Deep learning-based pneumoconiosis prognosis multi-outcome competition risk prediction method
Through deep learning methods and deep recurrent neural networks, combined with interpolation method and SMOTE technology, a multi-destination competition risk prediction model was constructed, which solved the problem of difficult interaction between multiple factors in the existing technology, and achieved more accurate pneumoconiosis disease risk prediction and personalized intervention strategies improve prediction accuracy and prognostic management effectiveness.
Patent Information
- Application Number
- CN202310535895.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-05-12
- Publication Date
- 2024-11-12
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing technology has failed to effectively consider the complex impact of multiple influencing factors on the occurrence and prognosis of pneumoconiosis, and mainly uses linear models, making it difficult to deal with high-dimensional interaction problems, which limits the accuracy and in-depth development of pneumoconiosis risk prediction models.
Using a deep learning-based approach, a multi-outcome competition risk prediction model for pneumoconiosis prognosis by acquiring and preprocessing data of multiple influencing factors, using deep recurrent neural networks (DNNs) and attention mechanisms, combining interpolation, SMOTE technology and index Smoothing method, deals with outliers and data imbalance problems, and optimizes model parameters to improve prediction accuracy.
It significantly improves the accuracy of pneumoconiosis risk prediction, identifies and quantifies the contribution of various risk factors, supports personalized intervention strategies and diagnosis and treatment plans, and reduces the labor capacity and economic losses caused by the occurrence of pneumoconiosis and adverse prognosis.
Smart Images

Figure CN118942686A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of machine learning and data processing, and in particular to a method for predicting multiple competing risks of pneumoconiosis prognosis based on deep learning. Background Art
[0002] Pneumoconiosis accounts for about 90.0% of occupational diseases in my country. Its occurrence and prognosis is an extremely complex issue, which is affected by multiple factors such as the patient's own clinical manifestations, environmental exposure, living habits, social conditions, and genetic susceptibility, and there may be complex nonlinear and high-dimensional interactions between various factors. On the one hand, previous studies did not consider the impact of multiple influencing factors on the occurrence and prognosis of pneumoconiosis. On the other hand, most of them adopted classical linear model assumptions such as regression models, which can only deal with low-dimensional interaction problems, limiting the development of pneumoconiosis risk prediction models in the direction of precision and in-depth development. Summary of the invention
[0003] The purpose of the present invention is to provide a method for predicting multiple competing risk outcomes of pneumoconiosis prognosis based on deep learning, which can predict the possible multiple outcomes of pneumoconiosis prognosis under the influence of multiple factors and their interactions.
[0004] To achieve the above object, the present invention adopts the following technical solution:
[0005] A method for predicting the prognosis of pneumoconiosis with multiple competing risks based on deep learning, comprising the following steps:
[0006] (1) Obtain basic data and build a database. The basic data include social status, living and hygiene habits, dust exposure factors, genetic susceptibility, disease status, examination indicators, and outcomes;
[0007] (2) Processing of basic data: Use interpolation to process outliers in the basic data, and then use synthetic minority oversampling technology to create a data set for the basic data; first, the data needs to be systematically preprocessed, including data cleaning and missing value filling. Then, the exposure data is normalized to eliminate the impact of inconsistent variable dimensions on the results. Finally, the exposure factors are preliminarily screened based on the univariate logistics or Cox risk regression model to control the interference of too many irrelevant variables on the model fitting effect;
[0008] (3) using the dataset constructed in step (2) to construct a pneumoconiosis outcome prognosis prediction model using a DNN neural network, and training the pneumoconiosis outcome prognosis prediction model to obtain a final prediction model;
[0009] A deep recurrent neural network (DNN) was used with the attention mechanism to learn and construct a multi-outcome competing risk prediction model for the prognosis of pneumoconiosis. Among them, various risk factors were the input layer, and the multi-outcome competing risk of pneumoconiosis was the output layer. The loss function was defined as the likelihood function of the competing risk survival time. The gradient backpropagation algorithm was used for parameter optimization. The existing data was used as the training set and cross-validation of parameter estimation was performed. New cases that were continuously included in the later stage were used as the validation set for external verification, and the parameters were dynamically optimized according to the applicability of the model.
[0010] Furthermore, the basic data originated from the pneumoconiosis medical big data platform. Through a systematic literature review and combined with expert consultation, a risk factor relationship network that may affect the prognosis of pneumoconiosis was constructed. Based on the potential risk factors suggested by the risk factor relationship network, data was captured in the pneumoconiosis medical big data platform. The missing variable information was further improved through telephone follow-up and supplementary collection.
