Work

Forecasting pneumonia hospitalizations under climate change scenarios in Southern Brazil using neural networks

Year: 2026
Type: article
Abstract: Pneumonia hospitalizations exhibit seasonal patterns, and climate-informed models can help explore future hospital demand under specified climate trajectories. This study modeled monthly pneumonia hospitalizations in Rio Grande do Sul, Brazil, using an autoregressive neural network with exogenous climate variables (NNETAR). Hospitalization records from January 2015 to June 2025 were combined with temperature, precipitation, and humidity simulations from MPI-ESM1-2-LR under SSP5-8.5 to generate projections through December 2049. Model performance was assessed using a 24-month holdout and expanding-window validation against seasonal naive, Auto-ARIMA, and neural-network models without climate predictors. The NNETAR model reproduced the historical series closely during fitting (RMSE = 133.3; R 2 = 0.987). Climate-conditioned projections maintained a pronounced annual cycle, with monthly point estimates ranging from 2262 to 5623 and an upper 95% prediction bound of 5901 hospitalizations. The ordinary Sen slope was positive but nonsignificant (0.325 hospitalizations/month; p = 0.561), while trend direction varied across network initializations. The upper initialization-sensitivity case, based on the 97.5th percentile of the slope distribution, corresponded to approximately 12,400 additional hospitalizations relative to a no-trend trajectory over the projection horizon. Simpler benchmark models were more accurate in short-term temporal validation, whereas the climate NNETAR enabled long-term scenario analysis using projected meteorological inputs. Overall, the results indicate persistent seasonal pressure on hospital services under the evaluated climate pathway and support climate-conditioned scenario analysis as a complement to operational short-term forecasting of future hospital demand.
Language: English
Open Access status: hybrid
Estimated APC paid: $2,790
OpenAlex ID: W7236127345
FWCI:
Cites: 26
Cited by:
Related to: 10