Comparison of the prediction effects of LSTM, SARIMA and SARIMAX models on the incidence of hand, foot, and mouth disease

Objective: To compare the effects of seasonal autoregressive integrated moving average (SARIMA) , seasonal autoregressive integrated moving average with exogenous regressors (SARIMAX) and long short-term memory neural network (LSTM) models in predicting the incidence of hand, foot, and mouth disease (HFMD). Methods: Monthly incidence data of HFMD in Kunming City from 2010 to 2019 were collected. SARIMA, SARIMAX and LSTM models were established using the monthly incidence of HFMD from 2010 to 2018 to predict the monthly incidence of HFMD from January to December 2019. The prediction performance of the three models was compared using mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE) and mean absolute percentage error (MAPE). The optimal prediction model was selected based on the principle of minimizing MSE, RMSE, MAE and MAPE. Results: The HFMD cases were reported every month in Kunming City from 2010 to 2019, with the incidence fluctuating between 188.27/105 and 363.15/105. The disease exhibited a biennial high-incidence bimodal distribution. Among the four evaluation indicators for the training and testing sets, the LSTM model had the smaller values: MSE was 63.182 and 102.745, RMSE was 7.949 and 10.136, MAE was 6.535 and 7.620, and MAPE was 46.726% and 31.138%. The LSTM model performed the better, followed by the SARIMA model, while the SARIMAX model had the relatively poorest performance. Conclusion: The LSTM model outperforms the SARIMA and SARIMAX models in predicting the incidence of HFMD.