Lagged Temperature Fluctuation Effects on Suicide Attempts: A Comparative Study Using DLNM and LSTM Networks

Suicide remains a critical global health concern, influenced by interactions between social, psychological, and environmental factors. This study aims to assess and forecast short-term lag effects of environmental variables on suicide attempts. We employ statistical modeling distributed lag non-linear model (DLNM) to identify lag-response associations and deep learning techniques (LSTM) to build a predictive model, facilitating epidemiological understanding and forecasting capability. Model performance was evaluated across lag structures (1–7 days) using multiple error metrics (RMSE, MAE, MSE, SMAPE), with subgroup analyses by sex and age. The LSTM model achieved optimal performance at a 3-day lag (RMSE: 0.1789), outperforming in prediction error metrics across demographic subgroups. While LSTM offered greater accuracy and flexibility in capturing intricate multi-variable interactions for forecasting, DLNM offered epidemiological interpretability of lag-response relationships. The short-term climatic exposures, particularly temperature fluctuations and humidity, significantly influence suicide attempts. The integration of DLNM and LSTM modeling provides complementary insights offering both interpretability and high predictive power.