臺灣類流感週資料分析與預測

Abstract

摘要 本研究使用衛生福利部疾病管制署資料開放平台所提供的類流感監測資料，預測2016年1月至2025年4月間類流感於「住院」、「門診」與「急診」三類就醫途徑中的每週病例人數。預測模型涵蓋2016年1月至2024年12月（共469週）之樣本內資料，並針對2025年前18週進行樣本外預測。所採用的方法包括：具外生變數的自我迴歸整合移動平均模式（ARIMAX）、霍爾特線性趨勢指數平滑法，以及針對重大事件進行的介入分析。考量春節對就醫行為的影響，模型中特別納入春節指標變數。針對急診資料，也額外考慮COVID-19 Omicron變異株首次社區流行期間的影響。針對2016年霸王級寒流與Omicron廣泛傳播等時點亦進行介入分析。本研究採用擴展式預測策略（Expanding approach），逐週更新模型並進行滾動式一步預測。預測表現評估採用四項常見指標：平均絕對誤差（MAE）、均方根誤差（RMSE）、平均絕對百分比誤差（MAPE）與平均絕對比例誤差（MASE）。結果顯示，在「住院」類別中，指數平滑法線性趨勢具有最佳表現；而在「門診」與「急診」資料中，ARIMAX模式提供最準確的結果。整體而言，春節對類流感病例具有顯著影響，而採用逐週更新的預測策略可有效提升預測準確度。

Abstract This study utilizes influenza-like illness (ILI) surveillance data from the Taiwan Centers for Disease Control Open Data Portal to analyze weekly case counts for hospitalizations, outpatient visits, and emergency department encounters. Forecasting models were trained on an in-sample period from January 2016 to December 2024 (469 weeks) and evaluated using out-of-sample forecasts for the first 18 weeks of 2025. The three methods applied include autoregressive integrated moving average model with exogenous input variables (ARIMAX), Holt’s linear trend exponential smoothing method, and intervention analysis. Given the influence of the Lunar New Year on healthcare utilization patterns, a specific indicator variable was included to account for this temporal effect. The emergency department model incorporated an additional explanatory variable to capture the impact of the COVID-19 pandemic, as the visits were affected by both the seasonal fluctuation of the Lunar New Year and the shock of the pandemic. Furthermore, for the emergency department data, intervention analysis was specifically performed for the extreme cold wave in 2016 and the widespread Omicron transmission period. An expanding approach was employed, with the models recursively updated on a weekly basis to generate rolling one-step-ahead forecasts. The evaluation criteria include mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE), and mean absolute scaled error (MASE). The results indicate that Holt’s linear trend method achieved the best predictive performance for hospitalizations, while the ARIMAX model provided the most accurate forecasts for outpatient and emergency department visits. Overall, the findings underscore the pronounced effect of the Lunar New Year on ILI case volumes and demonstrate that weekly model updating can enhance predictive accuracy.