Speaker
Description
Background & aims of study
With the increasing risk of malaria transmission driven by climate change and continued imported cases, understanding malaria dynamics from surveillance data has become important for disease control and public health preparedness. This study aims to infer incubation-related temporal structures and predict malaria incidence from time series using long short-term memory (LSTM) models with attention mechanisms, with validation using synthetic data from an extended SEIR model.
Methods & results
Weekly malaria incidence data from Korea spanning 2013–2024 were used to develop and evaluate the proposed models. The attention mechanism assigns varying importance to past observations, enabling identification of incubation-related temporal patterns. The analysis revealed two dominant peaks corresponding to short- and long-term incubation effects. Additional validation using synthetic epidemic data from an extended SEIR model with short and long incubation compartments showed that the attention-based analysis identified long-incubation components and reproduced shifts in temporal influence patterns under different incubation parameters. The models also showed reasonable predictive performance.
Implications
The proposed LSTM–attention framework provides a data-driven approach for linking surveillance time series with latent epidemiological structures such as incubation dynamics and may complement mechanistic epidemic models in mathematical epidemiology.
