Medical resource allocation fairness analysis method based on multi-source big data and regression model
By using a multi-source heterogeneous big data and MLP-RF combined regression model in the fairness analysis of medical resource allocation, the problems of insufficient accuracy and low cost and efficiency of medical resource allocation in the existing technology are solved, and the fine-grained fairness assessment of medical resource allocation is achieved, which improves the flexibility and reliability of evaluation.
Patent Information
- Application Number
- CN202411701354.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-26
- Publication Date
- 2025-03-14
AI Technical Summary
The existing fairness analysis method for the allocation of medical resources cannot accurately reflect the spatial differences in medical resources and their correspondence with population distribution, and the method based on administrative region statistical data has problems such as insufficient accuracy and low cost efficiency.
The fine-grained fairness analysis method of medical resource allocation based on multi-source heterogeneous big data and MLP-RF combined regression model is adopted. The fine-grained analysis of medical resource allocation is achieved through preprocessing of maps and medical data, analysis of optimal grid slicing granularity, construction of regional population prediction models, and construction of medical resource evaluation models.
It significantly improves the flexibility and reliability of the fairness assessment of medical resource allocation, can more accurately reflect the spatial differences and population distribution relationships of medical resources, and provides the government with more intuitive and effective reference for resource allocation decisions.
Smart Images
Figure CN119622661A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of fairness analysis of medical resource allocation Background Art
[0002] Fairness in the allocation of medical resources is an important foundation for protecting everyone's right to health. It is related to the convenience and satisfaction of the people in seeking medical treatment, and also reflects social fairness and justice. Ensuring that everyone, especially vulnerable groups and residents in remote areas, can enjoy equal medical services will help achieve public health goals and improve the overall health level of society.
[0003] Generally speaking, research on the fairness of medical resource allocation is mainly carried out through statistical panel data at the provincial and municipal levels, and the regional fairness of medical resources is analyzed using methods such as the Lorenz curve, the Theil index, and the Gini coefficient based on administrative districts.
[0004] However, the existing methods for analyzing the fairness of medical resource allocation generally use statistical yearbook data or questionnaire and survey data released by the government, and only analyze data related to medical resources, which cannot accurately reflect the spatial differences of medical resources and their corresponding relationship with population distribution. At the same time, the use of field surveys or questionnaire surveys obviously has problems such as high cost, low efficiency, and limited coverage; at the same time, the existing methods are mainly based on statistical panel data based on provinces, cities, and counties, and the annual geographic vector data released by the government can only be accurate to the county level. Its accuracy is difficult to support the needs of fairness research on medical services and resource allocation in towns (townships) and villages or even finer granularity. Summary of the invention
[0005] The present invention provides a fine-grained fairness analysis method for medical resource allocation based on multi-source heterogeneous big data and MLP (Multi-Layer Perception)-RF (Random Forest) combined regression model, which effectively overcomes the problem that traditional methods are limited by administrative district statistical data and significantly improves the flexibility and reliability of evaluation.
[0006] In an embodiment according to the present disclosure, the specific scheme of the fine-grained fairness analysis method for medical resource allocation based on multi-source heterogeneous big data and MLP-RF combined regression model is as follows:
[0007] A method for analyzing fairness of medical resource allocation based on multi-source big data and regression models, including a data processing and analysis structure and a model building structure, the method comprising the following steps:
[0008] Step 1: Preprocessing of maps and medical data to obtain medical resource data associated with map data. This step ultimately constructs the training dataset for population prediction, the dataset to be predicted, and the comprehensive dataset for medical resources.
[0009] Step 2: Analysis of the optimal grid segmentation granularity, combining data statistics with actual needs and application scenarios to determine the optimal grid segmentation granularity. This step can help choose the appropriate grid segmentation granularity, which is the key to evaluating the accuracy and reliability of medical resource allocation. Choosing the most appropriate grid segmentation granularity can improve the stability and reliability of the analysis method.