[0011] Furthermore, the processing of the basic data includes the following steps:
[0012] (21) The outliers were set as null values, which means that the outliers were treated as missing values, and the null values were interpolated using SPSS software to obtain new basic data;
[0013] (22) Create a data set for the new basic data using SMOTE technology;
[0014] In the basic data, for a sample a in the minority class, the distance from it to the remaining samples b in the minority class sample set is calculated using the Euclidean distance as the standard. Where Sa and Sb are two risk factor vectors, which contain the values of all risk factor variables;
[0015] (23) The k samples of the minority class with the closest calculated distance are set as c(near), and their vector is recorded as Sc(near), near∈{1,…,k}; then the difference in risk factors between sample c(near) and sample a is calculated as differ=Sc(near)-Sa; then a sample ci(near) is randomly selected from the k nearest neighbors, and a random number ζ between 0 and 1 is generated to synthesize a new sample Sc(new). For each randomly selected neighbor, a new sample S is constructed. c(new) =S a +ζ(S c(near) -S a );
[0016] (24) All data sets are constructed according to the above steps. Assume that the original data set has X patients who died and Y patients who did not die. For each patient who died, repeat steps (21) to (23) M times, and a total of X×M patients who died will be added. At this time, there are a total of X×(1+M) patients who died. The steps for patients who did not die are the same as above, and finally a data set with a ratio of 1:1 between patients who died from pneumoconiosis and patients who did not die is formed.
[0017] Furthermore, the final prediction model is implemented using the following steps:
[0018] (31) After some factors in the data set are transformed into more intuitive influencing factors by exponential smoothing method, a new data set is formed; the data in the data set are randomly divided into a training set and a test set, and the ratio of the training set to the test set is 8:2;
[0019] (32) The training set is input into the pneumoconiosis outcome prognosis prediction model constructed by using the DNN neural network, and various hyperparameters are set to start iterative training; after continuous iterations, a table is plotted to show the loss and accuracy. The loss continues to decrease, and the accuracy continues to increase. After both reach a stable value, the prediction model consisting of the weight w and bias b corresponding to the minimum loss and the highest accuracy is finally obtained;
[0020] (33) The data in the test set were input into the trained DNN neural network to construct a pneumoconiosis outcome prognosis prediction model to obtain the final result.
[0021] Furthermore, the keras library was used to build a DNN fully linked neural network structure, and various hyperparameters were set. The hidden layer was set to 6 layers, and the Relu activation function was added to the output of each hidden layer. Finally, the softmax function was added to the output layer for output. The batch size was set to 10, the Adam optimizer was used as the optimizer, the sparse classification cross entropy loss was used as the loss, and the number of epoch iterations was set to 100.
[0022] The present invention has the following beneficial effects:
[0023] (1) The original data set is first processed for outliers and then SMOTE is used to create a new data set, which handles the problems of incomplete and unbalanced data. At the same time, some factors in the new SMOTE data set are processed using exponential smoothing to make the factors more intuitive and obvious. A new prediction model is then built in combination with the DNN network. Compared with traditional data processing and linear models, its prediction ability is significantly improved.
[0024] (2) Identify, quantify and rank the contributions of various risk factors to the onset and prognosis of pneumoconiosis, and construct a precise prediction model to lay the foundation for the dynamic evaluation and adjustment of the diagnosis and treatment plan for subsequent personalized intervention strategies for pneumoconiosis, effectively prevent the occurrence of pneumoconiosis, and reduce the loss of labor capacity and economic benefits of pneumoconiosis patients by managing their adverse prognostic factors. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION
[0026] like Figure 1 As shown, the present embodiment provides a method for predicting the prognosis of pneumoconiosis based on deep learning and multiple competing risks, including the following steps:
[0027] (1) Obtain basic data and build a database. The basic data include social status, living and hygiene habits, dust exposure factors, genetic susceptibility, disease status, examination indicators, and outcomes;
[0028] Social conditions include personality, ethnicity, age level, medical insurance payment method, unit attributes, residential location, income level, education level, etc.; living and hygiene habits include smoking, exercise, nutritional status, etc.; genetic susceptibility includes gene polymorphism, histone modification status, epigenetics; dust exposure factors include dust type, dust exposure time, dust concentration, particle size and dispersion, personal protection, dust exposure age, occupational health examination, etc.; examination indicators include blood routine, liver function, lung function, chest imaging performance, infection indicators, etc.; morbidity includes pneumoconiosis stage, underlying disease, complications, acute onset, number of hospitalizations, standardized medication, etc.; outcomes include improvement, deterioration, disability, complications, and death.
[0029] The data of this embodiment comes from the medical big data platform for pneumoconiosis. Through systematic literature review and combined with expert consultation, a risk factor relationship network that may affect the prognosis of pneumoconiosis is constructed. Then, based on the potential risk factors suggested by the risk factor relationship network, data is captured in the medical big data platform for pneumoconiosis, and the missing variable information is further improved through telephone follow-up and supplementary collection. For different prognostic outcomes of pneumoconiosis (including improvement, deterioration, disability, complications and death), telephone or outpatient follow-up is used for continuous and dynamic collection, so as to form the required exposure and outcome data.
