A method and device for constructing a fundamental cause of death prediction model

Abstract

The present invention relates to a method and device for constructing a fundamental cause of death prediction model; the method comprises: obtaining a plurality of historical electronic death card data, and performing data preprocessing to obtain a training sample set; constructing a fundamental cause of death prediction model, the fundamental cause of death prediction model comprising an incremental prediction module, a recurrent neural network prediction module, a naive Bayes classification module, and a fusion module for weighted fusion based on the prediction results of each module arranged in parallel; using the data in the training sample set, the incremental prediction module, the recurrent neural network prediction module and the naive Bayes classification module are trained respectively; and based on the prediction accuracy of the fundamental cause of death of each module after training, the weight coefficients of each module for weighted fusion are calculated to complete the construction of the fundamental cause of death prediction model. The present invention solves the problem in the prior art that the fundamental cause of death disease information needs to be manually judged, resulting in low efficiency in the entry of the fundamental cause of death and high error rate of the entry result.

Description

Technical Field

[0001] The present invention belongs to the technical field of statistics and machine learning, and specifically relates to a method and device for constructing a root cause of death prediction model. Background Art

[0002] The registration of population death information is basic data for strengthening population management and evaluating the health level of residents, and filling in the root cause of death in the death chain is an important task in the registration of population death information.

[0003] At present, in order to analyze the chain of causes of death and determine the root cause of death, it is necessary to refer to a lot of information, such as the rules and notes for determining the root cause of death in the Guidance Manual for Population Death Cause Monitoring, or to manually analyze and classify according to the ICD-10 standard. The subjective factors of the classifiers and the complexity of the workload can easily lead to misjudgment, resulting in the root cause of death being entered incorrectly or not meeting the standards.

[0004] Therefore, it becomes a very meaningful task to train an intelligent prediction model for the root cause of death, reduce the workload of manual judgment, and improve the accuracy of the coding of the root cause of death. Summary of the invention

[0005] In view of the above analysis, the present invention aims to disclose a method and device for constructing a root cause of death prediction model, which solves the problem in the prior art that human judgment of root cause of death disease information is required, resulting in low efficiency in entering the root cause of death and high error rate in the entry results.

[0006] The purpose of the present invention is mainly achieved through the following technical solutions:

[0007] In one aspect, the present invention discloses a method for constructing a fundamental cause of death prediction model, comprising:

[0008] Acquire multiple historical electronic death card data, perform data preprocessing, obtain multiple cause of death chain feature data, and annotate each cause of death chain feature data with a root cause of death category label to construct a training sample set;

[0009] Constructing a fundamental cause of death prediction model, the fundamental cause of death prediction model comprising an incremental prediction module, a recurrent neural network prediction module, a naive Bayes classification module, and a fusion module for weighted fusion based on prediction results of each module;

[0010] The incremental prediction module, the recurrent neural network prediction module and the naive Bayes classification module are trained respectively using the data in the training sample set; and based on the prediction accuracy of the underlying cause of death of each module after training, the weight coefficient of each module for weighted fusion is calculated to complete the construction of the underlying cause of death prediction model.

[0011] Furthermore, the incremental prediction module is an incremental tree model based on multiple decision trees, which is trained using a k-fold cross-validation method;

[0012] The recurrent neural network prediction module is a prediction model based on the GRU model, which is trained by the Adam optimizer;

[0013] The naive Bayes classification module is a prediction module based on the naive Bayes classification algorithm. It uses the death cause chain feature data and the corresponding root cause of death category labels in the training sample set to calculate the prior probability of each category label and the conditional probability under each category label, and uses the smoothness parameter for optimization training.

[0014] Furthermore, during the training process, the incremental prediction module divides the training sample set into k training subsets, randomly selects one training subset as a test set, and trains each training subset as a training set to obtain multiple decision tree models; the decision tree model is used to output the predicted cause of death disease category;

[0015] A grid search is performed on the output of each decision tree model to obtain the optimal hyperparameters of the incremental prediction module after training.