[0010] Step 3: Build a regional population prediction model based on multi-source data and the MLP-RF model, and use the model to predict the population data of the grid. This step ultimately builds a high-performance regional population prediction model.
[0011] Step 4: Construct a medical resource evaluation model to obtain the medical resource index of different grids.
[0012] Based on the above, this paper breaks the obstacle of the original medical resource evaluation method being limited by the statistical data of administrative regions at all levels, optimizes the data granularity of population statistics and medical resource statistics, and demonstrates flexibility and reliability. At the same time, the fine-grained analysis results of the configuration of regional medical institutions provide an intuitive and effective reference for the government to make further resource allocation decisions. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] Figure 1 A process framework diagram of the fine-grained fairness analysis method for medical resource configuration of the present invention;
[0014] Figure 2 A schematic diagram of a grid for determining the optimal granularity of the study area in the present invention;
[0015] Figure 3 It is the overall architecture diagram of the MLP-RF combined regression model of the present invention;
[0016] Figure 4 This is the MLP model architecture diagram of the present invention;
[0017] Figure 5 This is a diagram of the RF model architecture of the present invention. DETAILED DESCRIPTION
[0018] In order to better understand the purpose, structure and function of the present invention, the following is a further detailed description of the fairness analysis method of medical resource allocation based on multi-source big data and regression model in conjunction with the accompanying drawings. The overall process framework of the method of the present invention is as follows Figure 1 As shown:
[0019] Step 1: Preprocessing of maps and medical data. The preprocessing process is divided into administrative area data extraction, map grid division and grid data extraction, and medical institution data processing.
[0020] Administrative area data extraction: First, extract the night light remote sensing data, and clip the VIIRS night light remote sensing data set to the administrative divisions of Zhejiang Province. Then, use the data zoning statistics tool to perform zoning statistics on the city-level administrative districts, county-level administrative districts, and township administrative districts, and obtain the sum of the night light attribute values of each unit to form the regional night light data set corresponding to the administrative district unit. Next, process the population activity data. First, it is necessary to convert the point data of the population activity data into raster data, where the value of the raster represents the population activity intensity value. Then, use the zoning statistics tool again to perform zoning statistics on each administrative district, and obtain the sum of the population activity values of each unit as the regional population activity data set corresponding to the administrative district unit. Finally, associate the night light data with the population activity data, and add permanent population data for each region based on the statistical yearbook and the seventh census data. The final result is used as the training data set for the machine learning model.
[0021] Map grid division and grid data extraction: Mainly using the fishing net tool, we generated 10 sets of grid data in Zhejiang Province, ranging from 1KM*1KM to 10KM*10KM. Subsequently, we analyzed different grid granularities and selected the optimal granularity as the basic research scale of this study. Different from the previous data extraction based on administrative districts, the next step is to extract night light remote sensing data, population activity data, residential POI data, and medical resource data based on the divided grids. Similarly, these data are subjected to correlation analysis, and the final results are used as the prediction data set for the machine learning model.
[0022] Medical institution data processing: The POI data obtained from the map data platform contains unavailable data and duplicate data, which need to be processed. Unavailable data refers to data related to animal medical services and health care services (such as skin care, massage) in the POI data. These data have a low correlation with the medical resources involved in this study and need to be eliminated. The reason for the appearance of duplicate data is that different buildings belonging to the same medical institution are also regarded as an independent medical institution, which often appears in hospitals of a certain scale. The distance between medical institutions is calculated according to the longitude and latitude, and the duplicate data is processed with manual verification.
[0023] Step 2: Analyze the optimal grid segmentation granularity, combining data statistics with actual needs and application scenarios to determine the optimal grid segmentation granularity.
