A combination of annual and nonannual forces drive respiratory disease in the tropics

Introduction It is well known that influenza and other respiratory viruses are wintertime-seasonal in temperate regions. However, respiratory disease seasonality in the tropics is less well understood. In this study, we aimed to characterise the seasonality of influenza-like illness (ILI) and influenza virus in Ho Chi Minh City, Vietnam. Methods We monitored the daily number of ILI patients in 89 outpatient clinics from January 2010 to December 2019. We collected nasal swabs and tested for influenza from a subset of clinics from May 2012 to December 2019. We used spectral analysis to describe the periodic signals in the system. We evaluated the contribution of these periodic signals to predicting ILI and influenza patterns through lognormal and gamma hurdle models. Results During 10 years of community surveillance, 66 799 ILI reports were collected covering 2.9 million patient visits; 2604 nasal swabs were collected, 559 of which were PCR-positive for influenza virus. Both annual and nonannual cycles were detected in the ILI time series, with the annual cycle showing 8.9% lower ILI activity (95% CI 8.8% to 9.0%) from February 24 to May 15. Nonannual cycles had substantial explanatory power for ILI trends (ΔAIC=183) compared with all annual covariates (ΔAIC=263) in lognormal regression. Near-annual signals were observed for PCR-confirmed influenza but were not consistent over time or across influenza (sub)types. The explanatory power of climate factors for ILI and influenza virus trends was weak. Conclusion Our study reveals a unique pattern of respiratory disease dynamics in a tropical setting influenced by both annual and nonannual drivers, with influenza dynamics showing near-annual periodicities. Timing of vaccination campaigns and hospital capacity planning may require a complex forecasting approach.


S1 Text: Time series detrending
There are 7 clinics showing a long-term downtrend in %ILI.To remove the long-term trend and get a stationary time series, we detrended the %ILI by dividing each daily value by a 365-day moving average centered at that day for each clinic.We refer to this transformation as a ζ ("zeta")-score as it is related to an exponentiated z-score.This removes trends longer than 365 days and preserves the ratios between sequential daily reporting numbers.For influenza, we calculated daily ILI + as the product of the mean of the daily ILI ζ-score across all the clinics and the influenza positivity rate on that day to get a time series representing influenza incidence.
Then we applied a 7-day moving average to smooth both the ILI ζ-score and ILI+ to filter out short-term noise.The 7-day moving smoothed ILI ζ-score and ILI+ are used for subsequent analysis.

S2 Text: ILI data from other surveillance systems
To compare ILI trends in HCMC with other regions, ILI data from the United States, four European countries (Belgium, France, Greece, Netherlands), Singapore, and Hong Kong were (https://www.chp.gov.hk/en/resources/29/304.html), from December 29, 2013 to December 28, 2019.All ILI data were converted into an ILI ζ-score time series with the same methods above.

S3 Text: Cyclic step function
We used simple step functions with k steps in a periodic cycle of length c.The function values and breakpoints were estimated by minimizing the Akaike Information Criterion (AIC) using normally distributed errors.We maximized a normal likelihood using the Nelder-Mead algorithm for step functions with cycle lengths, denoted c, ranging between 150 and 450 days (in increments of 5 days), and number of steps, denoted k, ranging from 2 through 8.Only the period that is longer than 15 days would be considered as one step in step function.The estimated parameters were selected from 100 optimized step functions with randomly selected initial parameters to ensure a global optimum was found.95% confidence intervals were obtained via likelihood profiling.

S4 Text: Regression Covariates
We collected climate data from the NASA POWER project.Based on the criterion to only include climate factors that were reported to be biologically or epidemiologically associated with ILI transmission, we collected daily temperature, absolute humidity, and precipitation.We also included the 1-week, 2-week, 3-week lagged version of all climate variables because the effect of climate on ILI and influenza can be delayed.Absolute humidity was calculated using temperature and relative humidity: Eq.4 Each climate predictor was normalized using z-score normalization with the mean and the standard deviation of the predictor time series and then smoothed using 7-day moving average.
The school-term categorical variable is one from August 15 to June 1, when schools in HCMC are in session and zero otherwise.

S5 Text: Predictor importance
In multiple linear regression, the importance of each predictor in regression was measured as the averaged difference in R 2 when adding the predictor in all possible subsets of predictors in the model 33 .It is also referred as LMG in dominance analysis 34 , defined as: We compare the number of influenza cases expected (blue) and the number of influenza cases reported (red) from 24 out of the selected 33 clinics.We used the influenza positive % in 21-day windows (among swabbed ILI patients, across all clinics) as a baseline, and calculated the number of expected influenza cases (blue) for each clinic using this baseline percentage.With the possible exception of clinic 19, the clinics did not show consistently higher or lower trends than their expected values, suggesting that there is no bias in ILI diagnosis.
BMJ Publishing Group Limited (BMJ) disclaims all liability and responsibility arising from any reliance Supplemental material placed on this supplemental material which has been supplied by the author(s)

Figure S2 .
Figure S2.Validation of variation in ILI diagnosis using laboratoryconfirmed influenza cases.We compare the number of influenza cases expected (blue) and the number of influenza cases reported (red) from 24 out of the selected 33 clinics.We used the influenza positive % in 21-day windows (among swabbed ILI patients, across all clinics) as a baseline, and calculated the number of expected influenza cases (blue) for each clinic using this baseline percentage.With the possible exception of clinic 19, the clinics did not show consistently higher or lower trends than their expected values, suggesting that there is no bias in ILI diagnosis.

Figure S3 .Figure
Figure S3.Circular plots of weekly ILI ζ -score from HCMC, Singapore, Hong Kong, the United States, France, and Belgium (data sources: S2 Text).The orange color indicates tropical or subtropical regions, the blue color indicates temperate regions.The data from each year was visualized in a circular time scale.The range of each plot is [0,6].The weekly ILI ζ -score from HCMC is calculated by averaging the daily ILI ζ -score within one week.

Figure S5 .
Figure S5.Pearson's correlation of ILI ζ -score between different years from 2010 to 2019.The highest observed correlation is seen between 2018 and 2019.

BMJFigure S6 .
Figure S6.Wavelet transform of the ILI ζ -score in ten HHS regions from the United States, four European countries (France, Belgium, Greece, and Netherlands), subtropical city Hong Kong, and tropical country Singapore, along with ILI+ from HCMC.Continuous annual seasonality (52 weeks) is only observed in temperate regions.The disruption of annual seasonality in the HHS regions can be seen during the 2009 H1N1 pandemic.

BMJFigure
Figure S7.Discrete Fourier transform of ILI ζ -score.The signal of 215-day cycle and 365-day cycle are equivalently strong.