[0030] (2) Processing of the basis: Use interpolation to process outliers in the basic data, and then use synthetic minority class oversampling technology to create a data set for the basic data; specifically, it includes:
[0031] (21) The outliers are set as null values, which means that the outliers are treated as missing values. The null values are interpolated using SPSS software to obtain new basic data.
[0032] (22) Create a data set for the new basic data using SMOTE technology;
[0033] In the basic data, for a sample a in the minority class, the distance from it to the remaining samples b in the minority class sample set is calculated using the Euclidean distance as the standard. Among them, Sa and Sb are two risk factor vectors, which contain the values of all risk factor variables.
[0034] (23) The k samples of the minority class with the closest calculated distance are set as c(near), and their vector is recorded as Sc(near), near∈{1,…,k}; then the difference in risk factors between sample c(near) and sample a is calculated as differ=Sc(near)-Sa; then a sample ci(near) is randomly selected from the k nearest neighbors, and a random number ζ between 0 and 1 is generated to synthesize a new sample Sc(new). For each randomly selected neighbor, a new sample S is constructed. c(new) =S a +ζ(S c(near) -S a ).
[0035] (24) All data sets are constructed according to the above steps. Assume that the original data set has X patients who died and Y patients who did not die. For each patient who died, repeat steps (21) to (23) M times, and a total of X×M patients who died will be added. At this time, there are a total of X×(1+M) patients who died. The steps for patients who did not die are the same as above, and finally a data set with a ratio of 1:1 between patients who died from pneumoconiosis and patients who did not die is formed.
[0036] This embodiment first needs to preprocess the data systematically, including data cleaning, missing value filling, etc., then normalize the exposure data to eliminate the impact of inconsistent variable dimensions on the results, and finally, preliminarily screen the exposure factors based on the univariate logistics or cox risk regression model to control the interference of too many irrelevant variables on the model fitting effect. In order to retain variables with predictive information as much as possible, this embodiment mainly intends to use the method of controlling the false discovery rate as the variable screening criterion.
[0037] (3) Using the dataset constructed in step (2) to construct a pneumoconiosis outcome prognosis prediction model using a DNN neural network, and training the pneumoconiosis outcome prognosis prediction model, the specific steps are as follows:
[0038] (31) After some factors in the data set are transformed into more intuitive influencing factors by exponential smoothing method, a new data set is formed; the data in the data set are randomly divided into a training set and a test set, and the ratio of the training set to the test set is 8:2;
[0039] (32) The training set was input into the pneumoconiosis outcome prognosis prediction model constructed using the DNN neural network, and various hyperparameters were set to start iterative training; the DNN fully linked neural network structure was built using the keras library, various hyperparameters were set, the hidden layer was set to 6 layers, and the Relu activation function was added to the output of each hidden layer. Finally, the softmax function was added to the output layer for output, the batch size was set to 10, the Adam optimizer was used as the optimizer, the sparse classification cross entropy loss was used as the loss, and the number of epoch iterations was set to 100.
[0040] After continuous iterations, a table is plotted to show the loss and accuracy. The loss keeps decreasing and the accuracy keeps rising. When both reach a stable value, it indicates that the training effect has reached the best. Finally, the prediction model consisting of weights w and bias b corresponding to the minimum loss and the highest accuracy is obtained.
[0041] (33) The data in the test set were input into the trained DNN neural network to construct a pneumoconiosis outcome prognosis prediction model to obtain the final result.
[0042] Since the medical big data of pneumoconiosis has the characteristics of longitudinal data, this embodiment adopts a deep recurrent neural network DNN and uses the attention mechanism to learn and construct a multi-outcome competing risk prediction model for the prognosis of pneumoconiosis. Among them, various risk factors are the input layer, the multi-outcome competing risk of pneumoconiosis is the output layer, the loss function is defined as the likelihood function of the competing risk survival time, and the parameter optimization adopts the gradient backpropagation algorithm. In order to ensure the internal and external consistency of the data, this project intends to use the existing data as the training set and perform cross-validation of parameter estimation, and the new cases continuously included in the later stage will be used as the verification set for external verification, and the parameters will be dynamically optimized according to the specific applicability of the model.
[0043] The above description is only a preferred implementation manner of the present invention, but the protection scope of the present invention is not limited thereto, and any modification and replacement based on the technical solution and inventive concept provided by the present invention should be included in the protection scope of the present invention.