[0016] Further, each training subset is used to perform training based on a preset number of training times; during each training, the size of the sliding window is adjusted, multiple feature extractions are performed on each data block, and a decision tree model for each training is generated respectively, so as to obtain multiple decision tree models corresponding to each data block; wherein,

[0017] In a training process, a preset proportion of sample data is randomly selected, and the selected sample data is divided into multiple data blocks; feature data is extracted from each data block based on a sliding window and information entropy and information gain are calculated, and a decision tree model obtained from this training is generated based on the information entropy and information gain.

[0018] Furthermore, during the training of the recurrent neural network prediction module, the following loss function is used for iterative updating to obtain a converged recurrent neural network prediction module:

[0019]

[0020] Among them, Q is the result of the loss function, m represents the number of samples, and y i represents the underlying cause of death category label of the i-th sample data, that is, the actual result value, x i Represents the i-th sample data.

[0021] Furthermore, during the training process of the naive Bayes classification module, the smoothness parameter alpha is set to 1.0 for smoothing optimization; and the Boolean parameter fit_prior is set to true to learn the prior probability of the label category.

[0022] Furthermore, the electronic death card data is preprocessed, including:

[0023] Based on the electronic death card data, the cause of death chain information within a preset range is obtained; the cause of death chain information includes information on multiple cause of death diseases and the time interval from the occurrence of each cause of death disease to death;

[0024] Obtain the ICD-10 code corresponding to the cause of death;

[0025] The letter part of the ICD-10 code is converted into ASCII value, and the value range of the ICD-10 code is converted by the following formula:

[0026] D=d ASCii *1000+d Dig *10;

[0027] Where D is the cause of death disease data value after the value range is converted; d ASCii Convert the letter part of the ICD-10 code to ASCII value, d Dig The value of the numeric part of the ICD-10 code;

[0028] Fill the data with 0 for the cause of death disease in the cause of death chain information that is empty;

[0029] The cause of death disease information after value range conversion is spliced ​​based on the time interval to complete the data preprocessing.

[0030] Furthermore, based on the prediction accuracy of the underlying cause of death of each module after training, the weight coefficients of each module for weighted fusion are calculated, including:

[0031] Obtaining the cause of death chain information of multiple electronic death cards within a preset time, and performing value range conversion to generate multiple feature data to be predicted;

[0032] Inputting the feature data to be predicted into the incremental prediction module, the recurrent neural network prediction module and the naive Bayes classification module respectively to predict the underlying cause of death;

[0033] The accuracy of the underlying cause of death inference of each module was calculated, and the accuracy of each module was converted into the weight of the model using the following formula:

[0034]

[0035] Among them, α jis the weight of the jth module, ACC j is the accuracy of the jth model, is the total accuracy of the three models.

[0036] Furthermore, the fusion module performs weighted fusion based on the predicted probabilities of the same death cause disease category output by each module through the following formula to obtain the predicted probabilities of each death cause disease category after weighted fusion:

[0037]

[0038] Among them, LP i is the predicted probability of the weighted fusion of the i-th cause of death disease category, α ij is the weight of the jth module, P ij The probability of the i-th cause of death disease category predicted by the j-th model.

[0039] In another aspect, a computer device is disclosed, comprising at least one processor, and at least one memory communicatively connected to the processor;

[0040] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the aforementioned method for constructing a root cause of death prediction model.

[0041] The present invention can achieve at least one of the following beneficial effects:

[0042] 1. Based on the characteristics of large amount of newly added electronic death card data in practical applications, the need to continuously update the categories of cause of death diseases, and the discreteness of cause of death chain data, the present invention adaptively adopts the incremental prediction module, the recurrent neural network prediction module and the naive Bayes classification module to perform feature extraction and classification prediction on the electronic death card data respectively, and obtains the final prediction result through data fusion, thus avoiding the problem that a single classification prediction model is biased towards certain features of the data when performing classification prediction, which easily leads to inaccurate prediction;

[0043] 2. This method obtains the cause of death disease information and corresponding time information related to the cause of death based on historical electronic death card data, and constructs training sample data based on multiple cause of death disease information and the time interval from the occurrence of each cause of death disease to death, and trains the underlying cause of death prediction model, thereby improving the efficiency and accuracy of the model in inferring the underlying cause of death;

[0044] 3. The present invention utilizes the characteristic of ICD-10 codes to classify similar diseases into the same category, and converts the value range of the ICD-10 codes corresponding to the cause of death diseases based on a specific method, converting similar diseases into similar numerical values, which is beneficial to model training and improves the accuracy of the inference of the underlying cause of death. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0046] Figure 1 is a flow chart of a method for constructing a fundamental cause of death prediction model in an embodiment of the present invention;

[0047] Figure 2 Schematic diagram of the training process of the underlying cause of death prediction model in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used to illustrate the principles of the present invention together with the embodiments of the present invention.

