Abstract / Summary
Hand, foot, and mouth disease (HFMD) is a common pediatric infection in Korea, tracked weekly through Korea Disease Control and Prevention Agency sentinel surveillance, but no domestic study has compared deep learning models with classical baselines using multi-seed evaluation and formal accuracy testing across the COVID-19 disruption.Five deep learning architectures (deep neural network, one-dimensional convolutional neural network, single-and multilayer long short-term memory [LSTM], and a convolutional neural network-LSTM hybrid) were compared with three classical baselines (naive, seasonal naive, and seasonal autoregressive integrated moving average [SARIMA]) for one-week-ahead forecasting of the Korean HFMD sentinel rate, using weekly data from 2011 to 2019 as the primary analysis and through 2025 as a robustness analysis.SARIMA achieved a test mean absolute error of 1.73, four to six times lower than the best deep learning model.The Diebold-Mariano test confirmed this advantage at P<0.001 under both loss functions, with the ranking robust to input window length and the COVID-19 disruption.Single-seed evaluation produced a misleading deep learning ranking, supporting multi-seed reporting in deep learning forecasting research.These findings support a deployable SARIMA-based alert pipeline keyed to the 20-per-1,000 outpatient threshold for childcare facility advisories, sentinel-clinic staffing, and hand hygiene campaigns.