| With 30 years of reform and opening to the rapid economic development, China's growing health care costs, 2007 and 2008, total health expenditure in China was 1.0966 trillion yuan, respectively, and 1.45354 trillion yuan, according to the "2009 Statistical Bulletin of China's health development," is expected to 2009 National Health The total costs amounted to 1.6119 trillion yuan, 1,192 yuan per capita health expenditure. Rapid growth in health care costs become a common problem, how to solve this problem, has aroused great concern. Effective control of rising health care costs, particularly the rising cost of hospitalization, first of all need factors of hospital costs the right way. Previous studies on the influencing factors of hospitalized patients reported more researchers use more appropriate tools of mathematical statistics on a variety of hospital costs data analysis and processing, including: using different multivariate statistical analysis techniques to explore the impact of factors of hospital costs , and to predict hospital costs; on the different types of classified and other hospital charges. The traditional regression analysis of hospital charges in the study there is some limitations. BP neural network can be due to its linear or nonlinear multivariable without preconditions in the case of statistical analysis, with the traditional statistical methods need to be analyzed to meet certain conditions compared to the variable has its own advantages, so more and more applied research hospital costs.Artificial neural network, is a simulation of biological neurons work the mathematical model, in many disciplines are widely used in pattern recognition, forecasting and prediction, numerical approximation and so on. Back-propagation, BP neural network is a neural network, it is a multi-layer perceptron network weights due to the adjustment of the rules using error back propagation algorithm, named after the BP algorithm is a neural network the current development of the most mature, most widely used network model. BP neural network for the type of data, no distribution requirements, and has certain fault tolerance, through self-study, self-adjustment of the input variables and output variables of the complex mapping, so you can consider using BP neural network its impact on hospital costs in the model factors, hospital costs and its influencing factors to achieve the relationship between the fitting.ObjectiveIn this study, Tangshan City, two general hospitals provided information on hospital costs, were established BP neural network model and multiple linear regression model, and the cost of hospitalization for conditions in the prediction accuracy of a comparative analysis of two different data distribution and promotion of capacity-type fitting, and then demonstrated the correct prediction of hospital charges. Meanwhile, BP neural network modeling process of the optimization of model parameters. In case of hospitalization costs for the continuous variable dependent variable, independent variables for the categorical variables or continuous variables and categorical variables are independent variables and the dependent variable and non-linear relationship between the application of statistical methods.Method The data from the Joint University Hospital of Hebei provided from 2006 to 2010, A group of acute appendicitis (ICD10 code K35) and B group with diabetes (ICD10 code E11) the patient's medical record information, a total number of cases of 2317 copies.Using BP neural network and multiple linear regression functions were fitted to model cost of hospitalization in the modeling process using T test for different statistical methods to compare different data distribution of data analysis, has been established based on BP neural network model sensitivity analysis on the use of impact factors of hospital costs. Above T test and multiple linear regression in SPSS software for, BP neural network modeling and sensitivity analysis were carried out by MATLAB programming software can be realized. ResultFirst, the model performance results The data from different data distribution BP neural network and multiple linear regression modeling, the following results: the results of acute appendicitis: BP training set R = 0.79, R~2 = 0.785; test set R = 0.72, R~2 = 0.704; diverse training set R = 0.598, R~2 = 0.357. Diabetes results: BP training set R = 0.96, R~2 = 0.869; test set R = 0.51, R~2 = 0.499; diverse training set R = 0.685, R~2 = 0.443.Second, BP neural network and multiple linear regression results of the comparison through the fitting and predictive value of the difference between the real value of the relative error obtained by comparing paired T test results are as follows: acute appendicitis in the training set T = 3.63, P = 0.000; test set T = 1.31, P = 0.190, diabetes training Set T = 9.65, P = 0.000; test set T =- 5.29, P = 0.000.Third, BP neural network data distribution in different types of comparison Through the fitting and predictive value of the difference between the real value of the relative error obtained by comparing the two cases of a more independent samples T test results are as follows: the training set T = 1.18, P = 0.237; test set T =- 1.34, P = 0.181 . Mean relative error of training set: 0.31 3.60 Acute appendicitis, diabetes 0.12 1.08 0.183 difference estimate; test set the mean relative error: 0.364 0.553 acute appendicitis, diabetes, 0.52 1.54, the difference estimate -0.156.Fourth, the results of sensitivity analysis Sensitivity analysis showed that the sensitivity of the factors in decreasing order of: acute appendicitis were operated cases (0.27), length of stay (0.224), treatment outcome (0.157), hospitalization (0.14), secondary diagnosis (0.058 ), admission illness (0.048), age (0.037), career (0.033), costs do not (0.025), gender (0.008), indicating that the greatest impact on hospital costs factor is the surgical case, the smallest gender. Length of stay in diabetic group (0.348), surgical conditions (0.162), age (0.131), hospitalization (0.073), treatment outcome (0.072), secondary diagnosis (0.062), career (0.051), costs do not (0.045), admission condition (0.038), gender (0.019), indicating that the greatest impact on hospital costs factor is the length of hospital stay, the smallest gender.ConclusionThe study uses BP neural network to fit the model of hospital costs, through the choice of different model parameters can be optimized to set the role model. BP neural network and multiple linear regression fitting and predictive ability compared to better, more stable performance. Discrete data distribution in the BP neural network because too many samples to learn the details of predictive ability decreased. Fitting ability of neural networks and generalization can not be reconciled, in the modeling should be based on the actual situation on the trade-off between the two. BP neural network model based on sensitivity analysis can be used for the analysis of the influencing factors of hospitalization. BP neural network model based on hospital charges for forecasting and influencing factors, the data type of linear model demanding nonlinear neural network to deal with the problem to overcome weaknesses in regression models for nonlinear problems and the Influencing Factors analysis provides an optimal method of modeling.Because this study selected the medical record information is limited, therefore the model prediction accuracy of hospital costs is limited, can provide a theoretical reference. BP neural network due to the different parameter settings on the network have different effects on the results, such as the different number of neurons in the hidden layer, different initial weights and thresholds, etc., is still a lack of theoretical support, pending further study. |
PDF Full Text Request