpatient resources is based on reliable prediction results of discharge patterns. Taking ischemic heart disease with a diagnosis code range of ICD I20-I25 as an example, the prediction of discharge patterns for inpatients with ischemic heart disease under the weekend effect is constructed. The specific implementation steps are as follows: Figure 1 described.

[0024] S1. Data collection and preprocessing steps:

[0025] First, the electronic medical records of hospitalized patients with ischemic heart disease were obtained from the medical database, and the clinical data of hospitalized patients were collected by date from the electronic medical records of patients. In the preprocessing stage, data screening was performed, and only data directly related to the three time points were selected. Data during the admission period were collected, including the patient's admission date, admission route, admission type, whether it was ST-segment elevation myocardial infarction (STEMI), admission diagnosis, admission situation, case classification, and personal information of patients such as age, gender, marital status, etc.; data during the surgery period were collected, including surgery date, elective surgery, surgery level, and surgery-related information; data during the discharge period were collected, including discharge date and discharge method. In order to improve data processing efficiency and optimize the data structure of hospitalized patients, including data merging and deduplication, marking working days and non-working days (including weekends and holidays) in date data to consider weekend effects, and integer value labeling processing was also performed on other data to simplify the data set structure and improve the efficiency of subsequent analysis. After data screening and structural optimization, the data of hospitalized patients with weekend effects were obtained.

[0026] S2. Feature selection and key factor identification steps:

[0027] First, feature engineering was performed to perform preliminary feature selection on all variables that may affect the discharge mode, including the patient's basic information, including age, gender, marital status, admission information, and surgical information. The data set was then divided into a training set and a test set, with 80% used for training and 20% for testing; ensuring that the distribution of various discharge modes in the training set and the test set were roughly the same to avoid data bias. The training set data was used to build a random forest model, and the parameters of the random forest model were set, such as the number of decision trees, maximum depth, etc. After the model was trained, the test set data was used to evaluate the performance of the model. Common evaluation indicators include accuracy, precision, recall, and F1 value.

[0028] Variable importance assessment is performed using the feature importance attribute in the random forest model. Assess the importance of each variable for predicting discharge patterns, and rank the importance of the variables in descending order to identify the variables that have the greatest impact on the prediction results. Based on the results of the variable importance assessment, select several of the most important variables as the final features. You can set a threshold and only select variables with importance scores that exceed the threshold. Retrain the random forest model using the selected key features and adjust the parameters, such as using grid search to optimize hyperparameters. Repeated training and validation ensures the stability and predictive performance of the model. Further validate the performance of the model through cross-validation, such as k-fold cross-validation, to ensure that it performs consistently on different data subsets.

[0029] S3. Causal analysis steps:

[0030] Based on the key features selected in the feature selection and key factor identification steps, the initial structure of the Bayesian causal network is constructed. Nodes representing features and edges representing causal relationships are set to preliminarily determine possible causal relationships. Then, the data of key features are used for training, and the parameters in the network are learned through maximum likelihood estimation or Bayesian estimation methods. The network structure is optimized through structural learning algorithms, such as greedy algorithms, to ensure that the network can best represent the causal relationships in the data. The performance of the Bayesian causal network is evaluated using the cross-validation method to ensure that it accurately reflects the causal probability of the features on the discharge mode.

[0031] After the Bayesian causal network determines the causal relationship, the causal forest model is used to further explore the nonlinear relationships and interactions in these causal relationships. The causal forest model is constructed based on the causal relationship features determined by the Bayesian causal network. The causal forest model deeply explores the complex relationships and interactions between features by using a decision tree-based method. When training the causal forest model, the key features and discharge mode are input into the model as target variables, and the model will learn the nonlinear relationships between features. The output of the causal forest model is the specific causal effect estimate of each feature on the discharge mode, including the average causal effect and the conditional causal effect, and reveals the nonlinear relationships and interactions between features. Special attention is paid to features that contain weekend effects, such as the impact of admission date, admission route, and surgery date on discharge mode, and the difference in the effects of these features on weekdays or non-working days is evaluated.

[0032] The admission date and surgery date are features that include weekend effects because they contain date data markers. The admission route usually refers to the patient entering the hospital through the emergency department, outpatient department, etc. Regular outpatient services may not be open on weekends, and more patients may enter the hospital through the emergency department. In addition, the medical resources on weekends, such as the number of medical devices available, the configuration of doctors and nurses, etc., will be different from weekdays. The patient's admission route or the admission method that can be chosen will affect the discharge method due to differences in variables such as treatment plans, medical results, and length of hospital stay. This can to some extent reflect the differences in medical service contact and usage patterns between weekends and weekdays, and thus it includes weekend effects.

