Cameroon’s sentinel influenza surveillance network has been operational since 2007. Like in many other countries, it aims to track the spread of influenza viruses and other respiratory viruses of public health importance, such as HRSV and, more recently, SARS-CoV-2. This surveillance system now includes 19 sites spread across all 10 administrative regions of the country. Some of these sites have an ILI (Influenza-Like Illness) surveillance program, some have SARI (Severe Acute Respiratory Infection) surveillance, and some have both.

For this study, we used data collected at sentinel sites in Yaoundé, the capital of Cameroon, over 3 years (from July 2020 to December 2022). Yaoundé has a population of around 4.5 million and a tropical climate profile, albeit tempered by altitude, with four seasons: two rainy seasons and two dry seasons. The sites used in our analysis included two SARI sites, the “Centre Hospitalier d'Essos” and the “Hôpital Jamot”, as well as three ILI sites, including the “Centre Médico-Social Ambassade de France”, the “Centre d’Animation Sociale et Sanitaire de Nkolndongo” and the “Centre Médical d'Etoudi”. All laboratory analyses to confirm HRSV cases were performed at the Centre Pasteur du Cameroun (CPC) (Figure 1).

All patients presenting with symptoms of ILI or SARI at the focal sites were recruited. As previously reported [10], the WHO case definition was used to identify cases of ILI or SARI. Patients of all ages with signs of ILI (an acute respiratory infection with fever measured at a temperature of 38 °C and cough occurring within the previous 10 days) or SARI (an acute respiratory infection with fever or fever measured at a temperature of 38 °C and cough occurring within the previous 10 days and requiring hospitalization) were, in brief, the site-level inclusion criteria [11].

Each enrolled patient provided a nasopharyngeal and/or oropharyngeal sample. After collection, samples were preserved and transported to the virology department of the CPC for immediate processing or storage at 4 °C until analysis according to WHO guidelines [12]. All procedures for handling biological samples have been previously documented [10]. Briefly, nasopharyngeal swabs were tested for influenza A and B, HRSV, and SARS-CoV-2 using real-time RT-PCR assays.

Meteorological data considered in this study included earth surface temperature (EaST), the temperature at 2 m (T_2M), wet bulb temperature at 2 m (WbT_2M), specific humidity at 2 m (SH_2M), relative humidity at 2 m (RH_2M), dew/frost point at 2 m (DFPt_2M), average corrected precipitation (AvCP), cumulative rainfall (CumRF), and solar radiation (SolRad). The variables were collected at monthly intervals from ground stations in the city of Yaoundé, located at latitude 3°50′29.04″N and longitude 11°29′31.56″, and downloaded from the National Aeronautics and Space Administration (NASA) database at https://power.larc.nasa.gov/data-access-viewer.

Data on positive cases of HRSV for this study were obtained from 2020 to 2022. Monthly counts of positive cases were used as dependent variables, while monthly values of 12 climate variables were used as independent variables. These data were combined in an Excel spreadsheet and then imported into SPSS software version 26.0 (IBM, Chicago, IL, USA) for analysis. A comparative analysis of two time series models was used: the Autoregressive Integrated Moving Average (ARIMA) model and Holt-Winters’s additive model. The ARIMA model was used to examine the interactions between each climate variable and overall viral activity, whereas the additive Holt-Winter model, which is designed to capture seasonal patterns, does not account for climate variables.

ARIMA models are suitable for analyzing and forecasting time series data, based on three fundamental components: (1) autoregressive (AR) models, (2) integration (I), which determines the level of stationarity of the variable under study, and (3) the moving average (MA), which expresses the current value of the variable in question. The ARIMA model has three parameters: P, D, and Q. These are usually written as (p, d, q) (P, D, Q)m. In this notation, p and P refer to the autoregressive lags, q, and Q to the moving average lags, d and D to the differencing orders, and m to the seasonal cycle.

ARIMA models used the Box-Jenkins method. We used the Augmented Dickey-Fuller (ADF) test to see if the monthly HRSV case series was stable. Differencing was not necessary (d = 0, D = 0). The seasonal period was set at 12 months to match the annual cycle. Potential orders were determined through the application of autocorrelation function (ACF) and partial autocorrelation function (PACF) plots. The final model selection was then determined using the corrected Akaike Information Criterion (AICc) and the principle of parsimony. This approach led to the selection of an ARIMA (0,0,0) (0,0,0) model, which is characterized by a constant structure, for the multivariate model that included all twelve climate variables. As a result, the climate variables were used as direct predictors, without any additional autoregressive or moving average components.

The Winter’s additive seasonal model was used to describe whether the HRSV data showed a consistent seasonal pattern over time with no trend. This model has zero orders of autoregression, one order of differencing and seasonal differencing, and orders 1, p, and p + 1 of the moving average, where p is the number of periods in a seasonal interval and p equals 12 for monthly data. The smoothing parameters—level (α), trend (γ), and seasonality (δ)—were used in a maximum likelihood estimation to determine them.

The performance of the time series models was compared using graphs, stationary R², and error metrics such as root mean square error (RMSE).

We carried out a final execution of the out-of-sample series test to evaluate the quality of the forecasts. First, we segmented the 30-month period, spanning from July 2020 to December 2022, into two distinct sets. The first part, covering 18 months from July 2020 to December 2021, was used to determine the model. The second part, on the other hand, constitutes a test set encompassing the last 12 months, from December 2021 to December 2022. This latter set is used to measure the accuracy of the formulated forecasts.

Supplementary Figure S2 also illustrates a flowchart that summarizes the entire analytical process, from data preprocessing to model training, followed by the evaluation of its performance.

Supplementary Table S1 shows the socio-demographic characteristics of the patients who participated in the study. All participants were residents of Yaoundé, and the year 2021 had a higher percentage of cases (57.0%) compared to other years. The 30-65 age group was the most represented, while the over-65 age group was the least represented, according to the population’s age distribution. Males and females represented 46.4% (823/1774) and 44.2% (784/1774) of the registered patients, respectively, while 9.4% (167/1774) were of unspecified sex. The HRSV positivity rate was 8.5%.

During the study, we observed a heterogeneous pattern of virus circulation from year to year. The temporal distribution of the virus is shown in Supplementary Figure S1. The months with the highest HRSV peaks were December 2020, June 2021, April 2022, and December 2022.

No significant correlation (p < 0.05) was observed between individual climatological variables and HRSV activity (Table 2). With a RMSE of 12.94, the model incorporating the full set of climatological factors demonstrated superior performance, accounting for 56.4% of the variance in HRSV activity at this site. Figure 3 provides a graphical representation of the ARIMA models with all climate variables included as independent predictors in Yaoundé, Cameroon.

The Holt-Winter’s Additive model was the best model for HRSV activity, explaining 72.6% of HRSV activity (Supplementary Table S2). Furthermore, a lower RMSE value for the Winters Additive model (7.289) indicates a better performance of this model in monitoring the spread of HRSV. In addition, a statistically significant correlation (P < 0.001) was observed between seasonal changes and HRSV activity. Figure 4 shows the graphical representation of the ARIMA models with all climate variables as independent predictors in Yaoundé, Cameroon.

To assess the generalization capacity of the forecasts, both models were employed to estimate the number of HRSV infection cases over a 12-month observation period. In this context, the additive Holt-Winters model demonstrated superior effectiveness compared to the ARIMA model in terms of predictive performance. The RMSE associated with the Holt-Winters model was 20.50, whereas that of the ARIMA model was 27.07, as shown in Supplementary Table S3. These results indicate that the seasonal model can predict HRSV activity in this study with high accuracy, its error rate being lower by 24%.