JMIR Publications

ABSTRACT

Background:

Background:

Enterovirus infections, particularly hand, foot, and mouth disease (HFMD), pose a significant public health burden in Taiwan, with seasonal hospitalization peaks in spring and early summer. Accurate forecasting of hospital burden is essential for effective resource allocation and public health preparedness, achieving 6.8% mean absolute percentage error for 1-week ahead forecasts in this study.

Objective:

Objective:

To develop and evaluate machine learning (ML) models using age-specific surveillance data to predict enterovirus-related hospitalizations 1-4 weeks in advance, informing public health decision-making.

Methods:

Methods:

We analyzed weekly surveillance data (2008-2024) from Taiwan's Centers for Disease Control, including outpatient, emergency, and hospitalization counts across five age groups. Support Vector Machine (SVM), XGBoost, and Random Forest (RF) models were trained on 85% of data (2008-2021) and tested on 15% (2022-2024). Performance was assessed using R-squared (R²), root mean square error (RMSE), and mean absolute percentage error (MAPE) for 1-, 2-, and 4-week-ahead predictions. Age-specific features and risk ratios were evaluated.

Results:

Results:

Random Forest outperformed SVM and XGBoost, achieving R²=0.951, RMSE=12.8, and MAPE=6.8% for 1-week-ahead predictions with age-specific features. Performance declined for longer horizons (2-week: R²=0.893, MAPE=11.4%; 4-week: R²=0.742, MAPE=22.8%). Outpatient visits in children aged 0-2 and 5-9 were the strongest predictors. Children aged 0-2 had 37.1× higher hospitalization odds (95% CI: 35.2-39.1) than adults (15+).

Conclusions:

Conclusions:

ML models using age-specific surveillance data enable accurate 1-2-week-ahead hospitalization forecasting for enterovirus infections. This approach supports proactive resource allocation and can be adapted for other infectious diseases, advancing digital epidemiology.