Outbreak pattern
The observation of the outbreaks of dengue, chikungunya and Zika in Colombia and Mexico suggests that arboviruses present a similar pattern at the start of the outbreak. The increase of the chikungunya and Zika curves was preceded by a dengue peak, possibly due to diagnostic confusion (see Figs.
4,
5). It can be assumed that during the first weeks of the chikungunya outbreak in Colombia, the unusual increase of dengue cases (172 cases in week 44 of 2014) was possibly due to chikungunya infections that remained unconfirmed by the laboratory. The later chikungunya cases were correctly diagnosed and the epidemic curve rapidly increased.
The same happened when the Zika outbreak occurred. Two weeks before the peak of Zika cases, there was an increase in “dengue cases”. Many of these were probably Zika infections. Only after the laboratory confirmation of the Zika virus in the national laboratory, the correct diagnoses were notified and the number of reported Zika cases increased. A similar but less pronounced phenomenon was observed in Mexico in 2014/15: at the beginning of the chikungunya outbreak, there was an increased dengue activity probably because the medical doctors did not consider chikungunya to be a diagnostic option. The same happened with the Zika outbreak in 2016 when first dengue case numbers increased and subsequently Zika was detected. The delay in the confirmation of cases can lead to the silent spread of the new disease, which makes control difficult. This situation highlights the need for an improved surveillance system with a focus on early outbreak warning.
The observation that Aedes borne arbovirus outbreaks seem to mutually suppress each other warrants further confirmation.
Dengue, chikungunya and Zika have a seasonal transmission pattern which is accounted for in the present EWARS model. However, there are also important variations over the years depending on response activities by the vector control services and climatic change. These variables are now taken care of in a further developed EWARS model, which is being tested at the moment and will be published later.
The propose of Early Warning and Response System (EWARS)
The EWARS tool is primarily aimed at supporting district health managers and national health planners to mitigate or prevent disease outbreaks, ideally using tools that are integrated in the national surveillance programs [
26]. To further ensure effective functions, the EWARS should be perceived as an information system designed to support the decision-making of national- and local-level institutions but also enable vulnerable groups in the society to take actions to mitigate the impacts of an impending risk [
26].
Recent analysis of the evidence indicates that early warning and response system that are capable of demonstrating evidence of prospective predictive ability and allows technical and practical adaptations of local public health responses while augmenting communications channels between users at central and district levels are tools that are more likely to be implemented into national surveillance programs [
27]. In this sense, EWARS has moved towards frameworks that facilitate low-cost IT maintenance and adapt to unskilled users. The aim is to form a tool that can be plausibly integrated into existing national systems [
25,
27].
Predictive abilities of the Early Warning and Response System
In an effort to more effectively prepare for and prevent arbovirus disease outbreaks, the EWARS appeared to adequately predict outbreaks of dengue (3 to 5 weeks), chikungunya (10 to 13 weeks) and Zika (6 to 10 weeks) ahead of time in Colombia. In Mexico, the lag time was relatively shorter for Zika with 4 to 5 weeks ahead of the outbreak (not tested for dengue and chikungunya). The variation of lag times for the three diseases may be due to different extrinsic incubation periods (i.e. the time the virus replication and passage from the mosquito gut to the salivary gland requires) but could also be due to the different delays caused by health services diagnose of the diseases [
10]. A lag time of 6 to 10 weeks ahead of disease outbreaks would allow timely public health services response with enhanced vector control measures. Nevertheless, shorter time periods for preparing response activities will require the definition of high transmission “hot spots” where interventions can be targeted. The better the control programmes are prepared to identify such high transmission areas and step up the response quickly, the better the chance to mitigate or even avert an outbreak. This has been shown for Mexico [
28] and has triggered the incorporation of risk maps into the EWARS tool.
For dengue outbreaks, the sensitivity of alarm indicators to correctly predict an outbreak varied in both countries, with a validity ranging between 74 and 92% using multiple meteorological indicators. The importance of meteorological alarm indicators underlines the characteristic climate sensitivity of vector borne diseases. This pattern is similar to that observed in other studies, which used EWARS for dengue outbreak prediction (83–99% sensitivity in Brazil, 50–99% in Malaysia and 79–100% in Mexico) [
14]. Likewise, the PPV of up to 68% in Colombia and up to 83% in Mexico was similar to those reported in previous reports (40–88% in Brazil, 71–80% in Malaysia and 50–83% in Mexico) [
14]. The sensitivity and PPV as statistical measures of EWARS performance have meaningful operational implications with the sensitivity indicating the validity of the tool in detecting and predicting outbreaks in time. The PPV, however, can inform local health district managers of potential economic consequences of failing positive alarms—e.g. for 70% PPV, one can be certain that about 30% of resources deployed would not be efficiently used. This is a crucial measure which has further directed the EWARS design to employ a step-wise approach of vector control and response to ensure more efficient application of the tool mainly in resource-limited settings (‘initial’ response: is declared when two consecutive alarm signals occur; ‘early’ response: is declared when three consecutive alarm signals occur; and ‘late’ response: is declared when more than three consecutive outbreak weeks take place). Nevertheless, even in the case of false alarms, resources on vector control are not spent in vain as they contribute in any case to keeping vector densities down.
