Climate factors influence seasonal influenza activity in Bangkok, Thailand

factor and the peak of an influenza virus activity (e.g., lag 1 represents 1 month difference). We found that there was no significant association between influenza A(H1N1)pdm09 activity with any meteorological variables (Fig 4A). Influenza A(H3N2) showed a moderately positive correlation with the mean temperature of lag 4 (r = 0.65, p<0.001), and with relative humidity and rainfall of lag 1 (r = 0.47 and r = 0.44, respectively) (p<0.001) (Fig 4B). Additionally, influ- enza B was positively correlated with temperature at lag 4 (r = 0.37, p<0.05) (Fig 4C). To deter- mine the correlation among influenza (sub)types, weakly negative correlation was found between influenza A(H1N1)pdm09 and influenza A(H3N2) at lag 4 (r = -0.26, p<0.05). In contrast, influenza A(H1N1)pdm09 showed strongly positive correlation with influenza B virus (r = 0.63, p<0.001). All influenza activity showed positive correlation with the mean tem- perature at lag 4 (r = 0.62, p<0.001), and relative humidity and rainfall at lag 1 (r = 0.32 and r = 0.3, respectively) (p<0.05) (Fig 4D). To identify the best-fitting model for the observed values, we performed the log10 transfor- mation on the dependent variables to obtain a stationary time series (S1 Fig). We then com- puted the suitable univariate and multivariate ARIMA models. For multivariate model, variables which provided significant correlation were selected as predictors (S2 Table). The appropriate ARIMA models for individual influenza (sub)type and all influenza viruses were

established (S2 and S3 Figs) by comparing the AICc values between the univariate and multi- variate time series. We found that multivariate ARIMA model using both influenza incidence and climate factors provided better goodness of fit than the univariate ARIMA model, which used only influenza incidence (S3 Table). It also performed better and had reduced errors (RMSE and MAPE) than the univariate model. For influenza A(H1N1)pdm09, the SARIMA (2,1,1)(0,1,1)12 of multivariate time series gave a better forecast (RMSE = 0.89, MAPE = 22.99%), while the SARIMA(1,0,2)(1,1,0)12 model combined with all climate factors for influ- enza A(H3N2) provided lowest forecasting error (RMSE = 0.77, MAPE = 22.47%). In addition, the multivariate ARIMA(3,0,3)12 fitted well for influenza B when combined with temperature and influenza A(H1N1)pdm09 (RMSE = 1.14, MAPE = 46.09%). The SARIMA(2,0,1)(1,0,0)12 with the mean temperature at lag 4, relative humidity at lag 1, and rainfalls at lag 1 was the selected multivariate model with the best forecast value, which took into account all influenza and climate factors (RMSE = 0.15, MAPE = 2.81%). Fitted graphs derived from the univariate and multivariate analysis of the time series (Fig 5) showed that the multivariate models were better able to capture the observed influenza activity during 2018 than the univariate models. Overall, the forecast model using all influenza activity and all climate factors performed better than the forecast model of individual (sub)type. Discussion

Influenza is prevalent throughout the year in some countries with tropical climate [6]. In Thai- land, increased influenza activity appears biannually typically in the rainy (June-August) and cool (December-February) seasons [20, 23, 28]. Our multi-year study confirmed the substan- tial decrease in influenza activity following the relatively hot and dry month of May, which precedes the rainy season and coincides with the yearly recommended influenza vaccination period in Thailand [40, 41]. We found that the first wave of influenza activity in a typical year involves A(H1N1)pdm09, and to a lesser extent influenza B virus, both of which peaked in February and again in August-September. The second and major wave of influenza activity occurs around August-September in which all influenza viruses co-circulate. The observation that influenza A(H1N1)pdm09 activity increases during the periods of low (February-May) and high (August-October) relative humidity has previously been reported and suggests that PLOS ONE Climate factors influence seasonal influenza activity in Bangkok, Thailand PLOS ONE | https://doi.org/10.1371/journal.pone.0239729 September 29, 2020 8 / 15