R02.2 Epidemic intelligence data and disease risk mapping: the case of Crimean-Congo haemorrhagic fever (CCHF)

Abstract

The Epidemic Intelligence from Open Sources (EIOS) (WHO, 2023) is a system used for scanning information from open sources to detect disease events. The initiative, led by the World Health Organization (WHO), was built on the “One Health, All Hazards principle”.

This study uses the EIOS for complementing data collection of neglected diseases thus improving the understanding of their geographic extent and level of risk. A case study on Crimean-Congo Haemorrhagic Fever (CCHF) in the European Region is presented.

Data on CCHF occurrence (January 2012–March 2022) were retrieved using an algorithm to detect mentions of CCHF cases in humans or vectors. A Bayesian additive regression trees (BART) model (Chipman et al., 2010) was built to map the risk of CCHF occurrence. Thirty-three explanatory variables retrieved from WorldClim, ENVIREM and PROBAV_V2.2.1. Every predictor was rescaled at a resolution of 4 x 4 km, aligned, and reprojected to EPSG:4326-WGS84. For each pixel a posterior distribution of predicted probabilities with the associated 95% credible intervals (CI) was obtained. The most important variables were selected by initially fitting the model using all variables. An automated stepwise reduction algorithm with 50 iterations and 10 trees was used to eliminate the variables with the lowest importance and obtain the model with the lowest root mean square error (RMSE). The final model was run with the reduced variable set, 200 trees, and 1000 posterior draws after a burn-in of 100 draws.

Overall, a general decreasing risk trend from south to north across the entire European Region was detected. The results show a positive association between all the temperature-related variables and the probability of CCHF occurrence, with an increased risk in warmer and drier areas. In particular, the Mediterranean basin and the areas bordering the Black Sea are the areas at the highest risk of CCHF occurrence under the current environmental conditions. The model performed well in terms of accuracy (AUC = 0.95), correct classification rate (CCR = 0.80), correct prediction of presences (sensitivity/recall = 0.99), correct prediction of absences (specificity = 0.79), true skill statistic (sTSS = 0.89), Cohen’s kappa (skappa = 0.69), but the positive predictive value was low (precision = 0.31). The low precision means that the model predicts positive pixels where there are no actual observations of human cases or virus isolation from ticks. These are represented by the locations which are likely to be favourable for CCHF occurrence, but where the disease has yet to be reported.

This work demonstrates the potential of internet-based data accessed through the EIOS system to create a disease risk map. The case study presented may constitute a model to combine epidemic intelligence tools and advanced analytical approaches to detect and assess changes in disease distribution, allowing prevention and early mitigation of disease events of veterinary and public health importance.