In the case of chikungunya, the sensitivity to correctly predict an outbreak varied among alarm indicators from 77 to 93%, being rainfall and the combination of meteorological indicators the alarm signals with the highest sensitivity. Likewise, the PPV of up to 85% was similar to those reported with dengue in other countries [
14]. Therefore, the EWARS would predict outbreaks 10 to 13 weeks in advance, providing adequate time to activate response actions. In the case of Zika, the sensitivity (50–100%) and PPV (11–100%) were similar to the prediction of dengue and chikungunya when using alarm indicators with the highest values for sensitivity and PPV. While several studies are existing in the literature to demonstrate the applications of prediction models for dengue outbreaks, the EWARS is also useful for Zika and chikungunya outbreak prediction. Based on a recent scoping review study [
27], five models showed outstanding performance in dengue outbreak prediction; (i) the dynamic risk maps absolute shrinkage and selection operator (LASSO) processing multiple meteorological information; (ii) the auto-regression integrated moving average (ARIMA) using meteorological information and Google Trends data; (iii) the Shewhart moving average regression model (SMAR) maintaining a combination of meteorological, epidemiological, and entomological alarm indicators, which is the model employed in our study; (iv) the seasonal autoregressive integrated moving average (SARIMA); and (v) the stochastic Bayesian maximum entropy (BME) model using a mix of meteorological and entomological alarm indicator. Despite the fact that LASSO models seem to be toping the tools performance, these models were declared unamenable to easy and direct interpretation and usually demand advance-level of historical data, which limits their applications. No studies were retrieved from the literature to support the statistical prediction performance of Zika and chikungunya outbreaks.
This is the first study to evaluate TDR-EWARS using chikungunya and Zika surveillance data against meteorological and entomological alarm information, and the overall study findings were promising towards implementing the tool at national level. However, outbreak prediction in low endemic areas is less optimal and should be carefully interpreted in relation to routine vector control and response. Despite the fact that 3-year data records were sufficient to demonstrate the applicability of EWARS to broader
Aedes borne arbovirus diseases, using longer historical surveillance records has the potential to affect the definition of outbreaks and consequently impact on the sensitivity and PPV. As the user-friendliness of the application has already been established [
14], the next step is to bring it to practical use in endemic countries and monitor its ability to predict outbreaks and trigger effective response.
Limitations of the study
This study was based on the cases registered by the public health surveillance system in Colombia, which classifies the cases as probable, confirmed and hospitalized [
1,
19‐
21]. All confirmed cases should present a positive laboratory test, but in the current Colombian database the confirmed cases were sometimes based on clinical criteria. For chikungunya in Colombia, 106,592 cases were reported by SIVIGILA in 2014, 98% of which were confirmed cases according to clinical criteria [
19‐
21]. This reduces the reliability of the case definition. For dengue, the use of hospitalized cases as outbreak indicator was feasible, but in chikungunya and Zika this was hardly the case due to the low proportion of “hospitalized cases”. This limitation was overcome by taking into account probable and confirmed cases as outbreak indicators a suggested by other authors [
14,
19‐
21,
23,
24]. Furthermore, there is a possibility of overestimation caused by the out-of-sample prediction, which may explains the case of 100% sensitivity observed. This can be reduced by employing a longer-data history of disease and alarm information.
In Cúcuta, more than 23,000 cases of chikungunya occurred, but the majority were diagnosed collectively (“collective reporting” i.e. all patients with fever and other symptoms in the waiting area of a health service are diagnosed as having chikungunya), and only a small proportion of cases had complete information from individualized examination [
19,
22] which is the routine approach in Mexico. Collective reporting demonstrates the impact of the chikungunya outbreak on the overstretched health system in Colombia. However, the data analyzed in this study was collected throughout the epidemic underlining the ability of the surveillance system to function under difficult circumstances [
19,
29].
The time period for retrospectively testing the validity of the EWARS tool was relatively short (3 years in the case of Zika); longer retrospective observation periods would have reflected better the usual pattern of the disease which is not possible in case of a newly emerging disease.