[0033] S4. Time series characteristics:

[0034] The surgical period was used as the mediating variable between the admission period and the discharge mode. A vector autoregression model was constructed. Multidimensional time series data of inpatients, including variable data of admission period, surgical period and discharge period, were input. The optimal order (number of lags) of the vector autoregression model was selected by the information criterion. The parameters of the vector autoregression model were estimated by the least squares method during the model training process. The performance of the vector autoregression model was estimated using the cross-validation method to ensure that the model could capture the dynamic change characteristics of each variable at the admission time point, surgical time point and discharge time point in the time series data. Multiple regression analysis was then used to predict the values ​​of each variable at each time step except the initial time step-admission time step. The dynamic change characteristics of the time series were constructed, which consisted of the original data of each variable at the admission time step, the predicted data of each variable at the surgical time step, and the discharge mode as the predicted data.

[0035] Dynamic time warping technology is used to measure the similarity of time series. The time series data of hospitalized patients' admission and surgery periods are input, and the similarity of patients in the time series is calculated. For each patient, a set of similarity features of the admission-surgery time-discharge method sequence is generated.

[0036] The dynamic change features extracted by the vector autoregression model are merged with the similarity features calculated by the dynamic time warping technology. The two sets of feature vectors are spliced ​​into a comprehensive feature vector through feature connection. The merged features are standardized to ensure that different features have the same dimension to avoid some features causing deviations in the subsequent prediction model training.

[0037] S5. Discharge mode prediction steps:

[0038] 5-1. Steps to integrate weekend effect causal feature data:

[0039] Merge key variable features, causal relationship features, and time series features to ensure that all feature dimensions are consistent and complete. Key variable features include basic information such as admission date, surgery date, and admission method. Causal relationship features come from Bayesian causal networks and causal forest models, including the causal probability and causal effect of features on discharge methods. Time series features include dynamic change features and similarity features generated by vector autoregression models and dynamic time warping techniques. Align the data in each feature according to the admission time point, surgery time point, and discharge time point to ensure that the data at all time points are correctly matched to avoid data time point misalignment. The alignment process needs to consider the timestamp information of the data to ensure that the feature data at each time step corresponds to the corresponding time point. Standardize the feature data, unify the dimension of the features, and ensure that the data is compared and analyzed on the same scale. Standardization ensures that the model training process is not affected by the difference in feature dimensions, thereby improving the model training effect and prediction accuracy.

[0040] 5-2. Long short-term memory network processing steps:

[0041] Recurrent neural networks are neural networks that process sequence data and are able to process sequences of arbitrary lengths, with information being passed from one unit to the next over time. Standard recurrent neural networks face the problems of gradient vanishing and gradient exploding, which limit their ability to learn on long sequences. Long short-term memory networks regulate the flow of information by introducing three gating structures, effectively preserving long-term dependencies and avoiding gradient problems. Long short-term memory networks are a special type of recurrent neural network with memory units and gating units that can better capture long-term dependencies in sequences. In a long short-term memory network, the output at each time point depends not only on the input and hidden state at the current time point, but also on the hidden state and cell state at the previous time point. This enables long short-term memory networks to better process sequence data, especially on long sequences.

[0042] In these gating mechanisms, the forget gate determines how much of the cell state information at the previous moment is retained. The output gate determines how much new information content is updated to the cell state. The candidate memory cell generates new information candidates. The new cell state in the state update is updated based on the output of the forget gate and the input gate, including the forgotten part of the information from the previous cell state and adding new information, and the output is guaranteed to be between -1 and 1 through the tanh function. The final hidden state is calculated based on the output gate and the current cell state, and determines the information content of the output. These operations ensure that the hidden state contains the information of the current time step and will be used as the input part of the next time step.

[0043] The long short-term memory network model processing enhances the understanding and memory of the time series of hospitalized patients during the admission period, surgery period and discharge mode, and combines the weekend effect characteristic data to improve the accuracy and efficiency of predicting the discharge mode of hospitalized patients. Through this method, the present invention ensures the high accuracy and robustness of analyzing the discharge mode under the weekend effect, and further provides a suitable treatment time plan for hospital management and patient treatment.