[0049] The embodiment of the present invention discloses a method for constructing a fundamental cause of death prediction model, such as Figure 1 As shown, the following steps are included:

[0050] Step S1: obtaining a plurality of historical electronic death card data, and performing data preprocessing to obtain a plurality of cause of death chain feature data, and labeling each of the cause of death chain feature data with a root cause of death category label to construct a training sample set;

[0051] Specifically, after obtaining multiple historical electronic death card data, the electronic death card data is first preprocessed by the following method:

[0052] Based on the electronic death card data, the cause of death chain information within a preset range is obtained, that is, a preset number of cause of death disease information closest to the death time is obtained; wherein the cause of death chain information includes multiple cause of death disease information and the time interval from the occurrence of each cause of death disease to death;

[0053] Obtain the ICD-10 code corresponding to the cause of death; convert the letter part of the ICD-10 code into ASCII value, and convert the value range of the ICD-10 code by the following formula:

[0054] D=d ASCii *1000+d Dig *10;

[0055] Where D is the cause of death disease data value after the value range is converted; d ASCii Convert the letter part of the ICD-10 code to ASCII value, d Dig The value of the numeric part of the ICD-10 code;

[0056] Fill the empty data of the cause of death disease in the cause of death chain information with 0;

[0057] The cause of death disease information after value range conversion is spliced ​​based on the time interval to complete data preprocessing.

[0058] More specifically, the electronic death card data records the cause of death chain information of the deceased, which generally includes multiple cause of death disease information (i.e., cause of death disease names) related to the cause of death of the deceased and the time interval from the occurrence of each cause of death disease to death; in this embodiment, the cause of death diseases in the acquired cause of death chain information are sorted from near to far according to the occurrence time, and multiple cause of death diseases with the shortest time interval to death are selected. Based on the disease names and death times of the selected multiple cause of death diseases, the cause of death disease information is converted into a value range, sorted by occurrence time, and spliced ​​in the following way to obtain the cause of death chain feature data to construct a training sample set for training the root cause of death prediction model:

[0059] [d1,t1,d2,t2,……d i ,t i ……d n ,t n ];

[0060] Among them, d i is the data value of the ith cause of death disease after the value range is converted, t i is the time interval from the occurrence of the ith cause of death to death.

[0061] It should be noted that ICD-10 (International Classification of Diseases, 10th Edition) coding is an international unified disease classification method developed by WHO. It classifies diseases according to their etiology, pathology, clinical manifestations, and anatomical location, making them an orderly combination and a system represented by coding methods. It consists of structures such as chapters, blocks, categories, and subcategories. The ICD-10 coding of existing diseases covers all codes from A00.0 to Y98. In order to convert the ICD-10 coding corresponding to the cause of death disease information into data that can be used for machine learning, the ord method of Python can be used to convert the letter part of the ICD-10 coding into ASCII values, multiply it by 1000, and then multiply the value of the numeric part by 10 for value range conversion to obtain a number between 65000 and 89980; if the cause of death disease is a null value, it is filled with 0. Because ICD-10 codes classify similar diseases into the same chapter blocks and categories, the above-mentioned value range conversion method will also classify similar diseases into similar values, which is beneficial to model training and prediction and improves prediction accuracy.

[0062] Step S2: constructing a fundamental cause of death prediction model, wherein the fundamental cause of death prediction model includes an incremental prediction module, a recurrent neural network prediction module, a naive Bayes classification module, and a fusion module for weighted fusion based on the prediction results of each module.

[0063] Specifically, the incremental prediction module is an incremental tree model based on multiple decision trees, which is trained using the k-fold cross-validation method. The incremental prediction module adopts the StreamingRFC algorithm. For newly added electronic death card data, the training sample set is added, and the partial_fit() method is called to perform incremental training of the model to adapt to the characteristics of the continuous increase of electronic death card data and the continuous updating of cause of death disease categories, thereby improving the accuracy of the model in inferring the underlying cause of death.

