Time Series Study Of Monthly Case Numbers Of Hemorrhagic Fever With Renal Syndrome In Liaoning Province Based On Long Short-term Memory Neural Network Model

Objective:To illustrate the characteristics of epidemic of hemorrhagic fever with renal syndrome disease and its pattern of change over time in Liaoning province in the course of from January 2004 to December 2017;To select all the meteorological variables correlated to the incidence of HFRS and to analyze the hysteresis effect between them;To compare the performances of long short-term memory neural network model and auto-regressive integrated moving average model in terms of accuracy of fitting and predicting of HFRS.Methods:Collected the HFRS data and meteorological data of Liaoning province over the course of from January 2004 to December 2017,the meteorological data included but not limited to monthly average temperature,monthly average pressure,monthly average humidity,monthly average maximum temperature,monthly average water press;And analyzed the characteristics and pattern of change of HFRS over time,and statistically described the characteristics of meteorological variables;Selected all the meteorological variables correlated to the incidence of HFRS and analyzed the hysteresis effect between them by using pre-whitening cross-correlation method;Compared the performance of long short-term memory neural network model and auto-regressive integrated moving average model in terms of accuracy of fitting and predicting of HFRS;In the study,we used EXCEL and R to help us analyse.Results:1.There are totally 22489 cases of HFRS reported between 2004 and 2017 in Liaoning province,and there has been consecutive report of cases every month in every year.In a general perspective,the incidence of HFRS shows a stable downtrend,in the period of from 2004 to 2007,the annual reported case slumped and after 2007,the amount of annually reported HFRS cases generally remained stable.In average,in September it saw the lowest amount of monthly reported cases while in March and November it saw the highest amount of monthly reported cases.2.Using the methods of pre-whitening cross-correlation function,we rejected meteorological variables with too much missing value and variables with no significant correlation to the HFRS,and selected three meteorological variables:monthly average water press with 1 month lag,monthly average relative humidity with 1 month lag and monthly average maximal temperature.These three variables showed a statistically significant correlation with the incidence of HRFS.3.Using the exhaustive method,we got the optimal seasonal multiplicative ARIMA model(ARIMA（2,1,0）×（0,1,1）₁₂)with the smallest AIC（-29.31）,and the residual is proven as white noise series.And we successfully predicted monthly reported case of HFRS with the selected ARIMA model,and calculated the fitting accuracy(MAE17.909,MAPE 18.168%,RMSE 26.601 and predicting accuracy（MAE 40.787,MAPE52.958%,RMSE 47.928）.4.LSTM neural network model outperformed ARIMA model in terms of accuracy of both fitting（MAE 14.502,MAPE 16.965%,RMSE 21.242）and predicting（MAE15.929,MAPE 23.659%,RMSE 19.035）of HFRS.Conclusion:1.The epidemic of HFRS in Liaoning province presented a general stable downtrend with a significant seasonality from 2004 to 2017,the tally of yearly reported cases of HFRS maintained stable after 2007,and the total amount of monthly reported cases normally peaked at winter and spring.2.Three meteorological variables,monthly average water press with 1 month lag（negative correlation）,monthly average relative humidity with 1 month lag（negative correlation）and monthly average maximal temperature（positive correlation）,showed a statistically significant correlation with the incidence of HRFS.3.LSTM neural network model outperformed ARIMA model in terms of accuracy of both fitting and predicting of HFRS.