Data

We extracted weekly observed dengue cases from 2014–2020 from the Brazilian Notifiable Diseases Information System (Sistema de Informação de Agravos de Notificação—SINAN) [24] for each of Brazil’s 27 Federative Units (26 states and the capital city, Distrito Federal) (Fig 1). Climate data were sourced from the Brazilian National Institute of Meteorology (Instituto Nacional de Meteorologia—INMET) over the same time period, specifically variables measuring rainfall, humidity, temperature (max, min, median) and wind speed [25]. To approximate changes in care-seeking behavior and care availability, we measured the number of treatments for conditions related to HIV infections (ICD-10 codes B20–24 and Z21) and elective hospitalizations from the Brazilian Ministry of Health (DATASUS) [26], respectively. We used Google Regional Mobility Reports to measure human mobility, specifically the change in time spent in transit stations during the pandemic, relative to a pre-pandemic baseline [27].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Study area.

Map of Brazil with state boundaries, colored by region. Map data from OpenStreetMap, Copyright: Creative Commons, https://www.openstreetmap.org/copyright.

https://doi.org/10.1371/journal.pntd.0012726.g001

Interrupted time series analysis

We developed an interrupted time series analysis (ITSA) [29, 30] to assess the causal impact of the COVID-19 pandemic on observed dengue cases. The interruption time was defined as the first week of March 2020, the start of the pandemic, when the largest mobility reductions were observed (S4 Fig). Accordingly, the data was split into a training set, consisting of all data up to the week of February 29th, 2020, and a forecast period, consisting of all data from the week of March 7th, 2020 until the wek of May 9th, 2020, resulting in a 10-week forecast horizon. Input features included 12 lags of reported dengue cases and all climate variables (rainfall; minimum, average, and maximum temperature; relative humidity; and average wind speed.

We performed four-fold expanding time series cross validation for model selection. The four validation sets each covered a full calendar year from 2016 to 2019, with training performed on all data preceding the validation year. For each validation fold, training and validation datasets were normalized using only information from the training sets, allowing for estimation of out-of-sample performance of the prediction models using mean absolute error (MAE).

During cross-validation, we compared several prediction algorithms, specifically random forest (number of trees: 150, maximum tree depth: 4, minimum number of samples per node: 3), XGBoost (maximum tree depth: 4, minimum number of samples per node: 3, learning rate: 0.05), gradient boosting regression (number of trees: 150, maximum tree depth: 4, minimum number of samples per node: 3, learning rate: 0.05), support vector machine regression (kernel: radial basis function, l2-regularization parameter: 10, tolerance ϵ = 0.1), a feed-forward neural network with single-step output (architecture: 3 layers with 128 units each, 128 x 128 x 128, activation function: rectified linear unit, optimizer: Adam, adaptive learning rate with starting value 0.001), a feed-forward neural network with multi-step output (architecture: 2 layers with 128 units each, 128 x 128, rectified linear unit activation, Adam optimization, and a learning rate of 0.00), a Bayesian structural time series model (using a Gamma prior distribution and previously described parametrization [31]), and a naïve model measuring the seasonal moving average, calculated as the historical average of weekly cases. We fitted separate models for each of the 10 forecast horizons, ranging from 1-week-ahead to 10-week-ahead predictions. From these models, we constructed 3 ensemble models, specifically (i) an average prediction of all models, (ii) an average prediction of the 3 best-performing models on the validation folds (random forest, XGBoost, gradient boosting regression), and (iii) a random forest meta-model that takes as inputs the predictions of all models. We constructed 95% prediction intervals of predictive performance using 200 bootstrapped samples of the predictions on the out-of-sample validation folds for each forecast horizon and state. A summary of the out-of-sample performance on the validation datasets is presented in S1 Table.

We selected the best-performing ensemble on the validation datasets—the average of the top 3 best models, which we trained on the full training set (2014–19) and used to make predictions on the forecast period (10 weeks from March 2020).

To measure the overall impact of the pandemic on dengue in 2020, we compared the predicted dengue cases from the ensemble model (i.e. those that we would have expected in the absence of the pandemic) with those that were actually observed. This expected-observed gap was calculated as the simple difference between observed and predicted cases. Two states, Amapá and Roraima, are excluded from the ITSA analysis, because of missing climate observations during part of 2020. We highlighted results for five sample states in the main text and presented full results in the supplement. The sample states were chosen to cover the five geographic regions of Brazil and to showcase a range of outcomes observed in our analyses.

Reporting adjustment

We implemented a second ITSA to measure the extent to which lower-than-usual dengue cases could be explained by changes in surveillance, specifically by excess under-reporting, driven by reduced access to and availability of care during the pandemic. We leveraged two proxy time series to estimate under-reporting. First, changes in health care availability were approximated by data on elective hospitalizations. These elective procedures declined primarily due to supply-side factors, such as guidance to reserve hospital capacity for treatment of COVID-19 infections [32]. We expect capacity for treatment of mild cases of dengue to follow similar trends as capacity for elective hospitalizations. Second, reductions in care-seeking behavior were approximated by data on HIV-related treatments, which declined primarily due to impediments to accessing care early in the pandemic [15]. We expect behavioral factors reducing demand for dengue treatment to follow similar patterns.

For each of these proxy time series, we compared expected and observed quantities from March until April 2020 using an ensemble of predictions from a multi-step neural network and a Bayesian structural time series model, with architectures and hyperparameters defined as in the first ITSA. The percentage change between expected and observed HIV-related treatments and elective hospitalizations was then used to compute adjusted dengue cases, which may be interpreted as the dengue cases that would have been observed in the absence of changes in care seeking and care availability, respectively. Formally, we computed reporting-adjusted dengue cases d*, as , with the percentage change in reporting u determined by , where d_t are observed dengue cases at time t, a_t is the observed quantity of the reporting approximation, and is the expected quantity of the reporting approximation predicted by ITSA. The under-reporting adjustment was performed only for states in which observed cases were lower than expected by the ITSA predictions.