[0024] The number of hospital beds in the grid area is used as a feature to measure the richness of medical resources in the grid. The standard deviation, skewness coefficient and zero content rate are combined to obtain an optimal grid granularity. The above statistical performances in ten groups of different grid scales from 1KM to 10KM are compared. The results are as follows: Figure 2As shown. Among them, the standard deviation is a measurement concept of the degree of dispersion of a set of values from the mean. A larger standard deviation means that most of the values are more different from their mean, and a smaller standard deviation means that these values are closer to the mean. The skewness coefficient is a characterization of the direction and degree of distribution skewness, also known as the deviation coefficient, which is used to describe the asymmetry of data distribution. The larger the absolute value of the skewness coefficient, the greater the degree of data deviation, that is, the more serious the skewness. The zero content rate is the ratio of the number of grids with 0 beds to the total number of grids. Finally, an optimal grid division granularity is determined, and the subsequent steps of the method use the grid of this granularity as the research unit, such as Figure 2 shown.
[0025] Step 3: Based on multi-source data and the MLP-RF model, a regional population prediction model is constructed. The model is used to predict the population data of the grid. The overall architecture of the regional population prediction model based on multi-source data and the MLP-RF model is as follows: Figure 3 shown.
[0026] The task of the model is to predict the permanent population of a given area based on the night light data, population activity data and residential POI data of the data set. The training data set of the model is the night light data, population activity data, residential POI data and permanent population in the administrative area. Finally, the corresponding permanent population of each grid needs to be predicted based on the data of the mapping data set. The MLP-RF model is a combined regression model composed of MLP and RF.
[0027] Among them, the MLP part is a basic feedforward artificial neural network structure. This part consists of multiple neuron layers, each layer is fully connected to the next layer. In addition to the input layer and the output layer, there can be multiple hidden layers in the middle of the model. As one of the foundations of deep learning, the MLP model has a wide range of applications in solving prediction problems. Compared with a linear model with only one hidden layer, the MLP model can better capture the complex nonlinear relationship between input features, thereby improving prediction performance. The random forest part consists of multiple decision trees, each of which is trained for prediction, and the final prediction result is determined by the set of these decision trees. When constructing each decision tree, features are randomly selected to ensure that each decision tree is trained based on a different subset of features, increasing the diversity of the model.
[0028] The architecture diagram of the MLP module is as follows Figure 4As shown in the figure, in the MLP module, the previously constructed dataset is fed as input features into a linear layer of the MLP, which uses the number of input features as the input size and maps it to a hidden representation with an intermediate dimension of 128. Next, a ReLU activation function is used to introduce nonlinear characteristics to capture the complex relationships in the input features. Subsequently, the intermediate hidden layer is further connected to three hidden layers with 256 neurons. Each hidden layer adopts the same structure, which effectively captures the abstract features in the data through linear transformation and ReLU activation function. Finally, the output of the hidden layer is mapped to a single output node through the last linear layer. The loss function is then used to calculate the gradient of the loss with respect to the weights of each layer. Starting from the output layer, the gradient is back-propagated to update the weights and biases of the model so that the loss gradually decreases. Subsequently, the training dataset is associated with the results of the MLP model to form a new dataset used by the RF module. The architecture diagram of the RF module is shown in the figure. Figure 5 As shown in the figure, after training 100 decision trees, the prediction result of the RF module will be determined by the average output of these decision trees. The prediction result of the random forest is the final result of the MLP-RF combined regression model.
[0029] At the same time, in order to prevent overfitting problems during training, the K-fold cross-validation method is introduced. The specific steps are as follows: First, the data set is evenly divided into K subsets according to the amount of data. The training and evaluation of the model are repeated K times, each time a different subset is selected as the test set, and the other subsets are used as the training set. In each iteration, the model is trained using the training set, and the performance indicators of the model are evaluated on the corresponding test set. Finally, the performance indicators evaluated in the K iterations are averaged as the final performance evaluation result of the model.
[0030] The learning rate of the MLP model was initially set to 0.01, and after 200 iterations, a large fluctuation in loss was observed. The learning rate was then reduced to 0.005, but loss fluctuations were still observed. Subsequently, the learning rate was further reduced to 0.000005, which made the loss drop smoother, but the final loss value increased. This indicates that further adjustments to the learning rate and number of iterations are needed. Finally, the learning rate was set to 0.00002 and the number of iterations was increased to 400, and the model loss finally stabilized.
[0031] Step 4: Construct a medical resource evaluation model, and use the evaluation model to obtain the medical resource index of different grids. The construction process of the medical resource evaluation model includes calculating the grid medical institution score, calculating the grid medical resource accessibility, calculating the grid medical resource index, calculating the grid medical resource index per thousand people, and calculating the grid medical resource concentration index.
[0032] Grid medical institution scoring calculation: The processed medical institution POI data is divided into different grades according to different institution types, medical resource richness and technical level for scoring, and a medical institution MI is defined. i Rated MIS i . Class III hospitals usually have a large number of beds, health technicians and advanced medical equipment, and have the highest richness in medical resources, so they have the highest score. Clinics and pharmacies have fewer resources and provide basic medical and drug needs for residents, so they have the lowest score. At the same time, this method is based on the medical and health data in the statistical yearbooks of each region, and adds the number of beds and health technicians to different types of POIs in each region. For a grid of medical institutions, the score, number of beds, and number of health technicians are equal to the sum of the medical attribute values of the POIs falling in the grid. A medical institution MI i The number of beds and health personnel are denoted as HB i and HT i , for Grid j , the sum of the scores of medical institutions in the grid is GMIS j The sum of the number of beds and the sum of health technicians in the grid are defined as GHB j and GHT j , the calculation formula is:
[0033]
[0034] Calculation of grid medical resource accessibility: Method Definition The grid medical accessibility index represents the average driving time from the grid center of each grid to the nearest tertiary hospital, general hospital and pharmacy, recorded as GMRA j , this value comprehensively considers the accessibility of medical resources in the grid to meet different levels of medical needs. The specific method is to use the geometric center of each divided grid as the starting point, use the map data platform to search, find the location of the medical institution closest to the starting point of the grid, and use this location as the target location. Then, through the path planning function of the map data platform, the driving time required from the starting point of each grid to the corresponding target location can be obtained. The shorter the driving time, the better the accessibility of the grid. This method calculates the average driving time from each grid center to the nearest tertiary hospital, general hospital and pharmacy, and obtains the overall average driving time.
[0035] Grid medical resource index calculation: This method converts a grid j The available medical resources are defined as the grid medical resource index, denoted as GMRI jThe higher the index, the richer the medical resources that the grid can enjoy. The previously calculated grid medical institution score, grid bed number, and grid health technician number are standardized, multiplied by the weight and summed, and finally multiplied by the standardized result of the grid medical resource accessibility to obtain the grid medical resource index. The calculation method is:
[0036] GMRI j =(ω MIS GMIS std +ω HB GHB std +ω HT ·GHT std ) GMRA std (4)
[0037] ω represents the weight of the standardized attributes of medical resources. The score of a medical institution usually reflects the quality of its facilities and services, including the level of medical technology and the quality of medical staff, so it should be given the highest weight of 0.4 in the comprehensive evaluation. The number of beds and technicians represent the actual service quality of medical resources and should also be considered as equally important indicators, with a weight of 0.3.
[0038] Since different medical resource attributes have different measurement units and ranges, in order to eliminate their influence, it is necessary to standardize the data. j GMIS Rating of Traditional Chinese Medicine Resources j , number of beds GHB j and the number of health technicians GHT j , the standardized calculation formula is:
[0039]
[0040] Among them, X min represents the minimum value of the medical resource attribute X, X max Indicates the maximum value of medical resource attribute X.
