Deep learning forecasting of scorpion envenoming incidence in Brazil to support early warning and prevention

Abstract Background Scorpionism is a growing public health problem in Latin America, with incidence rapidly increasing in Brazil. Climate change and land-use transformations are expected to intensify this trend in the coming decades. Accurately anticipating the spatiotemporal dynamics of the disease is essential to optimize resource allocation, antivenom stock management, and preventive strategies. In this study, we propose and implement global deep learning models for short-term incidence forecasting in each Brazilian state. Methods We developed local and global temporal forecasting models for Brazil’s five macro-regions using monthly incidence data from all 27 federal units. Four approaches were evaluated: a 12-month Naive model, ARIMA, Prophet, and a global deep learning model (N-BEATS), integrating the historical series of all states within each macro-region. Data from 2007 to 2024 were used for training and backtesting with Bayesian hyperparameter optimization, and the best-performing models were applied to unseen 2025 data. Results A total of 2,012,495 cases were reported between 2007 and 2024. Here we show that scorpionism incidence increased nationwide, particularly in the Central-West ( ≈ 15.6× over a decade). Global N-BEATS models achieved the best performance (median symmetric Mean Absolute Percentage Error, sMAPE: 13%, min = 3.08, max = 30.5), with the lowest errors in Alagoas and Pará. On the 2025 test set, the models accurately captured short-term variation (median sMAPE: 16.8%, min = 5.2, max = 39.9). Conclusions These results demonstrate 6-month-ahead predictions across Brazilian regions and states, supporting prevention strategies, antivenom allocation, and early-warning systems under ongoing environmental change.