Explainable Spatio-Temporal Graph Neural Networks for Dengue Outbreak Prediction: Environmental Data Integration and Policy Recommendation Framework in Yogyakarta, Indonesia
Diyah Utami Kusumaning Putri, Dinar Nugroho Pratomo, Guntur Budi Herwanto, Annisa Maulida Ningtyas, Aldilas Achmad Nursetyo, Ahmad Watsiq Maula, Anis Fuad, Citra Indriani, Riris Andono AhmadAbstract
Background
Dengue fever is among the most rapidly spreading vector-borne diseases globally and a critical public health challenge in Indonesia, requiring evidence-based decisions supported by interpretable artificial intelligence (AI) systems rather than black-box prediction models with limited practical utility.
Objective
To demonstrate how an explainable Spatio-Temporal Graph Neural Network (STGNN) can generate actionable policy recommendations for dengue control by integrating environmental and surveillance data through interpretable AI mechanisms.
Methods
We developed Explain Dengue, a web-based platform implementing an STGNN with multi-head temporal attention, a graph attention network (GAT), and bidirectional Long Short-Term Memory (LSTM) networks, trained on dengue surveillance data (2021–2025) from Yogyakarta Province, Indonesia, using a chronological 70/10/20 train/validation/test split and an explicit 4-week forecast horizon, validated on 1300 weekly observations across five regions.
Results
Across five independently trained models, the STGNN achieved a mean (SD) test MAE of 7.50 (±0.48) cases/week and R2 of 0.265 (±0.102), outperforming seasonal-naive, gradient-boosting, and LSTM-only baselines. Using a statistical endemic-channel risk classification method (binary alert/no alert), the system achieved 77.6% accuracy (84.6% recall, 45.2% precision for alerts). Learned spatial attention was highly reproducible across seeds (r = 0.95), but an ablation-based faithfulness check found it was not a reliable causal explanation of predictions (r = −0.11), a limitation we discuss transparently.
Conclusion
A strictly held-out STGNN can outperform baselines for weekly dengue forecasting, offering a more evidence-based foundation for deploying explainable AI in vector-borne disease control, while surfacing an important caveat about the causal interpretability of learned attention.