[0041] For Grid j Accessibility of Traditional Chinese Medicine Resources GMRA j For example, the corresponding standardized value GMRA std , the calculation formula is:
[0042]
[0043] Among them, GMRA min and GMRA max Represent the minimum and maximum values of the average reachability data respectively.
[0044] Calculation of the medical resource index per thousand people in the grid: The permanent population data of each grid (referred to as GP j ) is associated with the medical resource index of the grid to calculate the medical resource index per thousand people, recorded as GPTI j , which can be defined as the richness of medical resources that every thousand people in the grid can enjoy. The higher the index, the more medical resources per capita can enjoy. The calculation method is:
[0045]
[0046] Calculation of grid medical resource concentration: The maximum and average values of the medical resource index of the grids contained in each municipal administrative district in the study area are calculated. The medical resource concentration of a municipal administrative district is defined as the ratio of the maximum and average values of the medical resource index within the administrative district, which is recorded as CCMR. j , indicating the concentration of medical resources in the city-level administrative area. The higher the index is, the more concentrated the medical resources in the city-level administrative area are, and the greater the difference in resource allocation is. The calculation method is:
[0047]
[0048] Among them, GMRI max Indicates the maximum value of the GMRI of the grids contained in the city-level administrative area. avg It represents the average value of the grid GMRI contained in the city-level administrative area after removing zero values.
[0049] It is to be understood that the present invention is described by some embodiments, and it is known to those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the scope of protection of the present invention.
Claims
1. A method for analyzing fairness of medical resource allocation based on multi-source big data and regression model, characterized in that: The medical resource allocation fine-grained fairness analysis method includes: Step 1: Preprocessing of maps and medical data to obtain medical resource data associated with map data; Step 2: Analyze the optimal grid segmentation granularity, combining data statistics with actual needs and application scenarios to determine the optimal grid segmentation granularity; Step 3: Construct a regional population prediction model based on multi-source data and the MLP-RF model. The regional population prediction model predicts the population data of the grid; Step 4: Construct a medical resource evaluation model to obtain the medical resource index of different grids.
2. The fine-grained fairness analysis method for medical resource allocation according to claim 1 is characterized in that: The step 1 includes step 1.1: Obtain map data from the National Geographic Information Resource Directory Service System and Alibaba Cloud DataV Map Platform, and use the map data of Zhejiang Province and the administrative regions within the province as the data analysis unit; To obtain population data, the permanent population data of Zhejiang Province and its administrative regions in the study area were obtained from the statistical yearbooks published by various administrative regions in Zhejiang Province and the seventh population census data in the EPS database; to obtain medical resource data, the number of beds and the number of health technicians used were also mainly derived from the statistical yearbooks of various regions in Zhejiang Province; Obtain night light remote sensing data, which comes from the VIIRS night light remote sensing dataset; Obtain population activity data, which comes from Baidu Map Smart Population Geography Analysis System; Obtain POI data, obtain residential area POI and medical resource POI data, the data source is Baidu map data platform.
3. The fine-grained fairness analysis method for medical resource allocation according to claim 2 is characterized in that: The step 1 includes step 1.2: Extraction of comprehensive data of administrative regions: extract night light remote sensing data, population activity data and POI data from the map data of each administrative region, add permanent population data, and finally integrate the comprehensive data as the training data set for the subsequent machine learning model; Map grid division and grid comprehensive data extraction, grid division of the administrative area of Zhejiang Province, generating grid granularity from 1KM*1KM to 10KM*10KM in the study area Zhejiang Province, a total of 10 groups of grid data, and extracted night light remote sensing data, population activity data and POI data respectively, and added permanent population data, and finally integrated comprehensive data as the prediction data set for subsequent machine learning models.
4. The fine-grained fairness analysis method for medical resource allocation according to claim 3 is characterized in that: The step 2 comprises: The number of hospital beds in the grid area is used as a characteristic to measure the richness of medical resources in the grid. Combined with the standard deviation, skewness coefficient and zero rate, the statistics of ten groups of different grid granularities from 1KM to 10KM are compared. The optimal grid granularity is obtained by combining the statistical results and application scenarios.