[0044] The application of LSTM networks is to effectively capture and analyze the time series data of hospitalized patients during the period of admission, surgery, and discharge. The uniqueness of LSTM networks lies in their ability to establish forward and backward dependencies for each time step in the time series. Specifically, when the LSTM network analyzes the data of a specific treatment day, it not only takes into account the characteristics related to the weekend effect during the admission period, but also comprehensively considers the characteristics related to the weekend effect during the surgery period. This comprehensive consideration of the context enables the LSTM network to more comprehensively and accurately grasp the contextual relationship of the key moments in the treatment process of hospitalized patients under the influence of the weekend effect when dealing with and analyzing the discharge patterns of hospitalized patients with the weekend effect.

[0045] During the implementation, the processed feature data is input into the long short-term memory network. The long short-term memory network can process complex time series data, capture the dependency between the time steps of the admission period and the surgery period, and is suitable for processing data with time series characteristics. Through the long short-term memory network, we can better understand the time series data of the admission period and the surgery period, and combine the weekend effect feature data to improve the accuracy and efficiency of predicting the discharge mode of inpatients.

[0046] 5-3. Time distribution fully connected layer processing steps:

[0047] The temporal fully connected layer is used to output the predicted discharge mode of inpatients to achieve accurate prediction of the discharge mode.

[0048] Through the training data, the back propagation algorithm is used to adjust the network parameters and minimize the loss function. The long short-term memory network outputs the time series features of each time step, and then these features are input into the fully connected layer for further processing. The fully connected layer processes the time series features output by the long short-term memory network through linear transformation and activation function to extract high-level feature information. Based on the time series features output by the long short-term memory network, the fully connected layer predicts the patient's discharge mode. The prediction results can be used to obtain the specific discharge mode category through the Softmax function or other classifiers. The model performance is evaluated by indicators such as accuracy and F1 score. The prediction performance of the model is evaluated using the validation data set, and the model is further optimized through cross-validation and hyperparameter tuning. According to the model evaluation results, the network structure and parameters are adjusted to improve the prediction effect of the model, and finally the long short-term memory network model with the best performance is obtained.

[0049] By integrating the long short-term memory network structure and the application of the fully connected layer, the model we proposed can effectively process and analyze the time series data of hospitalized patients in terms of admission period, surgery period and discharge mode. This method is of great significance for the analysis of time series data in the medical field, and can help medical professionals better understand and predict the discharge mode of patients under the weekend effect, thereby optimizing treatment plans and providing better medical services for patients. Through the present invention, we can provide a more accurate and reliable tool for analyzing the impact of the weekend effect in the medical field, provide more powerful support for medical management decision-making to solve the negative impact of the weekend effect, and provide patients with more appropriate diagnosis and treatment time recommendations.

[0050] The model for predicting the discharge mode of inpatients under the influence of the weekend effect using the above steps includes random forest, Bayesian causal network, vector autoregression model, dynamic time warping technology, and long short-term memory network. When the data of the admission and operation period of historical inpatients are input into the discharge mode prediction model, after the above steps, the discharge mode prediction model can output the predicted discharge mode.

[0051] The embodiment selected the electronic medical record information of inpatients from a municipal tertiary hospital and performed multiple preprocessing steps to ensure the accuracy and relevance of the data.

[0052] The embodiment selected the electronic medical record data of inpatients with ischemic heart disease in a municipal tertiary hospital, and performed a number of preprocessing steps to ensure the accuracy and relevance of the data. Then, the random forest algorithm was used for feature selection to identify the key factors that affect the discharge mode in the weekend effect. The Bayesian causal network and causal forest model were used to explore the causal relationship between the key factors. Then, the surgical period was regarded as a mediating variable between the admission period and the discharge mode, and the vector autoregression model was used to analyze the time series data to capture the dynamic relationship between the variables; the dynamic time warping technology was used to measure the similarity of the admission and surgical process of different patients, and generate time series similarity features; the key factors were combined with the causal features, the features generated by the vector autoregression model and the dynamic time warping technology, and input into the long short-term memory network model to better predict the patient's discharge mode, and the generalization ability of the model was evaluated using cross-validation. The weekend effect in medical services predicts the patient's discharge mode, providing a scientific basis for clinical decision-making and supporting more accurate medical management and decision-making.

