Cross-national prediction of imported dengue risk in Northeast Asia using climate, hydrological, and viral epidemiological indicators

Abstract

Imported dengue remains a public health concern in non-endemic countries. Local transmission may occur when imported infections coincide with competent vectors and suitable environmental conditions. We developed a cross-national framework for predicting monthly imported dengue risk in Japan and the Republic of Korea. We analyzed monthly data from January 2005 to June 2025. The period from 2020 to 2022 was excluded because of coronavirus disease 2019-related structural changes. The models were fitted to the log-transformed travel-volume-adjusted imported dengue case risk, and predictions were back-transformed to monthly case counts using outbound travel volume. Candidate predictors included Pacific Ocean sea surface temperature indices, the Dipole Mode Index, monsoon indices, source-region Standardized Precipitation–Evapotranspiration Index values, and dengue serotype dominance persistence. Candidate lags were selected from prespecified predictor-specific lag windows using the Akaike Information Criterion. Elastic Net regression, Random Forest, Support Vector Machine, and eXtreme Gradient Boosting were evaluated across 72 predictor sets for each country. A trend-and-seasonality-only reference model was used to evaluate the additional predictive information provided by the candidate predictors. Selected base-model predictions were combined in stacked ensemble models using Elastic Net regression and Random Forest as meta-learners. The selected Elastic Net-based stacked ensemble outperformed the reference model in both countries. In Japan, root mean squared error decreased from 23.875 to 7.220, while mean directional accuracy increased from 48.3% to 65.5%. Similar improvements were observed in the Republic of Korea, with root mean squared error decreasing from 28.812 to 6.313 and mean directional accuracy increasing from 44.8% to 65.5%. Within the fitted base models, the 6-month-lagged Oceanic Niño Index in Japan and the 5-month-lagged Niño 3 Index in the Republic of Korea showed relatively high predictive relevance. Thailand Standardized Precipitation–Evapotranspiration Index also ranked highly within several selected base models. An interactive R Shiny application was developed for scenario-based prediction. A framework combining climate, hydrological, and viral epidemiological indicators performed better for predicting imported dengue than a model based on temporal trends and seasonality alone. These findings support a multidimensional approach to preparedness for imported dengue in non-endemic Northeast Asia.