| Background:Road traffic injuries(RTIs)have become a serious social and public health problem.According to the World Health Organization(WHO),RTIs have become the 8th leading cause of death for people of all ages and the 1st leading cause of death for children and young adults aged 5-29 years worldwide.The number of RTIs remains unacceptably high,with 1.35 million people dying and 20 million to 50 million people suffering non-fatal injuries globally every year.Most of these road traffic deaths occur in developing countries.As the largest developing country in the world,China is facing enormous challenges in improving road traffic safety and reducing road traffic casualties with the improvement of highway construction,the increasing number of private cars and the improvement of degree of motorization.RTIs,which mainly occur among productive populations aged 16 to 45 in China,have a serious impact on the potential economic and human capital.Compared with developed countries,there are some problems in China,such as lack of road traffic safety facilities,road users violating traffic rules and poor management.In order to achieve the global and national road safety goals,China should pay more attention to RTIs and take certain measures and strategies to promote the development of road traffic safety.Objectives:1.To estimate epidemiological characteristics of RTIs in China;2.To forecast short-term RTIs in China using different prediction methods;3.To explore practice and application of different artificial neural network models in road traffic injury prediction;4.To compare different prediction models comprehensively and provide scientific basis for the selection of road traffic safety prediction models.Methods:1.According to the data from Bureau of Traffic Management of the Ministry of Public Security of People’s Republic of China,epidemiological characteristics of RTIs in China in 2019 was described and analyzed.The prevalence trends and internal rules of RTIs in China is clarified,including the characteristics of crash time,region,cause and type of road,as well as the victim(s)’characteristics.2.Taking the data of monthly road traffic crashes,fatalities and nonfatal injuries in China from 2000 to 2018 as sample data,the prediction models were established by using regression analysis prediction method,time series prediction method,Back Propagation Neural Network(BPNN)and Elman recurrent neural network(ERNN),respectively,to forecast the monthly RTIs in 2019.3.Different prediction models were compared and analyzed to explore the application scope.Results:1.RTIs mostly occurred in the daytime,and the number of road traffic crashes varies in different seasons and months.More RTIs occurred on weekdays were than on weekends.The province with the highest number of crashes was Hubei,followed by Guangdong and Guangxi.The province with the highest number of fatalities was Guangdong,followed by Hubei and Guangxi.The province with the highest number of nonfatal injuries was Hubei,followed by Guangxi and Guangdong.Shanghai had the highest crash fatality ratio(the average number of deaths per crash)at 0.99,followed by Hebei,Liaoning,Gansu and Yunnan.In about 83.24%of the crashes,the vehicle drivers were believed to be responsible for the crashes.Among these crashes,drunk driving accounted for 9.41%.the accidents were mainly between vehicles,accounting for 70.16%.Deaths between 21 and 55 years old accounted for 48%of the total number of fatalities.The fatalities of crashes were mainly male,accounting for 70.13%.The fatalities of the crashes were mainly agricultural household registration.Road traffic crashes mainly occurred in plain terrain,accounting for 69.05%of the total crashes.Crashes mainly occurred on straight roads(81.49%).Most of the crashes occurred in the motorway(62.01%).Road traffic crashes mostly occur on secondary main roads,and crash fatality ratio of highways is the highest.2.The regression model for the crashes was=0.009~3-1.438~2-318.855+64422.785.The regression model for the fatalities was=0.002~3-0.714~2+33.064+8235.751.The regression model for the nonfatal injuries was=0.016~3-4.854~2+236.579+39038.028.The coefficient of determination was0.900,0.774 and 0.884,respectively.The prediction fitting mean absolute percentage error(MAPE)of crashes,fatalities and nonfatal injuries by regression models was38.71%,28.94%and 36.25%,respectively.3.The Seasonal Autoregressive Integrated Moving Average(SARIMA)model was established with the same data.The best models of crashes(SARIMA(0,1,2)(0,1,1)12),fatalities(SARIMA(2,1,2)(2,1,2)12)and nonfatal injuries(SARIMA(0,1,1)(0,1,1)12)were selected from several candidate models.The prediction MAPE of models of crashes,fatalities and nonfatal injuries were 5.22%,4.88%and 5.29%,respectively.The predictive values of RTIs in 2019 in China consistent with the observed values,and the actual values were within the 95%CI of predictive value.The prediction MAPE of crash model,fatality model and nonfatal injury model were 19.96%,5.05%and15.08%,respectively.4.The BPNN model provided the best result when the input neurons were set to 3.The optimal structure of BPNN model for crash,fatality and nonfatal injury was 3-15-1,3-12-1,3-6-1,respectively.After comparing different functions,the(69)4)2)function was selected as the activation function of the hidden layer,and the0)7)4)9)0)function was selected as the output layer.The(64)9)(82)function was selected by comparing several common training functions.The prediction MAPE of crash model,fatality model and nonfatal injury model were 9.03%,4.84%and 6.69%respectively;the prediction mean absolute error(MAE)were 1840.67,254.42,1422.08 respectively;the prediction root mean square error(RMSE)were 2415.79,304.73 and 1768.03respectively.5.ERNN models provided the best result when the input neurons were set to 3.The best model structure of ERNN model for crash,fatality and nonfatal injury was 3-8-1,3-5-1,3-12-1.Similar to BPNN model,the(69)4)2)function was selected as the hidden layer activation function after comparison.The best performance network training function(64)9)2)(9 was selected.The prediction MAPE of the final crash model,fatality model and nonfatal injury model were 7.47%,3.89%and 6.61%,respectively;the prediction MAE were 1566.17,205.08,1383.08,respectively;the prediction RMSE were 2223.22,260.87 and 1976.29,respectively.6.The comparison of artificial neural networks with different structures showed that both BPNN model and ERNN model have the best prediction accuracy in NET 4(predicted RTIs in the current month using the deaths in the same period of the previous3 years),especially for the forecasting model of road traffic fatalities.The prediction MAPE of BPNN and ERNN were lower than 5%.7.The comprehensive comparison of the prediction results of four different models showed that the ERNN model had the best prediction accuracy,followed by the BPNN model and the SARIMA model.The regression model predicted the worst effect.Conclusions:1.In 2019,RTIs in China remain a serious public health problem.2.The prediction accuracy of the four prediction models for crash,fatality and nonfatal injury based on monthly RTIs data from 2000 to 2019 in China was ERNN model>BPNN model>SARIMA model>regression model.3.When it comes to the selection of different prediction methods,the sample size,data regularity,seasonality and so on should be taken into consideration.Regression analysis model can be selected when the sample data size is large and has strong regularity.The data should be enough sample size and have obvious seasonality for SARIMA model.When the data has obvious seasonality and the sample size is small,the artificial neural network model can be selected.The prediction effect of ERNN model is slightly better than BPNN model. |
