The Prediction Of Incidence Of Hemorrhagic Fever With Renal Syndrome Based On Long Short-term Memory Neural Network Model

Objective:To understand the spatial and temporal distribution of hemorrhagic fever with renal syndrome(HFRS)in different years in some areas of Liaoning Province and the host surveillance of HFRS outbreaks in various regions,and to analyze the influence of host animal factors and meteorological factors on HFRS.According to the correlation between host animal factors,meteorological factors,and the incidence of HFRS,the Long short-term memory(LSTM)neural network model was established for univariate and multivariate time series prediction,and compared with the traditional time series model.To explore the applicability of LSTM neural network model in predicting the incidence of HFRS,so as to provide reference for the prevention and control of the infectious disease.Methods:The monthly incidence data of HFRS from 2009 to 2016 were collected from the National Health Commission(http://www.nhc.gov.cn),and relevant literature was found from CNKI,Wanfang,Yuexiu and other websites,and the monthly incidence data and some host surveillance data(rat density,rat poison rate,rat poison index)of HFRS from January 2005 to December 2017 in Shenyang City and January 2005 to December 2020 in Dandong City were collected.Meteorological data were collected from the National Meteorological Science Data Center(http://data.cma.cn)and Huiju Data Network(http://hz.hjhj-e.com),and the monthly meteorological factors included:precipitation from A20 to 20(i.e.cumulative precipitation from 20 pm to 20 pm),average air pressure,average 2-minute wind speed,average temperature,average water pressure,average relative humidity,average minimum temperature,average maximum temperature and sunshine duration.The time distribution characteristics of HFRS incidence in Shenyang City and Dandong City were described,and the Spearman rank correlation between host surveillance data and HFRS incidence was analyzed,and the Spearman rank correlation between meteorological factors and HFRS incidence was analyzed with a lag of 0-6 months,and the test level was α=0.05.According to the correlation between host animal factors,meteorological factors and the incidence of HFRS,univariate and multivariate LSTM neural network prediction models were constructed to fit and predict the previous incidence.The prediction results were compared with the Autoregressive Integrated Moving Average model(ARIMA).The prediction results were compared with the Autoregressive Integrated Moving Average model(ARIMA).The RMSE、MSE、MAE were used to compare and evaluate the predictive effect of the two types of models.Results:1.There are obvious differences in the incidence level of the two cities,from 2009 to 2016,the incidence of HFRS in Dandong City was significantly higher than the national level,and the incidence of HFRS in Shenyang City was comparable to the national level before 2012,and gradually lower than the national incidence level after 2012;The incidence of HFRS has obvious seasonal characteristics,with Shenyang City showing high incidence in spring and autumn,and Dandong City showing high incidence in autumn and winter.2.In the correlation study between host animal factors and HFRS incidence,the correlation coefficients between the incidence rate and rat density in Shenyang City,the rat poison rate in residential areas and the rat poison index were statistically significant,with correlation coefficients of 0.946,0.852,0.864,0.877,0.542,0.505,0.846 and 0.740,respectively,and the correlation between the incidence rate and the density of rats in the wild in autumn was statistically significant,with a correlation coefficient of 0.881.3.In the lagging correlation analysis between meteorological factors and HFRS incidence of 0-6 months,the correlation between the monthly incidence of HFRS in Shenyang City and the precipitation from A20 to 20 with a lag of six months was statistically significant and the strongest correlation,with a correlation coefficient of 0.219,which was statistically significant and the strongest correlation with the mean air pressure,average temperature,average minimum temperature and average maximum temperature with a lag of two months,and the correlation coefficients were 0.332,-0.343,-0.317,-0.352,which was statistically significant and the strongest correlation with the average 2-minute wind speed of the month,the correlation coefficient was 0.516,and the correlation coefficient was statistically significant and the strongest correlation with the average water pressure with a lag of one month,the correlation coefficient was-0.315.Which was significantly correlated with the average relative humidity and sunshine duration with a lag of four months,and the correlation coefficients were 0.334 and-0.393,respectively.The correlation between the monthly incidence of HFRS in Dandong City and precipitation from A20 to 20,average 2-minute wind speed,average temperature,average water pressure,average relative humidity,average minimum temperature and average maximum temperature was statistically significant and the correlation was the strongest in Dandong City,with correlation coefficients of 0.472,-0.418,0.532,0.547,0.518,0.555 and 0.499,respectively.Which was significantly correlated and most strongly correlated with average air pressure with a lag of four months,with a correlation coefficient of-0.532.Which was significantly correlated and most strongly correlated with sunshine duration with a lag of six months,with a correlation coefficient of 0.232.4.Compared with the ARIMA model,the LSTM neural network prediction model had a better fitting and prediction effect on the incidence of HFRS in Shenyang City,and in the prediction of HFRS incidence in Shenyang City,univariate and multivariate time series RMSE,MSE and MAE were 0.036,0.001,0.030,0.032,0.001 and 0.026,respectively,while in the prediction of HFRS incidence in Dandong City,the ARIMA model had better fitting and prediction effects for univariate time series.The RMSE,MSE and MAE were 0.231,0.053 and 0.159,respectively,and the LSTM neural network prediction model had better fitting and prediction effect for multivariate time series,RMSE,MSE and MAE were 0.176,0.031 and 0.095,respectively.In addition,the prediction accuracy of LSTM neural network models for multivariate time series was better than univariate time series.Conclusion:1.Host animal factors are closely correlated with the incidence of HFRS,and the impact of mouse surveillance information on incidence is different in different habitat types and seasons.2.The correlations between the incidence of HFRS and precipitation from A20 to 20,average air pressure,average 2-minute wind speed,average temperature,average water pressure,average relative humidity,average minimum temperature,average maximum temperature and sunshine duration were statistically significant.3.The LSTM neural network model has good predictive ability for HFRS incidence,and the prediction of multivariate time series is better than univariate time series,and the time series prediction of HFRS incidence based on LSTM neural network model has certain application value.