All analyses were performed in Python version 3.9.13. We used the scikit-learn implementations for all machine learning models and the causalimpact implementation of the Bayesian structural time series (BSTS) model. All code is available at https://github.com/KRoster/impact-of-COVID-19-on-dengue.

Southeast

States in the Southeast observed fewer dengue infections than expected (Table 1), except for Minas Gerais, where observed cases differed from expected cases by less than 1%. In São Paulo, the season started with higher-than-typical incidence in the first half of February (Fig 2 and S1 Fig), but cases began to decline in epidemiological week 8 (February 22nd), resulting in a cumulative gap of 37 thousand fewer cases (34%) than expected in March and April of 2020 (Table 1 and Fig 3). Adjusting for reduced care availability in São Paulo resulted in a case count that closely followed the predicted time series, while reduced care seeking accounted for only 40% of the expected-observed gap (Figs 4 and 5). In Rio de Janeiro, the gap of 3.6 thousand cases (67%) was only partially explained by under-reporting, suggesting that other factors are needed to explain the low observed case count (Fig 4).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Observed and expected dengue cases in five sample states.

Observed dengue cases (black) and 1- to 10-week forecasts (teal) with 95% prediction intervals (shaded area).

https://doi.org/10.1371/journal.pntd.0012726.g003

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Proportion of the observed-expected gap explained by under-reporting.

Total gap between expected and observed cases in March and April 2020 (black boxes) and the proportion of the gap that is explained by under-reporting adjustments using HIV-related treatments (orange) and elective hospitalizations (purple) proxies. Columns show (i) states where observed cases are lower than expected cases, and at least one of the two adjustments is within the observed-expected gap (Observed < adjusted < expected) (left column) and (ii) states where the adjustment over-corrects for the gap, such that the adjustment is greater than the prediction (Observed < expected < adjusted) (right column). All x-values are expressed as shares of the observed-expected gap.

https://doi.org/10.1371/journal.pntd.0012726.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Observed, expected, and under-reporting-adjusted dengue cases over time in sample states.

Observed dengue cases (dark gray), predicted dengue cases (teal), and under-reporting-adjusted cases using HIV-related treatments (orange) and elective hospitalizations proxies (purple).

https://doi.org/10.1371/journal.pntd.0012726.g005

The government of Espírito Santo stopped using the online disease notification system in 2020 [24], leading to clearly evidenced and sudden under-reporting (S1 Fig), which helped validate our ITSA results. The models identified this break in the data generation mechanism and predicted higher cases than reported (S2 Fig).

South

All three states in the South reported much larger dengue outbreaks in 2020 than in prior years (Fig 2). Santa Catarina observed 3,856 (307%) more cases than expected and Rio Grande do Sul observed 890 (54%) more than expected (Table 1 and Fig 3). Paraná also reported atypically high dengue cases (Fig 2), though the excess cases occurred largely before the start of the pandemic, peaking in the week of March 14th (S1 and S2 Figs).

Center-West

Dengue cases in three of four federative regions in the Center-West were higher than observed in previous years (in Mato Grosso, Distrito Federal, Mato Grosso do Sul, see S1 Fig). However, cases were still lower than predicted from the ITSA in Mato Grosso do Sul (29%) and in Mato Grosso (21%) (Table 1 and S2 Fig). Observed cases were roughly as expected in Goiás (4% more than expected) and Distrito Federal (6% fewer than expected) in spite of a notable dip in cases around the onset of the pandemic (Table 1 and S2 Fig). Under-reporting either fully explained (e.g. adjustment for care access in Mato Grosso) or over-corrected for the observed-expected gap, especially in Goiás (Fig 4 and S3 Fig).

Northeast

Eight of the nine states in the Northeast experienced fewer cases than predicted (Table 1) and than had been seen in prior years (S1 Fig). In half of these states, under-reporting fully explained and over-corrected for the expected-observed gap (Ceará, Piauí, Rio Grande do Norte, and Pernambuco), while under-reporting only partially explained the gap in the other half of the states (Maranhão, Paraíba, Alagoas, Sergipe) (Fig 3). Ceará, for example, began the season with typical incidence, followed by a dip in cases in mid-March, and remained below typical incidence for the rest of the dengue season (S1 Fig). The gap between expected and observed cases in March and April amounted to 2,630 cases (27%) (Table 1), which was fully explained by both reporting adjustments (Fig 4). Bahia deviated from the other states in the Northeast, with its higher-than-expected incidence (Table 1). However, the outbreak size in Bahia was not as unprecedented as in the South, since 2020 case counts were almost as large as during the 2016 outbreak and the outbreak timing was comparable to the 2019 season.

North

The Northern region experiences more variation in the timing of dengue cases than other parts of Brazil and outbreaks tend to be less intensive and less seasonal, making prediction of dengue cases in this region more challenging [12, 19]. Relative to other regions, outbreaks in the Northern states tend to peak earlier (December-February) [19] (S1 Fig) and are thus less likely to have been impacted by the pandemic. Observed time series were comparable to previous years in all states except for the state of Amazonas, which observed higher case counts than usual in January (S1 Fig). Observed cases in all states were both higher and lower than expected at some time points in the predictions (S2 Fig), with total observed cases lower than expected in all states except for Rondônia (Table 1). The size of the gap ranged from 6% in Tocantins to 48% in Acre (Table 1). Acre was the only state in the North where under-reporting adjustment did not explain the expected-observed gap (Fig 3).