Clustering Of Hepatitis B Related Diseases And Prediction Of Medical Expenditures Based On Neural Network Model

ObjectivesHepatitis B virus(HBV)infection is one of the major public health problems in the world.Not only does it seriously endanger human health,but it also imposes a heavy economic burden on patients.Meanwhile,the rational allocation of health resources and the reduction of patients,families and sociol-economic burdens are one of the key issues for medical security work.Therefore,scientific and accurate prediction of medical expenses is the premise of the work.Based on the data of the hospital electronic medical records is easily accessible and can reflect the characteristics of a real word,it has become an ideal data source for the evaluation of health economics.However,the cost distribution shows the characteristics of skewed and heavy-tail.Moreover,the factors influencing of cost distribution are more complex.The distribution of these factors in patients also presents high dimensions,types and multiple collinearities.Therefore,the traditional statistical analysis method has been limited,and simply group by some observable variables to analyze the patient's medical expenses,it is difficult to fully reflect the characteristics of the patient and the heterogeneity of the group.If cost prediction can be made on the basis of effective identification of patient population heterogeneity,it is possible to improve the accuracy of prediction.This study was based on SOM neural network cluster to detect the heterogeneity of patients admission status and hospitalization treatment.And the prediction model of hospitalization total cost was established to analyze the influencing factors.At the same time,the performance of BP neural network model and linear regression model was compared.To evaluate the feasibility of predictingthe medical cost of hospitalized patients with CHB related diseases using SOM and BP neural network models.Provide scientific and reasonable data for diagnosis and treatment of CHB related diseases,and provide support for the application of neural network methods in the application of clinical cost data.MethodsThis study was based on the information system of an infectious disease hospital of Guangzhou and included 5194 patients with CHB related diseases in 2014 and2015(2014:2522;2016:2672).The demographic characteristics of the two-year patient,the results of the laboratory tests on admission,the treatment plans during the hospital stay and the medical expenses has been described.Unsupervised SOM neural network was used to cluster the two year hospitalized patients' hospitalization treatment plan and the laboratory examination results on admission to explore the heterogeneity of the hospitalized patients.Furthermore,combined with the supervised BP neural network method,a hospitalization cost prediction model was established and its influencing factors are analyzed.According to the different input variables,there are four BP neural network prediction model were established for the total cost of inpatients(NN-cost-1,NN-cost-2,NN-cost-3,NN-cos-t-4).Correspondingly four linear regression prediction model were established(LR-cost-1,LR-cost-2,LR-cost-3,LR-cost-4).At the same time,four BP neural network model for predicion of the total cost of the logarithm were established(NN-lgcost-1,NN-lgcost-2,NN-lgcost-3,NN-lgcost-4).Correspondingly four linear regression prediction model were established(LR-lgcost-1,LR-lgcost-2,LR-lgcost-3,LR-lgcost-4).Then,compare the prediction performance of each model.In this process,hospitalized patients with CHB and related diseases in 2014 were used as a training set,and 2015 inpatients were used as a validation set.ResultsAmong the 5194 hospitalized patients with CHB and related diseases were analyzed in this study,male dominated(78.17%),mean age was 46.0±14.2 years;patients with self-paying accounted for the highest(60.74%),and medical insurance patients accounted for 35.79%.Most patients with CHB(40.62%);the average hospital stay was 15.2±13.0 days;17.50% of patients had had hospital records in theprevious year.The total average cost for patients with CHB and related diseases in2014 and 2015 was 10837.8 and 1087.83 yuan respectively.Among them,the highest drug costs were 4510.9 and 4024.7 yuan,followed by laboratory examinations costs,3499.2 and 3626.9 yuan respectively.SOM clustering to explore the heterogeneity of patients treatment shows that the treatment methods of patients mainly include four categories.(1)Antiviral combined symptomatic combined complications treatment model,(2)Symptomatic combined complications treatment model,(3)Antiviral combined hepatoprotective treatment,(4)hepatoprotective treatment.The total hospitalization costs for these four treatment models in 2014 were 20618.7,13393.3,8743.4,and 6043.0 yuan respectively,and 21096.6,13801.3,9107.7 and 5626.0 yuan respectively in 2015.And the difference in various types of cost components are statistically significant.There are four disease status of patients admitted to hospital by SOM clustering,(1)Abnormal liver function with hepatic synthesis dysfunction with high HBV-DNA viral load,(2)Abnormal liver function with liver synthesis dysfunction,(3)Abnormal liver function combined with high HBV-DNA viral load,(4)Abnormal liver function.The hospitalization expenses for the four patients admitted to the hospital in 2014 were: 18012.9,14380.1,8194.6 and 7527.8 yuan respectively,and 16810.1,11934.7,8610.1 and 6093.8 yuan respectively in 2015.the difference in other cost components were also statistically significant.In terms of the neural network model predicts the total cost(NN-cost),The model prediction accuracy is about 83% and 77% respectively in the training set and the verification set.Compared with NN-cost-1,the variables formed by the SOM clustering to reflect the status of the patient at the time of admission and hospitalization treatment model were used as predictors(NN-cost-2,NN-cost-3,NN-cost-4),the prediction accuracy of the model is improved,and the MAPE is also reduced.The same result appears in the neural network model to predict the total cost of the logarithm(NN-lgcost).This study found that the NN-cost is better than the LR-cost model,the prediction accuracy is improved by about 3%,and the MAPE is reduced.This shows that the neural network model is superior to the linear regression model in predicting the total cost of hospitalization.Similarly,The NN-lgcost model is superior to the LR-lgcost model,suggesting that neural networks also show strong advantages in the prediction of total log cost.This study also found that the four models of NN-lgcost have improved the prediction accuracy of the model compared with the corresponding NN-cost model(about 3%).This suggests that the logarithmictransformation of the total cost of hospitalization improves the accuracy of model prediction.Similarly,the LR-lgcost model is also better than the LR-cost model,indicating that the cost after logarithm conversion is used as the forecast target,which improves the prediction accuracy of the linear regression model.In addition,The importance of predictor variables in the neural network model was analyzed.After the SOM clustering,a variable reflecting the heterogeneity of the population was used as a predictor,which was more important in cost forecasting.The neural network model is used to predict medical costs,which effectively avoids the problems such as the correlation between independent variables in the linear regression process,improves the effect of the independent variables in the linear regression process,and improves the model prediction accuracy.ConclusionsThe SOM neural network cluster analysis was used to detect the patients heterogeneity in treatment methods and hospital admission status.The differences in the total cost of hospitalization and cost components between the heterogeneous patients were statistically significant.It shows that SOM clustering can better distinguish the patient's group heterogeneity,and provide the premise for the hospital costs prediction.The model of BP neural network to predict the total cost of hospitalization is better than the linear regression model.Among them,based on SOM neural network cluster analysis,variables used to reflect the heterogeneity of the patients were generated as predictors,which improved the predictive performance of the model.Compared with the prediction of the total cost,in both the neural network model and the linear regression model,the performance of the model is significantly improved in predicting the total cost of the logarithmic transformation.