Prediction of hand, foot and mouth disease incidence trends in Shenzhen based on meteorological lag effect and LSTM model

Objective To predict the incidence trends of hand, foot and mouth disease (HFMD) in Shenzhen city, Guangdong province based on the meteorological lag effect and the long short-term memory (LSTM) model, thereby providing methodological support for establishing a more scientific and accurate meteorological disease risk early warning system.

Methods Data from 523 weeks (January 1, 2015 to December 31, 2024) of HFMD case reports in the China Disease Control and Prevention Information System were collected. A total of 353 046 HFMD cases in children aged ≤5 years were aggregated weekly. Meteorological parameters, including average weekly temperature (Temp), relative humidity (RH), temperature-humidity index (THI), apparent temperature (AT), and wind speed (WS), were integrated. Significant correlated variables were screened by Spearman rank correlation analysis to construct three types of LSTM models: a baseline model (case count only), single-factor models (case count + Temp/RH lagged 1–3 weeks), and composite models (case count + THI/AT lagged 1–3 weeks). Model performance was evaluated and compared via the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²).

Results Comparative analysis on the test set showed that all the models incorporating meteorological factors outperformed the baseline model (RMSE = 211.109, MAE = 117.654, R² = 0.942). Among the single-factor models, the model with Temp lagged by 2 weeks (Temp-lag2) performed best (RMSE = 179.915, MAE = 108.290, R² = 0.958), and the model with RH lagged by 2 weeks (RH-lag2) was optimal among all the models (RMSE = 177.457, MAE = 97.873, R² = 0.959). Among the composite models, the model with THI lagged by 2 weeks (THI-lag2) performed best (RMSE = 185.061, MAE = 103.817, R² = 0.956) but was slightly inferior to the RH-lag2 model. The RH-lag2 model predicted a bimodal pattern (peaks at weeks 10 and 15) for the HFMD incidence trends in the first 20 weeks of 2025, which closely matched the spring peak pattern of 2024 and showed a significant overall shift forward compared to the average peak period (weeks 9–35) from 2015 to 2024.

Conclusions The LSTM model constructed based on RH lagged by 2 weeks effectively characterizes the promoting effects of a hot and humid environment on the survival and transmission of enteroviruses, significantly improving the prediction accuracy of HFMD incidence trends.