Structural time series modelling for weekly forecasting of enterovirus outpatient, inpatient, and emergency department visits.

OBJECTIVES

Enteroviruses pose a substantial public health challenge in Taiwan, often leading to increased healthcare visits. This study utilizes Taiwan CDC databases to analyse weekly enterovirus case data from emergency departments (EDs), as well as outpatient and inpatient settings. The objectives are to understand infection patterns through model fitting, forecast future visits for proactive epidemic management, and improve forecast accuracy by incorporating holiday effects. This approach enhances the reliability of predictions, supporting timely and effective surveillance and early detection of significant case surges.

METHODS

This study divides the time series data into an in-sample period (2016-2023) and an out-of-sample period covering weeks 1 to 27 in 2024. Using an expanding window approach, the analysis applies Bayesian structural time series (BSTS) models, exponential smoothing, and random forest to forecast one-week-ahead cases over the 27 weeks in 2024. The study evaluates forecast accuracy using five key metrics and identifies significant surges in cases by detecting values that exceed the 95% prediction intervals, enhancing anomaly detection.

RESULTS

The results demonstrate that BSTS models, which incorporate trends, seasonal variations, summer, and Lunar New Year holiday effects, achieve superior forecasting accuracy. Specifically, by accounting for the Lunar New Year holiday within the out-of-sample period, the models attain mean absolute percentage error (MAPE) values of 6.509% for non-ED visits and 12.645% for ED visits.

CONCLUSIONS

The BSTS model effectively addresses nonlinearity and non-stationarity and adapts well to structural changes. This study highlights the importance of holiday adjustments, particularly for the Lunar New Year, in improving forecast accuracy during periods of unusual healthcare demand. These adjustments enhance the BSTS model performance for predicting irregular healthcare service demand.