Inferring unobserved vector dynamics for dengue forecasting using physics-informed neural networks and mechanistic transmission models

Abstract

Accurately characterising mosquito infection dynamics is essential for effective dengue prevention and control, yet these dynamics are rarely observable through routine surveillance. Here, we integrate a reduced SI-SIR transmission model with monthly dengue incidence data using physics-informed neural networks (PINNs) to develop a Dengue-Informed Neural Network (DINN). The DINN simultaneously fits reported dengue cases from 15 countries between January 2014 to April 2025 and reconstructs the unobserved time series of infected mosquitoes. Using the inferred mosquito infection dynamics together with five-month-lagged climatic variables, we subsequently train recurrent neural networks (RNN, GRU and LSTM) to forecast mosquito infection levels one month ahead. For each country, the optimal forecasting model is selected based on out-of-sample performance, and model behaviour is interpreted using SHAP analysis. We show that the DINN captures multi-wave outbreak dynamics across diverse epidemic regions and enables the development of country-specific vector forecasting models. When integrated with the transmission model, the predicted mosquito trajectories support two-year projections of dengue incidence, providing a quantitative framework for early warning and evidence-based control strategies. Author summaryMechanistic models of dengue transmission often require assumptions about mosquito population sizes and temperature-dependent biological traits that are difficult to specify accurately and rarely observed in practice. These limitations can reduce the reliability of outbreak prediction and hinder the evaluation of control strategies. Here, we introduce a Dengue-Informed Neural Network (DINN) that integrates deep learning with a reduced SI-SIR transmission model to address these challenges. Without requiring prior specification of mosquito initial conditions or biological trait functions, DINN infers latent time series of infectious mosquitoes directly from reported human case data and reconstructs complex, multi-wave transmission dynamics. Applying this approach to dengue surveillance data from 15 countries, we uncover substantial regional variation in inferred mosquito infection patterns, reflecting differences in environmental variability and vector control. By combining the inferred mosquito dynamics with climatic information, we further develop interpretable forecasting models and generate two-year projections of dengue incidence. Our results demonstrate that DINN provides a flexible and interpretable framework for inferring unobserved vector dynamics and offers practical support for early warning and evidence-based control of mosquito-borne diseases.