[0064] The recurrent neural network prediction module is a prediction model based on the GRU model, which is trained by the Adam optimizer; it is used to process sequence data to capture the long-term dependency of the death cause chain feature data and solve the gradient vanishing problem. The recurrent neural network prediction module of this embodiment includes a two-layer GRU model. The first layer is a GRU layer with 64 units; the second layer is a GRU layer with 32 units, and a fully connected layer with 10 units and a softmax activation function is added to the output layer to predict the category of the underlying cause of death.

[0065] The naive Bayes classification module is a prediction module based on the naive Bayes classification algorithm. It uses the cause of death chain feature data and the corresponding root cause of death category labels in the training sample set to calculate the prior probability of each category label and the conditional probability under each category label, and optimizes the training using the smoothness parameter. The naive Bayes classification module is a naive Bayes classifier used in the Scikit-Learn library for processing classification features. It is suitable for processing classification problems of multiple discrete features. It assumes that each feature is independent. It works better when the features are all classification variables. Therefore, it is suitable for the root cause of death inference based on the cause of death chain information in this embodiment.

[0066] The fusion module is used to perform weighted fusion on the predicted probabilities of the same disease category labels in the prediction results of each module based on the accuracy of each prediction module, and obtain the disease category with the highest predicted probability as the final prediction result of the root cause of death prediction model. This avoids the problem of using the output of a single classification prediction model as the final prediction result, in which the model will focus on certain features of the data, which may easily lead to inaccurate predictions.

[0067] Step S3: using the data in the training sample set, respectively training the incremental prediction module, the recurrent neural network prediction module and the naive Bayes classification module; and based on the prediction accuracy of the underlying cause of death of each module after training, calculating the weight coefficient of each module for weighted fusion, and completing the construction of the underlying cause of death prediction model.

[0068] Specifically, Figure 2 As shown, based on multiple copies of historical death card information obtained in advance, after the value range conversion and splicing of the cause of death chain data, the incremental prediction module, recurrent neural network prediction module, and naive Bayes classification module are trained respectively, and the weights are calculated based on the prediction accuracy of each module to obtain the final root cause of death prediction model.

[0069] In the training process, the incremental prediction module divides the training sample set into k training subsets, randomly selects one training subset as a test set, and trains each training subset as a training set to obtain multiple decision tree models; the decision tree model is used to output the predicted cause of death disease category;

[0070] A grid search is performed on the output of each decision tree model to obtain the optimal hyperparameters of the incremental prediction module after training.

[0071] Further, each training subset is used to perform training based on a preset number of training times; during each training, the size of the sliding window is adjusted, multiple feature extractions are performed on each data block, and a decision tree model for each training is generated respectively, so as to obtain multiple decision tree models corresponding to each data block; wherein,

[0072] In a training process, a preset proportion of sample data is randomly selected, and the selected sample data is divided into multiple data blocks; feature data is extracted from each data block based on a sliding window and information entropy and information gain are calculated, and a decision tree model obtained from this training is generated based on the information entropy and information gain.

[0073] That is, this embodiment uses a combination of k-fold group cross-validation and grid search to determine the hyperparameters of the incremental prediction model. First, initialize a Fold object. The parameter n_splits represents the number of folds of the data set, which is set to k. Shuffle is set to true, which means that the data will be shuffled to the greatest extent possible to make the data evenly distributed. The Fold object divides the sample data set into k folds to obtain k sample subsets. Each sample subset is used as a training set for model training. During each iterative training, a sample subset is randomly selected as a test set for model testing. The value of k is k=n / 50000 (where n is the number of samples in the data set). In practical applications, as the amount of data increases, the value of k can be adaptively adjusted.

