Scholarlink

<p>Dengue fever in Vietnam is increasingly influenced by climate variability, yet most early warning systems still rely on single-model, province-level forecasts that overlook lagged climate drivers, spatial spillover, and extreme-outbreak risk. This study develops and evaluates an integrated, climate-informed dengue forecasting framework for Ba Ria–Vung Tau (BRVT), a coastal province in southern Vietnam, using weekly district-level dengue surveillance data and 17 satellite-derived meteorological variables from 2003 to 2022. We first applied distributed lag nonlinear models, mutual information, and Granger-style causality tests to identify a compact set of 11 lagged climate predictors—dominated by humidity, soil moisture, wet-bulb temperature, and rainfall at 1–4-week lags—that are both statistically robust and epidemiologically plausible drivers of dengue. Spatiotemporal decomposition (STL, wavelet analysis) and hotspot mapping (Getis–Ord Gi* with kriging) revealed a rising trend in incidence, strengthening annual and 2–3-year cycles, and persistent high-risk clusters centred on Vung Tau and Ba Ria cities. Building on these diagnostics, we compared a suite of machine-learning and deep-learning models using a fixed chronological split (training: January 2003–December 2019; validation: January 2020–December 2022). Classical tree-based ensembles, Gaussian process regression, and recurrent networks achieved MAE ≈ 4.5–5.4 and R² ≈ 0.46–0.57, whereas the proposed Time-Series Transformer and graph neural networks (GNNs) substantially improved performance (Transformer: MAE 3.56, RMSE 5.99, R² 0.813; GNNs: MAE 2.83, RMSE 4.92, R² 0.874), with similar training and validation errors indicating limited overfitting. Finally, Extreme Value Theory was used to fit a Generalized Extreme Value distribution to annual maxima of weekly incidence, yielding Weibull-type tail behaviour, return levels, and exceedance probabilities that quantify the expected frequency of rare but severe dengue seasons. By combining causally screened climate predictors, spatiotemporal deep learning, and tail-risk metrics within a single pipeline, this framework provides operationally useful weekly forecasts and extreme-outbreak risk indicators to support early warning, spatially targeted vector control, and surge planning in dengue-endemic settings.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Causal and spatiotemporal deep learning for dengue forecasting and extreme outbreak risk under climate variability: a framework from Vietnam

  • 期刊论文

  • 发表日期：

    2026年3月20日

查看全文
  • Dang Anh Tuan

摘要

Dengue fever in Vietnam is increasingly influenced by climate variability, yet most early warning systems still rely on single-model, province-level forecasts that overlook lagged climate drivers, spatial spillover, and extreme-outbreak risk. This study develops and evaluates an integrated, climate-informed dengue forecasting framework for Ba Ria–Vung Tau (BRVT), a coastal province in southern Vietnam, using weekly district-level dengue surveillance data and 17 satellite-derived meteorological variables from 2003 to 2022. We first applied distributed lag nonlinear models, mutual information, and Granger-style causality tests to identify a compact set of 11 lagged climate predictors—dominated by humidity, soil moisture, wet-bulb temperature, and rainfall at 1–4-week lags—that are both statistically robust and epidemiologically plausible drivers of dengue. Spatiotemporal decomposition (STL, wavelet analysis) and hotspot mapping (Getis–Ord Gi* with kriging) revealed a rising trend in incidence, strengthening annual and 2–3-year cycles, and persistent high-risk clusters centred on Vung Tau and Ba Ria cities. Building on these diagnostics, we compared a suite of machine-learning and deep-learning models using a fixed chronological split (training: January 2003–December 2019; validation: January 2020–December 2022). Classical tree-based ensembles, Gaussian process regression, and recurrent networks achieved MAE ≈ 4.5–5.4 and R² ≈ 0.46–0.57, whereas the proposed Time-Series Transformer and graph neural networks (GNNs) substantially improved performance (Transformer: MAE 3.56, RMSE 5.99, R² 0.813; GNNs: MAE 2.83, RMSE 4.92, R² 0.874), with similar training and validation errors indicating limited overfitting. Finally, Extreme Value Theory was used to fit a Generalized Extreme Value distribution to annual maxima of weekly incidence, yielding Weibull-type tail behaviour, return levels, and exceedance probabilities that quantify the expected frequency of rare but severe dengue seasons. By combining causally screened climate predictors, spatiotemporal deep learning, and tail-risk metrics within a single pipeline, this framework provides operationally useful weekly forecasts and extreme-outbreak risk indicators to support early warning, spatially targeted vector control, and surge planning in dengue-endemic settings.