5. The fine-grained fairness analysis method for medical resource allocation according to claim 4 is characterized in that: The step 3 includes step 3.1: Based on the established training data, the correlation between the selected features and the data of the resident population was analyzed, and the Pearson correlation coefficient was used to measure the strength of the correlation between the response variables; Draw a scatter plot of the training features and labels, and remove extreme abnormal data that obviously do not meet the linear relationship between the features and labels. By analyzing the distribution of feature values, some extremely abnormal values are adjusted or removed, and adjustments and corrections are made based on statistical yearbooks and census data.
6. The fine-grained fairness analysis method for medical resource allocation according to claim 5 is characterized in that: The step 3 includes step 3.2: Construct an MLP-RF combined regression model, which is composed of an MLP model and an RF model. In the MLP module, the previously constructed dataset is fed into a linear layer of the MLP as input features. This layer uses the number of input features as the input size and maps it to a hidden representation with an intermediate dimension of 128. Then, a ReLU activation function is used to introduce nonlinear characteristics to capture the complex relationships in the input features. Subsequently, the intermediate hidden layer is further connected to three hidden layers with 256 neurons. Each hidden layer adopts the same structure, and the abstract features in the data are effectively captured through linear transformation and ReLU activation function. Finally, the output of the hidden layer is mapped to a single output node through the last linear layer, and then the loss function is used to calculate the gradient of the loss to the weights of each layer. Starting from the output layer, the gradient is back-propagated to update the weights and biases of the model to gradually reduce the loss. Subsequently, the training dataset is associated with the results of the MLP model to form a new dataset used by the random forest module. After training 100 decision trees, the prediction results of the random forest module will be determined by the average output of these decision trees. The prediction results of the random forest are used as the final results of the MLP-RF combined regression model.
7. The fine-grained fairness analysis method for medical resource allocation according to claim 6 is characterized in that: The step 3 includes step 3.3: The total training data set was randomly divided into training set and validation set according to the ratio of 85% and 15%. In order to ensure the reliability of the results, the total data set was randomly divided three times according to the above method, and divided into three training sets and three validation sets; The three regression evaluation indicators, MAE, RMSE and RMSLE, are used as indicators to measure the accuracy of the model.
8. The fine-grained fairness analysis method for medical resource allocation according to claim 7 is characterized in that: The step 4 includes step 4.1: Establish a medical resource scoring model to score the processed POI data according to different institution types, medical resource richness and technical level; At the same time, the accessibility of medical resources is used as one of the characteristics of the distribution of medical resources in the study area. The geometric center of each divided grid is used as the starting point. The map data platform is used for retrieval to find the location of the medical institution closest to the starting point of the grid, and this location is used as the target location. Then, through the path planning function of the map data platform, the driving time required from each grid starting point to the corresponding target location can be obtained. The shorter the driving time, the better the grid accessibility.
9. The fine-grained fairness analysis method for medical resource allocation according to claim 8 is characterized in that: The step 4 includes step 4.2: The medical resources that a grid can enjoy are defined as the grid medical resource index. The higher the grid medical resource index, the richer the medical resources that the grid can enjoy. Different weights are assigned to the scores of medical institutions, the number of beds, and the number of technicians. The three data are integrated in each grid and standardized to obtain the grid medical resource index of each grid. The population predicted by the machine learning model is associated with the grid medical resource index to obtain a medical resource index per thousand people in a grid, which indicates the richness of medical resources available to every thousand people in the grid. The maximum and average values of the medical resource index of each grid within different municipal administrative regions are calculated, and the maximum value divided by the average value represents the concentration of medical resources in the municipal administrative region. The higher the data, the more concentrated the medical resources in the municipal administrative region and the greater the difference in resource allocation.