Objective To evaluate the performance of the Prophet model in predicting the daily incidence of hand, foot, and mouth disease (HFMD) in Shenzhen city, to analyze the impact of the COVID-19 pandemic, public holidays, and school vacations (summer/winter) on HFMD predictions, and to provide new insights for HFMD surveillance and early warning systems.

Methods Using daily incidence rate data of hand, foot, and mouth disease (HFMD) in Shenzhen city from 2011 to 2023 as a training set, we constructed different Prophet models based on two factors: (1) whether data from the COVID-19 epidemic period were included, and (2) whether holiday effects were adjusted. These models were then used to predict the daily incidence rate from January to July 2024, and the predictions were compared with the actual observations. Model performance was evaluated using four metrics: mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), and symmetric mean absolute percentage error (SMAPE).

Results Compared to the Prophet model trained on data during the COVID-19 pandemic, the baseline Prophet model (without pandemic data) achieved reductions in daily average MSE, MAE, RMSE, and SMAPE of 45.64%, 27.63%, 25.89%, and 7.97%, respectively. After controlling for holiday effects, the MAE, RMSE, and SMAPE values remained unchanged. When the prediction horizon exceeded 4 months, the daily MSE decreased by 3.81%. For predictions within 4 months, the model yielded MSE, MAE, RMSE, and SMAPE values (with 95% confidence intervals) of 0.42 (0.33, 0.90), 0.51 (0.45, 0.75), 0.65 (0.57, 0.95), and 0.42 (0.24, 0.44), respectively. In contrast, the predictions at 7 months showed higher errors: MSE 1.78 (0.39, 29.61), MAE 0.94 (0.48, 3.05), RMSE 1.33 (0.63, 5.44), and SMAPE 0.40 (0.33, 0.43).

Conclusions During the COVID-19 pandemic (2020-2022), the daily HFMD incidence rate significantly affected the performance of the Prophet model. While controlling for holidays and school vacations (winter/summer) had a limited effect on short-term predictions, it improved the model′s accuracy for long-term predictions.