[0053] The implementation scheme has been verified through a series of experiments, which proves its remarkable feasibility and advantages in predicting the discharge mode of hospitalized patients under the influence of weekend effect. The experimental results show that the model is highly effective in mining the causal characteristics of the data of hospitalized patients associated with weekend effect, showing the excellent ability of the model in extracting causal characteristics in time series from medical data. In the deep learning model training stage, the random forest, Bayesian causal network, vector autoregression model, dynamic time warping technology, and long short-term memory network model architecture demonstrated its powerful mining of causal information in medical data associated with weekend effect, and predicted the discharge mode in terms of key performance indicators such as accuracy, precision, recall rate and F1 score. The significant accuracy reflects its superior learning and generalization capabilities.

[0054] In the validation phase, the model maintained high accuracy in predicting discharge patterns for new and unknown inpatients’ time series data sets of admission and surgery periods, demonstrating its good adaptability and generalization capabilities. In the testing phase, the model maintained high levels of accuracy and stability even when processing diverse inpatient electronic medical record data. These results indicate that our model can not only accurately predict inpatient discharge patterns under the influence of the weekend effect, but also maintain its prediction stability and accuracy on electronic medical record data of users with different diagnosis types.

[0055] In summary, the experimental verification results of random forest, Bayesian causal network, vector autoregression model, dynamic time warping technology and long short-term memory network model architecture strongly prove the application value of the present invention in the field of predicting the discharge mode of inpatients under the influence of weekend effect. It is not only innovative in theory, but also shows significant effects and advantages in practical application. These advantages make the present invention an efficient and reliable tool for predicting the final discharge mode of inpatients affected by weekend effect in medical practice, which helps to promote the comprehensive progress of improving medical quality and effectively utilizing medical resources related to weekend effect, thus making an important contribution to the management of inpatients in the medical field.

Claims

1. A method for allocating hospitalization resources under the weekend effect based on machine learning, characterized in that: Includes steps: Data collection and preprocessing steps: Inpatient data were collected from the medical service database in date order, and the date variables in the inpatient data were marked as working days and non-working days; the inpatient data included the patient's personal information, data during admission, data during surgery, and data during discharge; Key feature selection step: Use the random forest algorithm to construct a decision tree to evaluate the importance of each variable in the inpatient data, and identify the key variables that affect the discharge mode as key features; Causal feature selection step: Based on the selected key feature - discharge mode, a Bayesian causal network is constructed to explore the causal relationship between the key features including the weekend effect and the discharge mode. The nonlinear relationship and interaction between the key features of the causal forest model based on the decision tree are used to obtain the causal relationship features for predicting the discharge mode of inpatients under the weekend effect. Steps for generating time series features: Use the time series data consisting of the data of the hospitalized patients during the admission period, the data of the surgery period, and the data of the discharge period to train the vector autoregression model. The vector autoregression model learns to capture the dynamic change relationship of each variable at each time point during the admission period, the surgery period, and the discharge period. The dynamic time warping technology is used to measure the similarity of the admission and surgery processes of different patients. The trained vector autoregression model is used to receive the input time series data of the hospitalized patients during the admission period and the surgery period, calculate the similarity of each hospitalized patient in the time series, and output a set of time series features of the admission-surgery-discharge mode corresponding to each patient; Discharge mode prediction steps: merge key features, causal relationship features and time series features, preprocess the merged features, and then input the preprocessed features into the long short-term memory network. The long short-term memory network outputs the time series features of each time step to the fully connected layer; The fully connected layer outputs the predicted discharge mode at each time step based on the input time series features; Inpatient resource allocation steps: Inpatient resources are allocated based on predicted discharge patterns.

2. The method according to claim 1, characterized in that: The specific method of preprocessing the merged features is: align the variables in the merged key features, causal relationship features, and time series features according to the admission time point, surgery time point, and discharge time point; then standardize the feature data to unify the dimension of the features.

3. The method according to claim 1, characterized in that: The data during the admission period include the patient's admission date, admission route, admission type, whether it is ST-segment elevation myocardial infarction (STEMI), admission diagnosis, admission condition and case classification.

4. The method according to claim 1, characterized in that: Data from the surgical period included surgical date, elective surgery, surgical level, and surgical-related information.

5. The method according to claim 1, characterized in that: Data on discharge included the date and method of discharge.

6. The method according to claim 1, characterized in that: Key characteristics that account for the weekend effect include admission date, admission route, and surgery date.

7. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to claim 1 are implemented.

Images