Scholarlink

Background <p>Dengue transmission in Indonesia is shaped by interacting climatic, environmental, and socio-demographic factors, yet most forecasting systems remain static and vulnerable to data shifts. There is a critical need for adaptive, data-driven early-warning frameworks that integrate multiple predictor domains while preventing methodological biases such as information leakage. This study aimed to develop a Random Forest (RF)–based predictive model embedded within an Agile System Development workflow to forecast monthly dengue case counts in Yogyakarta.</p> Methods <p>Monthly dengue case counts from five districts (2017–2022) were modeled using multi-domain predictors. All preprocessing steps—including imputation, standardization, correlation screening, VIF diagnostics, and Negative Binomial GLM–based feature screening—were performed exclusively on the 2017–2021 training subset, with parameters applied unchanged to the 2022 test set. The GLM served solely as a leakage-free exploratory screening tool. A Random Forest model was trained using optimized hyperparameters (500 trees, max depth 10) and evaluated through temporal testing. Model reliability was assessed using calibration curves, prediction-interval metrics, and a one-month early-warning classification evaluated with sensitivity, specificity, PPV, and NPV.</p> Results <p>The RF model achieved strong predictive performance (R² = 0.86; RMSE = 5.72), exceeding the GLM benchmark (R² = 0.64). Rainfall lag-1, temperature, and humidity emerged as dominant predictors, complemented by built-up area and population density. Calibration indicated good agreement across routine transmission ranges, with reduced reliability during outbreak peaks. The early-warning component demonstrated high sensitivity (0.82) and strong negative predictive value (0.86), supporting its use as a decision-support indicator of elevated transmission risk.</p> Conclusion <p>The proposed Agile–AI framework demonstrates the potential to deliver accurate dengue risk predictions with interpretable uncertainty estimates within a flexible, multi-domain early-warning architecture. While external validation and further refinement are required, the framework offers a scalable foundation for adaptive dengue surveillance and targeted vector-control decision support in dynamic tropical 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

AI-based prediction of dengue incidence using climatic, environmental, and socio-demographic factors: an ensemble random forest approach with agile system development

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月4日

查看全文
  • Tri Baskoro Tunggul Satoto,
  • Nur Alvira Pascawati,
  • Roger Frutos,
  • Erizal,
  • Triwibowo Ambar Garjito,
  • Marko Ferdian Salim

摘要

Background

Dengue transmission in Indonesia is shaped by interacting climatic, environmental, and socio-demographic factors, yet most forecasting systems remain static and vulnerable to data shifts. There is a critical need for adaptive, data-driven early-warning frameworks that integrate multiple predictor domains while preventing methodological biases such as information leakage. This study aimed to develop a Random Forest (RF)–based predictive model embedded within an Agile System Development workflow to forecast monthly dengue case counts in Yogyakarta.

Methods

Monthly dengue case counts from five districts (2017–2022) were modeled using multi-domain predictors. All preprocessing steps—including imputation, standardization, correlation screening, VIF diagnostics, and Negative Binomial GLM–based feature screening—were performed exclusively on the 2017–2021 training subset, with parameters applied unchanged to the 2022 test set. The GLM served solely as a leakage-free exploratory screening tool. A Random Forest model was trained using optimized hyperparameters (500 trees, max depth 10) and evaluated through temporal testing. Model reliability was assessed using calibration curves, prediction-interval metrics, and a one-month early-warning classification evaluated with sensitivity, specificity, PPV, and NPV.

Results

The RF model achieved strong predictive performance (R² = 0.86; RMSE = 5.72), exceeding the GLM benchmark (R² = 0.64). Rainfall lag-1, temperature, and humidity emerged as dominant predictors, complemented by built-up area and population density. Calibration indicated good agreement across routine transmission ranges, with reduced reliability during outbreak peaks. The early-warning component demonstrated high sensitivity (0.82) and strong negative predictive value (0.86), supporting its use as a decision-support indicator of elevated transmission risk.

Conclusion

The proposed Agile–AI framework demonstrates the potential to deliver accurate dengue risk predictions with interpretable uncertainty estimates within a flexible, multi-domain early-warning architecture. While external validation and further refinement are required, the framework offers a scalable foundation for adaptive dengue surveillance and targeted vector-control decision support in dynamic tropical settings.