The Research On Predictive Model Of Influenza Incidence Based On Multi-source Data

[Purpose] In view of the critical nature and social harmfulness of influenza,it is urgent to prevent,monitor and warn influenza outbreaks.However,the factors affecting the outbreak and epidemic of influenza are complicated and interrelated.At present,the research on the incidence prediction of influenza based on the data of multi-source influenza is still insufficient.Starting from the multi-source data affecting the influenza epidemic,this study established a non-linear fitting incidence prediction model to predict and warn influenza more comprehensively and accurately,and timely carry out disease prevention and intervention control,so as to reduce the social harm and economic loss brought by influenza as far as possible.[Methods] Using EXCEL 2016 to weekly collect electronic health records data sets,drug sales data set and meteorological data set all three sources of data items recorded by centers for disease control and prevention in Yichang City of Hubei Province in 2017-2019,and compare with the percentage of influenza-like cases ILI % do correlation analysis,find out the main factors of increase in the number of influenzalike cases;Then in the application of the python language environment to take both short-term and long-term memory(LSTM)model based on attention mechanism flu nonlinear prediction model is set up,and at the same time establish LSTM model and support vector regression model,finally using historical data to validate the accuracy of the model,to comparative analysis of the model according to the results of the validation.[Results] After delay correlation analysis of screening with a total of 12 factors into the model,respectively is 8 days in advance of flu etiology positive cases,6 days in advance of illness associated with influenza cases,7 days in advance of antipyretic analgesic drugs daily sales,6 days in advance of antimicrobial agents,sales,five days ahead of influenza virus drugs daily sales,12 days in advance of antihistamine drugs sales,cough medicines for 7 days in advance sales,23 days in advance of wind speed,precipitation,22 days ahead of 15 days in advance,the highest temperature,daily minimum temperature of 15 days in advance,the minimum humidity 23 days of days in advance.The mean square error(MSE)of the three influenza prediction models on the training set showed that the Att-LSTM model with a predicted time domain of 8 days,the LSTM model with a hidden layer node of 128 days and the SVR model with a kernel function of RBF were used as the optimal fitting model.On the test set,The Att-LSTM model’s three evaluation indexes MAE,RMSE and MAPE are the smallest,and the predicted performance reaches the expected level.[Conclusions] Weather,drug sales and electronic medical records in combination with the three influenza related data source,on the basis of nonlinear time sequence forecast early warning model is established,in full consideration of historical information embedded at the same time the attention mechanism to measure the weight of the correlation factor,achieved good prediction effect,for short-term forecasting warning of a pandemic and advance public health intervention,develop prevention plan provides a certain support.