Claims
1. A method for predicting the prognosis of pneumoconiosis with multiple competing risks based on deep learning, characterized in that: The steps include: (1) Obtain basic data and build a database. The basic data include social status, living and hygiene habits, dust exposure factors, genetic susceptibility, disease status, examination indicators, and outcomes; (2) Processing of basic data: Use interpolation to process outliers in the basic data, and then use synthetic minority oversampling technology to create a data set for the basic data; first, the data needs to be systematically preprocessed, including data cleaning and missing value filling. Then, the exposure data is normalized to eliminate the impact of inconsistent variable dimensions on the results. Finally, the exposure factors are preliminarily screened based on the univariate logistics or Cox risk regression model to control the interference of too many irrelevant variables on the model fitting effect; (3) using the dataset constructed in step (2) to construct a pneumoconiosis outcome prognosis prediction model using a DNN neural network, and training the pneumoconiosis outcome prognosis prediction model to obtain a final prediction model; A deep recurrent neural network (DNN) was used with the attention mechanism to learn and construct a multi-outcome competing risk prediction model for the prognosis of pneumoconiosis. Among them, various risk factors were the input layer, and the multi-outcome competing risk of pneumoconiosis was the output layer. The loss function was defined as the likelihood function of the competing risk survival time. The gradient backpropagation algorithm was used for parameter optimization. The existing data was used as the training set and cross-validation of parameter estimation was performed. New cases that were continuously included in the later stage were used as the validation set for external verification, and the parameters were dynamically optimized according to the applicability of the model.
2. The method for predicting the prognosis of pneumoconiosis based on deep learning and multiple competing risks according to claim 1, characterized in that: The basic data are derived from the pneumoconiosis medical big data platform. A risk factor relationship network that may affect the prognosis of pneumoconiosis is constructed through a systematic literature review and combined with expert consultation. Based on the potential risk factors suggested by the risk factor relationship network, data are captured in the pneumoconiosis medical big data platform. The missing variable information is further improved through telephone follow-up and supplementary collection.
3. The method for predicting the prognosis of pneumoconiosis based on deep learning and multiple competing risks according to claim 1, characterized in that: The processing of basic data comprises the following steps: (21) The outliers were set as null values, which means that the outliers were treated as missing values, and the null values were interpolated using SPSS software to obtain new basic data; (22) Create a data set for the new basic data using SMOTE technology; In the basic data, for a sample a in the minority class, the distance from it to the remaining samples b in the minority class sample set is calculated using the Euclidean distance as the standard. Where Sa and Sb are two risk factor vectors, which contain the values of all risk factor variables; (23) The k samples of the minority class with the closest calculated distance are set as c(near), and their vector is recorded as Sc(near), near∈{1,…,k}; then the difference in risk factors between sample c(near) and sample a is calculated as differ=Sc(near)-Sa; then a sample ci(near) is randomly selected from the k nearest neighbors, and a random number ζ between 0 and 1 is generated to synthesize a new sample Sc(new). For each randomly selected neighbor, a new sample S is constructed. c(new) =S a +ζ(S c(near) -S a ); (24) All data sets are constructed according to the above steps. Assume that the original data set has X patients who died and Y patients who did not die. For each patient who died, repeat steps (21) to (23) M times, and a total of X×M patients who died will be added. At this time, there are a total of X×(1+M) patients who died. The steps for patients who did not die are the same as above, and finally a data set with a ratio of 1:1 between patients who died from pneumoconiosis and patients who did not die is formed.
4. The method for predicting the prognosis of pneumoconiosis based on deep learning and multiple competing risks according to claim 1, characterized in that: The final prediction model is implemented using the following steps: (31) After some factors in the data set are transformed into more intuitive influencing factors by exponential smoothing method, a new data set is formed; the data in the data set are randomly divided into a training set and a test set, and the ratio of the training set to the test set is 8:2; (32) The training set is input into the pneumoconiosis outcome prognosis prediction model constructed by using the DNN neural network, and various hyperparameters are set to start iterative training; after continuous iterations, a table is plotted to show the loss and accuracy. The loss continues to decrease, and the accuracy continues to increase. After both reach a stable value, the prediction model consisting of the weight w and bias b corresponding to the minimum loss and the highest accuracy is finally obtained; (33) The data in the test set were input into the trained DNN neural network to construct a pneumoconiosis outcome prognosis prediction model to obtain the final result.
5. The method for predicting the prognosis of pneumoconiosis based on deep learning and multiple competing risks according to claim 4, characterized in that: Use the keras library to build a DNN fully linked neural network structure, set various hyperparameters, set the hidden layer to 6 layers, add the Relu activation function to the output of each hidden layer, and finally add the softmax function to the output layer for output. The batch size is set to 10, the Adam optimizer is used as the optimizer, the sparse classification cross entropy loss is used as the loss, and the number of epoch iterations is set to 100.
