Title : An explainable Long Short-Term Memory Architecture-Based Dengue Prediction at 1-KM resolution
Abstract:
The surge of dengue cases globally, with nearly half of the world’s population at risk of contracting it, remains global public health concern. Malaysia, a tropical country from Southeast Asia, experiences recurrent outbreaks despite of ongoing surveillance and vector control efforts. Considering current system is reactive, it necessitates the development of accurate, and fine-resolution dengue prediction model to support early and effective intervention. Hence, this study aimed to develop Long Short-Term Memory (LSTM) architecture at 1-km spatial scale and weekly dengue prediction in Petaling district, Malaysia. Dengue cases were collected from Ministry of Health, Malaysia spanning 2015 to 2024. Climate and environmental variables were collected from satellite data. The collected data were cleaned and standardised to weekly and 1-km scale. Log transformation was performed for dengue case counts. Min-Max scaler was used for normalisation of data. Hyperparameters were optimised using Optuna. Sliding window of 4-, 8-, 12- and 16-weeks was evaluated for both single layer and stacked LSTM architectures. Each spatial unit was represented with unique grid identifier which resulted 276,667 grid-week observations for Petaling. The lowest validation loss was observed at 16-weeks of sliding window for both single layer and stacked LSTM architectures. However, only 0.13% improvement was observed in stacked LSTM architectures than single layer LSTM. The single-layer LSTM achieved a slightly lower root mean squared error (RMSE = 0.981) than the stacked LSTM (RMSE = 0.984), implies marginally better predictive performance. Shapley Additive exPlanations (SHAP) analysis demonstrated consistent predictor hierarchies across both the architectures. Previous dengue incidence was the dominant predictor in both models, followed by mean rainfall and maximum rainfall. The Spearman correlation between the both architectures SHAP rankings was 0.983. It implies strong agreement between architectures. This study highlights extending the sliding window improved the dengue prediction performance. While increasing the depth of architecture resulted comparatively small improvement in performance. Hence, considering the additional computational time for deeper architecture, the single layer LSTM may offer a better practical approach for real time dengue prediction than increasing the depth of an architecture.
