Research On Forecasting Model Of Hospitalization Expenses Based On Support Vector Machine

Background With the development of our country’s economy, the acceleration of the urbanization process and the continuous progress of medical technology, the medical expenses of our residents are rapidly increasing year by year. The problem of "difficult and expensive" has become the focus and difficulty of the health policy reform in China. At the same time, gastric cancer is one of the most common gastrointestinal tract tumors in China. The etiology of gastric cancer is not clear and easy to relapse. As one of the chronic and long-lasting diseases, gastric cancer is costly, further aggravate the economic burden on the patients. At present, the research methods to hospitalization expenses are most based on traditional statistics, the data mining methods include neural network and decision tree. Only a few studies available on the hospitalization expenses modeling by support vector machine.Objectives This paper aims to model the hospitalization expenses which was analysed the hospitalization expenses of patients with gastric cancer as an example, discussion on reasonable classification rule of hospitalization expenses, setting proper kernel functions and parameters to model hospitalization expenses based on support vector machine, measuring the impact degree of various factors on the hospitalization expenses and improving classification accuracy of SVM. It can help health administrators, policy-makers and medical insurance workers make the right decision. At the same time, this paper is explored to a new method to model the hospitalization expenses, in order to give some references for the methodology to model the hospitalization expenses.Methods This study takes hospitalization expenses data in the Affiliated Hospital of Ningxia Medical University for five years which are from year 2009 to 2011 and from 2013 to 2014 as samples, two modeling methods of hospitalization expenses ware proposed. One model of hospitalization expenses built based on K-means clustering and SVM, the other one is a new method to optimize the feature subset and SVM parameters based on genetic algorithm. Classification accuracy, sensitivity and specificity are using for evaluating the effect of classification results.Results For the first method based on clustering and SVM, the experimental result showed that the clustering accuracy of K-Means by year was increased by 13.13% compared to only by distribution characteristics. The gauss kernel function based SVM was the most stable model, with a classification accuracy rate of 92.11% when the penalty factor C and parameter γ were set to be 10 and 1, respectively. And the classification result of this algorithm is better than 2-logistic and BP neural network.For the second method to optimize the feature subset and SVM parameters based on genetic algorithm, the experimental result showed that GA can find the optimal subset of features and SVM parameters as soon as quickly. And the results of comparison experiments showed that it was highest than other algorithms from classification accuracy, sensitivity and specificity. At the same time, the time complexity of proposed method was least. Therefore, it showed that the algorithm proposed is an efficient method.