|
In recent years, medical costs soaring, high medical costs, particularly hospital costs to the people brought serious economic burden. Control excessive medical costs, optimize the allocation of health resources, the establishment of a multi-level medical security system in recent years has become a concern of the community hot spot. Control hospital costs should proceed from its influencing factors, through the establishment of an effective hospital costs fitting model, and its influencing factors on hospital costs analysis of the relationship, which is crucial to study hospital costs. The most commonly used method for the analysis of hospital cost modeling multiple linear regression, but its data have certain requirements, such as independence, normality, homogeneity of variance, linear, etc., and hospitalization expenses and its influencing factors may exist between non-linear relationship between the various factors that may exist multicollinearity, it should consider using another suitable hospital cost data modeling features. Artificial neural networks, referred to as neural networks, is a simulation of biological neurons work a mathematical model in a number of subject areas has been widely used in pattern recognition, predictions, numerical approximation and so on. Back-propagation (Back propagation, BP) neural network is a neural network in one, it is a multi-layer perceptron (multilayer perception), due to the network weights adjustment rules using error back propagation algorithm, namely BP algorithm and named, is a neural network is currently the most mature, the most widely used network model. As BP neural network for data types, distribution, etc. without any requirements, and has a certain fault tolerance, through self-learning, self-adjustment for input and output variables mapping relationship between the complex, so you can consider using BP neural network at its influencing factors for hospitalization expenses for modeling, hospitalization costs and its influencing factors to achieve the relationship between the fitting. Objective: In patients with cerebral infarction hospitalization cost analysis, for example, set the appropriate parameters to establish based on BP neural network hospital costs fitting model. In the established BP neural network model based on the influence of various factors on the cost of hospitalization to measure the degree of influence. In this study, the use of the BP neural network modeling methodology study provides a frame of reference, and through the hospitalization expenses and its influencing factors analysis helps health management and health insurance industry decision-makers to make the right decisions and analysis. Materials and Methods: Tangshan City, a three-level hospitals with cerebral infarction from 2007 to 2008 a total of 2538 cases of medical record information, excluding missing cases illogical cases, a total of 2218 cases of effective cases, 87.39% of the total number of cases . BP neural network modeling function fitting hospital costs, the modeling process using analysis of variance for different training algorithms, different number of hidden layer neurons comparative analysis, the model has been established based on the use of sensitivity analysis Factors Affecting the hospitalization expenses. The above analysis of variance in SAS software for, BP neural network modeling and sensitivity analysis are carried out through the MATLAB software programming can be achieved. Main results: one, modeling parameters of comparison results by the number of hidden layer neurons, respectively 10, 15 and 20 networks are used LM algorithm, BR algorithm, OSS algorithm, SCG algorithm for training, each network each Random training 100 times the number of hidden layer neurons obtained 10, 15 and 20, OSS fitting algorithm both in ability or capacity to promote optimum. Second, the establishment of model performance and parameters set by trying different algorithms, different number of neurons in the hidden layer and the different weights and closing the initialization value, the eventual establishment of a cerebral infarction and its influencing factors inpatient hospital costs of BP network model. Test set and the training set at 0 down residuals basic random fluctuations model fit better ability. Network structure of the model parameters are: single hidden layer, hidden layer 15 neurons, neurons in the input layer 8, an output layer neurons; network training parameters: use a combination of early stopping strategy OSS algorithm for training, set the learning speed of 0.01, set SSE error performance indicators for the network, the network of train stops 10 iterations, SSE reached 2.38631; training set fitting result is: R = 0.83954, R2 = 0.70483, Radj = 0.64783, RMSE = 0.05022; simulation results for the test set: R = 0.85523, R2 = 0.73142, Radj = 0.71649, RMSE = 0.04504. Third, the sensitivity analysis The sensitivity analysis showed that the sensitivity of each factor in descending order: Age (0.94897), length of stay (0.16101), treatment outcome (0.15227), the rescue frequency (0.14537), hospitalization ( 0.09421), marital status (0.08733), payment method (0.06751), gender (0.01391), showing the greatest impact on hospital costs factors are age, the smallest gender. Conclusion: Through this study, we can draw the following conclusions: BP neural network can be achieved by fitting model for hospital costs, through the selection of different model parameters can be set to optimize the role model; fitting ability of neural networks and promotion capabilities are not attained, when fitting ability good, because too many training samples to learn the information, the network's generalization ability will drop, so the model should be based on the actual situation on the balance between the two, a good neural network model should first have a good generalization ability, or even better fitting ability of the sample can not promote such a model is of no significance; through a sensitivity analysis on the BP neural network model can be displayed based on the input variables on the output variables The degree of influence. Since the institute selected medical record information is limited, and taking into account the promotion of network capacity and fitting ability, so the model prediction accuracy on hospital costs is limited, you can provide a theoretical reference. In the model sensitivity analysis on the basis of input variables to measure the degree of influence inside size is feasible. Due to the different parameter settings for the network have different effects on the results, such as different number of hidden layer neurons, different initial weights and thresholds, etc., is still lacking theoretical support, pending further study.
|