Prediction for trend and seasonal variation of incidence of hemorrhagic fever with renal syndrome: time series decomposition analysis

  Objective  To establish a predictive model for incidence trend and seasonality of hemorrhagic fever with renal syndrome (HFRS) in China and to evaluate the efficacy of the model based on time series decomposition analysis on the data between 2011 – 2016.

  Methods  Data on reported monthly HFRS incidences during 2011 – 2016 across China were collected via the website of National Health and Family Planning Commission of People’s Republic of China. We analyzed incidence trend and seasonality of HRFS during the period using time series decomposition. We performed modeling using the data with seasonal variable (S₂) having been removed; then the constructed model was multiplied by S₂ to establish a final model. The prediction accuracy was evaluated using back-substitution method.

  Results  The incidence of HFRS rose first and then declined, with an obvious seasonality, during the 6-year period in China. The HFRS incidence showed a bimodal distribution during a 12-month period, with the first peak from May to June of a year and the second from November of a year to January of the next year. The parameters of the established autoregressive integrated moving average (ARIMA) (2, 1, 1) were as following: Akaike information criterion = 866.4, autoregression at-lag-1 (AR1) = – 0.3543, AR2 = – 0.3543, moving average = – 0.7441 (P < 0.01); the residual of the established model was a white noise sequence (Q(20) = 16.364, P = 0.694). The formula for prediction of monthly number of HFRS was S₂ × ARIMA (2, 1, 1). The results of back-substitution revealed a mean absolute error (MAE) of 71.31 and an average absolute percentage error (MAPE) of 7.00% for predicted monthly incidence number of HFRS during the 6-year period across China.

  Conclusion  Time series decomposition can be used to predict the trend and seasonal variation HFRS incidence and applicable model for the prediction of monthly HFRS incidence could be established using data with seasonal variable (S₂) having been removed.