From global to local: A transfer learning framework for municipal dengue prediction in Brazil

Dengue fever poses a growing global public health challenge, particularly in Brazil, where municipal-level forecasting is limited by high epidemiological variability and data scarcity. To address this problem, we propose a scalable deep transfer learning framework that leverages global dengue incidence data to improve short- and medium-term municipal forecasts. The approach employs CNN–LSTM architectures pre-trained on international data and fine-tuned using local information from the metropolitan regions of São Paulo, Rio de Janeiro, and Minas Gerais. Model performance is evaluated against classical statistical approaches, including Poisson, Negative Binomial, Integer-valued Conditional Heteroskedasticity, and SARIMA models, using MAPE, MAE, and MSE as complementary accuracy metrics. The results show that deep learning models substantially outperform traditional methods, with error reductions exceeding 75% in most settings. Moreover, transfer learning yields additional gains of up to 36% relative to models trained solely on local data, consistently across forecasting horizons of one to four weeks. By integrating global epidemiological patterns with local transmission dynamics, the proposed framework improves predictive robustness in data-limited municipal contexts and offers a practical tool to support early warning systems and public health decision-making.