Establishment And Optimization Of Prediction Model On COVID-19 Epidemic

Background and PurposeEmerging infectious disease is a global public health problem that cannot be ignored because of its rapid onset,high infectivity and high mortality.The efficiency of the prediction and early warning of infectious diseases in the past is not high,which is mainly limited by the lack of historical data and the lag of reporting time.With the advent of the era of big data,machine learning methods are widely applied in epidemic prediction on.Coronavirus Disease 2019 was firstly reported in Hubei Province of China at the end of December 2020,coinciding with the Spring Festival.The large scale of population migration provided conditions for the early transmission of the epidemic.In order to contain the epidemic,the China government has taken unprecedented measures,such as cutting off all traffic out of Wuhan.Aiming to grasp the epidemic trend in a timely manner,a number of prediction models have been published,but without full consideration of the impact of prevention and control measures on the epidemic.Therefore,this study intends to establish SEIR model and LSTM model to predict the development of early COVID-19 in China based on the situation of prevention and control measures,and further optimize the prediction model with TFT framework,so as to improve the accuracy and applicability of the model.Data and research methods（1）Assessment on epidemiological characteristicsIn this study,number of COVID-19 cases in various provinces of China from January 17 to February 16,2020 were collected.Tableau software was used to conduct spatio-temporal analysis of the cases.The crude case fatality rate and basic regenerative number R0 were calculated by formula method and exponential growth method respectively.（2）Construction of SEIR and LSTM modelThis study mainly collected the data of Baidu migration and the SARS epidemic in 2003,and used the Epimodel package in R Version 3.4.1 software to construct the differential equation.We output the forecast data based on the SEIR model,and use the Py Torch learning package in Python to build the LSTM model.（3）Model optimization based on TFT architectureIn this part,the coefficients of multinational policy response were collected from the COVID-19 Government Response Tracking System,and the TFT framework was used to learn the national control coefficients to fit the infection indexβin the SEIR model,so as to conduct epidemic prediction with SEIR model based on TFT deep learning method in many countries.Results（1）Epidemiological characteristics of COVID-19In early February,Hubei province reported more than 10,000 cases of COVID-19,while Zhejiang and Guangdong provinces have become the two highest incidence regions outside Hubei province with more than 1,000 cases.The crude case fatality rate of COVID-19 was about 2.51%,The basic productive number of COVID-19 in Hubei,Guangdong and Zhejiang provinces was R₀=3.10（95%CI:3.07-3.13）,3.20（95%CI:2.81-3.63）and 2.66（95%CI:2.35-3.01）.（2）Comparison on prediction of SEIR and LSTMAccording to the SEIR model,the epidemic in Hubei Province leveled off in late April,with the total number of cases estimated to be 59,578（95%CI:39,189-66,591）.Assuming that the quarantine measures were relaxed in Hubei Province（with a non-zero migration index）,the final total outbreak size would reach 73,180（95%CI:51,308-85,839）.In this study,Guangdong and Zhejiang provinces are expected to peak on 20 February,with the final estimated outbreak size reaching 1,511（95%CI:1,097-1,948）and 1,491（95%CI:1,066-1,851）confirmed cases,respectively.The national epidemic leveled off after the end of April,with the estimated final size of the epidemic totaling 122,122（95%CI:89,741-156,794）.Our model suggests that a five-day delay in the implementation of control measures will triple the size of the epidemic.According to the prediction results of the LSTM model,the epidemic gradually leveled off after April,and the turning point of the epidemic appeared.By the end of the month,it was estimated that there would be 95,000 confirmed cases.From the perspective of prediction accuracy,MAPE of SEIR model is 11.34,RSME is 1560.45,LSTM model is slightly lower than SEIR model,MAPE is 9.40,RSME is 651.40.（3）Model optimization based on TFT architectureFrom the perspective of model fitting degree,the mixed model has a good fitting degree for the epidemic data of 7 countries including the United States,Italy,India,Spain,Russia,Turkey and the United Kingdom,but has a poor fitting degree for the three countries of Brazil,Germany and France,with a large number of outliers.From the perspective of model accuracy,compared with the SEIR model of population flow,RMSE and MAPE of the mixed model decreased by 1.33*103 and 11.46,respectively.Conclusion（1）Among the three highly pathogenic coronaviruses,SARS-Co V,SARS-Co V-2 and MERS-Co V,SARS-Co V-2 had the highest infectivity and the lowest lethality;（2）Timely control measures taken by the Chinese government have effectively reduced the scale of the outbreak.For the early stage of the epidemic,the accuracy of LSTM method is higher than that of SEIR model.（3）From the perspective of accuracy,the SEIR model based on TFT architecture is superior to the SEIR model and LSTM model of population flow.In addition,the hybrid model is conducive to improving the compatibility of the model and expanding the applicability of the model.