Objective To investigate the phase-specific impacts of non-pharmaceutical interventions (NPIs) during the coronavirus disease 2019 (COVID-19) pandemic on the transmission of hand, foot and mouth disease (HFMD) and evaluate the applicability and dynamic responsiveness of different prediction models in an intervention context.
Methods Monthly HFMD incidence data from 2014 to 2023 in Fengyang county, Anhui province, China were analyzed. Theil–Sen robust regression was employed to estimate the annual percent change (APC) and the average annual percent change (AAPC) before and after COVID-19. Baseline scheme (counterfactual stress test): the model was trained on the data from January 2014 to December 2019 and validated on the data from January 2020 to December 2023, without incorporating any pandemic intervention variables. Intervention scheme (adaptive prediction): the model was trained on the data from January 2014 to November 2022, incorporating phase-specific intervention dummy variables, and validated on the data from December 2022 to December 2023. The prediction performance of seasonal autoregressive integrated moving average model with exogenous regressors (SARIMAX), Prophet (a time-series decomposable additive model), and long short-term memory (LSTM) was compared based on mean absolute percentage error (MAPE), root mean square error (RMSE), and mean absolute error (MAE). Two novel indices, the phase sensitivity index (PSI) and the resilience coefficient (REC), were proposed to assess model adaptability in terms of dynamic responsiveness and resistance to perturbations.
Results In the pre-pandemic period from 2014 to 2019, HFMD incidence declined (APC = −15.97%, 95%CI: −22.06% to −6.80%, P = 0.004). In the post-pandemic period from 2020 to 2023, it increased (APC = 36.17%, 95%CI: 8.84% to 54.72%, P = 0.002). Over the full period from 2014 to 2023, the overall trend still decreased (AAPC = −12.28%, 95%CI: −16.49% to −7.37%, P < 0.001). During the relaxation period from December 2022 to December 2023, the intervention scheme outperformed the baseline scheme in out-of-sample prediction: the MAPE of SARIMAX decreased from 73.77% to 44.54% (a reduction of 29.23 percentage points); the MAPE of LSTM decreased from 46.18% to 41.85%; whereas the MAPE of Prophet remained high at 69.73%. The PSI/REC indicated that the SARIMAX and Prophet intervention models were more sensitive to the transition from routine control to relaxation, with PSI greater than 1 and recovery within 1 month (REC equal to 1); LSTM showed a PSI close to 1 but slower recovery (REC equal to 0.33) in about 3 months.
Conclusions Non-pharmaceutical interventions exert phase-specific effects on HFMD transmission with a risk of rebound. Statistical models and deep learning models are complementary for surveillance under intervention scenarios. Statistical models are sensitive to exogenous shocks and recover quickly, whereas deep learning models excel in short-term fitting but require enhanced generalization and adaptive updating after relaxation of control measures. Integrating multiple models is recommended to optimize infectious disease early warning systems under complex dynamics in the presence of interventions.
DownLoad: