Real-time dynamic prediction of HFMD transmission using SEIRQ-ARIMA hybrid model optimized by multi-stage ABC-GWO algorithm

In Hand, Foot, and Mouth Disease (HFMD) control, conventional SEIR/SIR models rely on static parameters, which limits their adaptability to dynamic real‑world conditions. To address this, we propose a SEIRQ–ARIMA hybrid model that integrates a quarantine‑enhanced SEIR framework (SEIRQ) with an ARIMA time‑series model, whose dynamic parameters are optimized via a multi‑stage Artificial Bee Colony–Grey Wolf Optimization (ABC–GWO) algorithm. This study contributes three key novelties to the field: (i) an IoT-driven dynamic SEIRQ parameterization with explicit intervention tracking via time-varying isolation rate δ(t); (ii) a multi-stage ABC–GWO calibration strategy that improves stability and mitigates premature convergence in nonlinear parameter estimation; and (iii) an interpretable SEIRQ–ARIMA fusion that combines mechanistic dynamics with statistical residual learning to support both accurate forecasting and policy-oriented evaluation. The core innovation is that, unlike traditional SEIR models that ignore quarantine interventions, our SEIRQ framework is designed to dynamically calibrate the isolation rate δ using real-time Internet-of-Things (IoT) surveillance streams. In this study, we validate the framework using historical surveillance records from Guangxi, China (2014–2020). the model achieves substantial forecasting improvements,reducing RMSE by 94.6% compared to the standalone SEIRQ model and reducing MAE by 94.1% compared to ARIMA.In practical application, this study provides the quantitative estimation of HFMD quarantine effects: at the optimal isolation rate (δ = 0.413), the infection peak is reduced by 52.7%, with an overall peak reduction of 40–65% and a cost–benefit ratio of 1:8.7. Sensitivity analysis identifies δ ∈ [0.3, 0.5] as a critical isolation range. Remaining limitations include simplified economic assumptions; future work will incorporate more detailed cost‑effectiveness modeling and Transformer‑based prediction modules.