[0074] Then define the hyperparameter combination interval of the incremental tree. The hyperparameter spf_sample_prop represents the proportion of samples randomly selected each time the incremental training partial_fit() method is called, and the interval is 0.1-0.9, which means that each time the partial_fit() method is called, 10% to 90% of the samples will be randomly selected from the training set used for training to perform partial fitting of the model. spf_n_fits is used to specify the number of times the partial_fit() method is called, and the interval is set to 10-50. By calling the partial_fit() method multiple times, data can be provided step by step and the model can be updated step by step. This embodiment adopts the StreamingRFC algorithm, and the training data in each training process is divided into multiple data blocks. The parameter n_estimators_per_chunk is used to specify the number of decision trees in each data block (chunk), and each data block contains a portion of sample data and corresponding labels. The n_estimators_per_chunk parameter determines the number of decision trees to be trained in each data block, and the interval is set to 2-10. max_n_estimators is used to specify the maximum number of decision trees. In this embodiment, this parameter is set to np.inf, that is, the maximum number of decision trees is unlimited.

[0075] After training multiple decision trees, grid search and cross-validation are performed based on the output of each decision tree to obtain the optimal hyperparameters of the incremental prediction model. Then, based on the optimal hyperparameters, the incremental learning method is called in batches to learn and the final incremental prediction module can be obtained.

[0076] Furthermore, the recurrent neural network prediction module is a classification prediction model based on the GRU model (gated recurrent unit). The first layer of the recurrent neural network prediction module of this embodiment is a GRU layer with 64 units, and return_sequences=True is set to pass the sequence output to the next layer. The second layer is a GRU layer with 32 units, and a fully connected layer with 10 units and a softmax activation function is added to the output layer to predict the category of the underlying cause of death.

[0077] The recurrent neural network prediction module is trained using the Adam optimizer, and the loss function uses sparse_categorical_crossentropy to handle the multi-classification problem of integer labels in this embodiment, and calculate appropriate losses during model training. The metric uses accuracy.

[0078] The calculation formula of the sparse_categorical_crossentropy function is as follows:

[0079]

[0080] Among them, Q is the result of the loss function, m represents the number of samples, and y i represents the underlying cause of death category label of the i-th sample data, that is, the actual result value, x i Represents the i-th sample data.

[0081] Furthermore, the generation steps of the naive Bayes classification module are as follows:

[0082] Initialize the CategoricalNB model and set the smoothness parameter alpha to 1.0 to control the estimated prior probability and solve the problem of data sparsity. Set fit_prior to true to indicate that the prior probability of the label category needs to be learned. Then, pass the pre-built sample data set and the underlying cause of death category label into the model, call the fit training function for training, and you can get the final naive Bayes classification module.

[0083] Furthermore, after the training of the underlying cause of death inference model is completed, the weights of the incremental prediction module, the recurrent neural network prediction module and the naive Bayes classification module are set by the following method:

[0084] The cause of death chain information of multiple electronic death cards within a preset time is obtained. Preferably, in this embodiment, electronic death card data within one month is selected for weight calculation, and the selected electronic death card data is data with a fundamental cause of death label; the cause of death disease information in the selected electronic death card data is converted into a value range, and the cause of death disease information is spliced ​​based on the death time to generate multiple feature data to be predicted; the feature data to be predicted is respectively input into an incremental prediction module, a recurrent neural network prediction module, and a naive Bayes classification module to predict the fundamental cause of death; based on the prediction results and the pre-marked fundamental cause of death label, the accuracy of the fundamental cause of death inference of each module is calculated, and the accuracy of each module is converted into the weight of the model by the following formula:

[0085]

[0086] Among them, α j is the weight of the jth module, ACC j is the accuracy of the jth model, is the total accuracy of the three models.

[0087] The fusion module performs weighted fusion on the predicted probabilities of the same cause of death category predicted by each classification prediction model using the following formula:

[0088]

[0089] Among them, LP iis the predicted probability of the weighted fusion of the i-th cause of death disease category, α ij is the weight of the jth module, P ij The probability of the i-th cause of death disease category predicted by the j-th model.

[0090] In practical applications, the model can also be updated based on a preset cycle. In this embodiment, the update cycle is set to 1 month, that is, the cause of death chain data within the last month is collected, and the cause of death chain data is organized into training data according to the above-mentioned value range conversion method, and is respectively passed into the above-mentioned three modules. The incremental prediction module and the naive Bayes classification module respectively call the predict_proba function to generate the probability of each cause of death disease category label result, and then use the np.argsort function to sort the probability array and take the highest three category labels. The recurrent neural network module calls the softmax function to generate the results and probabilities of each cause of death disease category label, and then sorts them according to the probability to obtain the three labels with the highest probability and probability. Then, the final probability of each label is calculated according to the weights of the three modules, and the probabilities of the same labels are summed to obtain the cause of death disease category label with the highest probability, that is, the final prediction result of the fundamental cause of death inference model.

[0091] Furthermore, it should be noted that when the underlying cause of death prediction model constructed by this method is used to predict the underlying cause of death, the underlying cause of death category label output by the model is the coded data after the value range conversion by the aforementioned method. After obtaining the prediction result, it is also necessary to restore the predicted underlying cause of death category label to the cause of death disease category by the following method to obtain the underlying cause of death disease category corresponding to the electronic death card data to be processed;

[0092] Divide the cause of death disease category label data value output by the underlying cause of death prediction model by 1000 to obtain an integer part and a remainder part;

[0093] The integer part is converted into letters of ASCII value, and concatenated with the number obtained by dividing the remainder part by 10 to obtain the ICD-10 code corresponding to the category of the cause of death disease.

[0094] In summary, the method and device for constructing a fundamental cause of death prediction model of the present invention, based on the characteristics of a large amount of newly added electronic death card data in practical applications, the need to continuously update the categories of cause of death diseases, and the discrete type of cause of death chain data, adaptively adopts an incremental prediction module, a recurrent neural network prediction module, and a naive Bayes classification module to perform feature extraction and classification prediction on the electronic death card data respectively, and obtains the final prediction result through data fusion, thereby avoiding the problem that a single classification prediction model is biased towards certain features of the data when performing classification prediction, which easily leads to inaccurate prediction; and the present invention is based on the electronic death card. The card data obtains the cause-of-death disease information and the corresponding time information related to the cause of death, and based on multiple cause-of-death disease information and the time interval from the occurrence of each cause-of-death disease to death, uses the pre-trained underlying cause of death inference model to perform intelligent inference of the underlying cause of death, thereby improving the efficiency and accuracy of the underlying cause of death inference; when training the model, the characteristic of ICD-10 codes that classify similar diseases into the same category is utilized, and the value domain of the ICD-10 codes corresponding to the cause-of-death diseases is converted based on a specific method, so that similar diseases are converted into similar numerical values, which is conducive to model training and improves the accuracy of the underlying cause of death inference.

[0095] Another embodiment of the present invention further discloses a computer device, comprising at least one processor, and at least one memory communicatively connected to the processor;

[0096] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the aforementioned method for constructing a root cause of death prediction model.

[0097] Those skilled in the art will appreciate that all or part of the process of the method in the above embodiment can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0098] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for constructing a fundamental cause of death prediction model, characterized in that: include: A plurality of historical electronic death card data are obtained, and data preprocessing is performed to obtain a plurality of cause of death chain feature data, and a root cause of death category label is marked for each of the cause of death chain feature data to construct a training sample set; wherein the data preprocessing of the electronic death card data includes: Based on the electronic death card data, the cause of death chain information within a preset range is obtained; the cause of death chain information includes multiple cause of death disease information and the time interval from the occurrence of each cause of death disease to death; the ICD-10 code corresponding to the cause of death disease is obtained; the letter part of the ICD-10 code is converted into an ASCII value, and the value range of the ICD-10 code is converted by the following formula: ； in, is the cause of death disease data value after the value range is converted; Convert the ASCII value of the letter part of the ICD-10 code to The value of the numeric part of the ICD-10 code; Fill the data with 0 for the cause of death disease in the cause of death chain information that is empty; The cause-of-death disease information after the value range conversion is sorted by occurrence time and concatenated as follows to obtain the cause-of-death chain feature data: ； in, is the i-th cause of death disease data value after the value range is converted, is the time interval from the occurrence of the i-th cause of death to death; Constructing a root cause of death prediction model, the root cause of death prediction model comprising an incremental prediction module of an incremental tree model based on multiple decision trees, a recurrent neural network prediction module, a naive Bayes classification module, and a fusion module for weighted fusion based on the prediction results of each module; The incremental prediction module, the recurrent neural network prediction module and the naive Bayes classification module are trained respectively by using the data in the training sample set; and based on the prediction accuracy of the underlying cause of death of each module after training, the weight coefficients of each module for weighted fusion are calculated to complete the construction of the underlying cause of death prediction model; when the underlying cause of death prediction is performed using the constructed underlying cause of death prediction model, the underlying cause of death category label output by the model is the coded data after the value domain is converted. After obtaining the prediction result, the predicted underlying cause of death category label needs to be restored to the cause of death disease category by the following method to obtain the underlying cause of death disease category corresponding to the electronic death card data to be processed: the cause of death disease category label data value output by the underlying cause of death prediction model is divided by 1000 to obtain an integer part and a remainder part; the integer part is converted into letters of ASCII value, and is concatenated with the number obtained by dividing the remainder part by 10 to obtain the ICD-10 code corresponding to the cause of death disease category.

