Comparison of autoregressive integrated moving average model and generalised regression neural network model for prediction of haemorrhagic fever with renal syndrome in China: a time-series study

Abstract

Objectives: Haemorrhagic fever with renal syndrome (HFRS) is a serious threat to public health in China, accounting for almost 90% cases reported globally. Infectious disease prediction may help in disease prevention despite some uncontrollable influence factors. This study conducted a comparison between a hybrid model and two single models in forecasting the monthly incidence of HFRS in China.

Design: Time-series study.

Setting: The People's Republic of China.

Methods: Autoregressive integrated moving average (ARIMA) model, generalised regression neural network (GRNN) model and hybrid ARIMA-GRNN model were constructed by R V.3.4.3 software. The monthly reported incidence of HFRS from January 2011 to May 2018 were adopted to evaluate models' performance. Root mean square error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE) were adopted to evaluate these models' effectiveness. Spatial stratified heterogeneity of the time series was tested by month and another GRNN model was built with a new series.

Results: The monthly incidence of HFRS in the past several years showed a slight downtrend and obvious seasonal variation. A total of four plausible ARIMA models were built and ARIMA(2,1,1) (2,1,1)₁₂ model was selected as the optimal model in HFRS fitting. The smooth factors of the basic GRNN model and the hybrid model were 0.027 and 0.043, respectively. The single ARIMA model was the best in fitting part (MAPE=9.1154, MAE=89.0302, RMSE=138.8356) while the hybrid model was the best in prediction (MAPE=17.8335, MAE=152.3013, RMSE=196.4682). GRNN model was revised by building model with new series and the forecasting performance of revised model (MAPE=17.6095, MAE=163.8000, RMSE=169.4751) was better than original GRNN model (MAPE=19.2029, MAE=177.0356, RMSE=202.1684).

Conclusions: The hybrid ARIMA-GRNN model was better than single ARIMA and basic GRNN model in forecasting monthly incidence of HFRS in China. It could be considered as a decision-making tool in HFRS prevention and control.

Keywords: autoregressive integrated moving average; generalized regression neural network; hemorrhagic fever with renal syndrome; prediction.

Figure 1

Monthly incidence of HFRS in China from January 2011 to December 2017. HFRS, haemorrhagic fever with renal syndrome.

Figure 2

The ACF and PACF graphs of differenced HFRS incidence series. ACF, autocorrelation function; HFRS, haemorrhagic fever with renal syndrome; PACF, partial autocorrelation function.

Figure 3

Residual white noise test. ACF, autocorrelation function.

Figure 4

The selection of basic GRNN model and hybrid ARIMA-GRNN model. ARIMA, autoregressive integrated moving average; GRNN, generalised regression neural network; RMSE, root mean square error.

Figure 5

The fitting and forecasting curves of three models and the actual HFRS incidence series. ARIMA, autoregressive integrated moving average; GRNN, generalised regression neural network; HFRS, haemorrhagic fever with renal syndrome.