2. The method for constructing the underlying cause of death prediction model according to claim 1, characterized in that: The incremental prediction module is trained using a k-fold cross validation method; The recurrent neural network prediction module is a prediction model based on the GRU model, which is trained by the Adam optimizer; The naive Bayes classification module is a prediction module based on the naive Bayes classification algorithm. It uses the death cause chain feature data and the corresponding root cause of death category labels in the training sample set to calculate the prior probability of each category label and the conditional probability under each category label, and uses the smoothness parameter for optimization training.

3. The method for constructing the underlying cause of death prediction model according to claim 2, characterized in that: During the training process, the incremental prediction module divides the training sample set into k training subsets, randomly selects one training subset as a test set, and trains each training subset as a training set to obtain multiple decision tree models; the decision tree model is used to output the predicted cause of death disease category; A grid search is performed on the output of each decision tree model to obtain the optimal hyperparameters of the incremental prediction module after training.

4. The method for constructing the underlying cause of death prediction model according to claim 3, characterized in that: Using each training subset, training is performed based on a preset number of training times; during each training, the size of the sliding window is adjusted, multiple feature extractions are performed on each data block, and a decision tree model for each training is generated respectively, so as to obtain multiple decision tree models corresponding to each data block; wherein, In a training process, a preset proportion of sample data is randomly selected, and the selected sample data is divided into multiple data blocks; feature data is extracted from each data block based on a sliding window and information entropy and information gain are calculated, and a decision tree model obtained from this training is generated based on the information entropy and information gain.

5. The method for constructing the underlying cause of death prediction model according to claim 2, characterized in that: During training, the recurrent neural network prediction module uses the following loss function for iterative updating to obtain a converged recurrent neural network prediction module: ; Among them, Q is the result of the loss function, m represents the number of samples, Represents the root cause of death category label of the i-th sample data, that is, the actual result value, Represents the i-th sample data.

6. The method for constructing the underlying cause of death prediction model according to claim 2, characterized in that: During the training process of the Naive Bayes classification module, the smoothness parameter alpha is set to 1.0 for smoothing optimization; the Boolean parameter fit_prior is set to true to learn the prior probability of the label category.

7. The method for constructing a basic cause of death prediction model according to claim 1, characterized in that: Based on the prediction accuracy of the underlying cause of death of each module after training, the weight coefficients of each module for weighted fusion are calculated, including: Obtaining the cause of death chain information of multiple electronic death cards within a preset time, and performing value range conversion to generate multiple feature data to be predicted; Inputting the feature data to be predicted into the incremental prediction module, the recurrent neural network prediction module and the naive Bayes classification module respectively to predict the underlying cause of death; The accuracy of the underlying cause of death inference of each module was calculated, and the accuracy of each module was converted into the weight of the model using the following formula: ； in, For the The weight of each module, For the The accuracy of the model, is the total accuracy of the three models.

8. The method for constructing the underlying cause of death prediction model according to claim 7, characterized in that: The fusion module performs weighted fusion based on the predicted probabilities of the same death cause disease category output by each module through the following formula to obtain the predicted probabilities of each death cause disease category after weighted fusion: ； in, is the predicted probability of the weighted fusion of the i-th cause of death disease category, is the weight of the jth module, The probability of the i-th cause of death disease category predicted by the j-th model.

9. A computer device, characterized in that: comprising at least one processor, and at least one memory communicatively connected to the processor; The memory stores instructions executable by the processor, and the instructions are used to be executed by the processor to implement the method for constructing the underlying cause of death prediction model according to any one of claims 1 to 8.

Images