氣溫對於急診類流感就診人次之影響－時間序列分析

Abstract

背景及目的： 2016年1月24日，帝王寒流來襲，台北溫度降到4度。當月底，流感疫情即開始升溫，在二月春節假期爆發達到高峰。流感爆發加上春節連假，各大醫院發生了急診室壅塞不堪負荷的情況。為改善對未來流感疫情的預測能力，我們發展時間序列統計模式，探討氣溫、放假天數、及交通流量是否為急診類流感就診率的顯著預測因子。 方法： 本研究資料來源為疾病管制署傳染病統計資料、中央氣象局氣溫紀錄、國道高速公路局交通流量統計資料、以及歷年行事曆。利用傳染病統計資料查詢系統資料庫取得2007年第1周至2016年第52周之每周急診類流感就診率 (每周類流感就醫人次除以每周急診就醫總人次)，連結氣象資料、各年行事曆、與交通流量統計，進行時間序列分析。時間序列統計模型1的因變項為台北區急診類流感就診率，自變項為每週最低溫及每週國定假日天數。時間序列統計模型2的因變項為全國急診類流感就診率，自變項為全國每周百萬車公里。 結果： 台北區急診類流感就診率之ARMA(1,1)模型 (模型1a)，移動平均參數估計為0.11688 (p-value=0.0169)，自迴歸係數為0.91729 (p-value<.0001)。加入輸入變量低溫和放假天數的台北區急診類流感就診率之ARMA(1,1)模型 (模型1b)，移動平均參數估計為 -0.13852 (p-value=0.0051)，自迴歸係數為0.90485 (p-value<.0001)，低溫係數為-0.04536 (p-value=0.0387，延遲1)，放假天數係數為0.52171 (p-value<.0001, 延遲0)。模型1b較模型1a好，模型1a D^2=25.72，模型1b D^2=18.85，顯然模型1b和實際值的差異較小。 另外，全國百萬車公里的每周最大值為全國每周急診類流感就診率的輸入變量的ARMA(2,2)模型2，移動平均參數估計MA1,1為 0.73192 (p-value <.0001)，移動平均參數估計MA1,2為 -0.19545 (p-value = 0.0011)，自迴歸係數AR1,1為1.66721 (p-value<.0001)，自迴歸係數AR1,2為-0.70858 (p-value<.0001)，全國百萬車公里係數為0.07306 (p-value <.0001，延遲0)。 結論： 本研究發現前一週最低氣溫、當週放假天數、及當週交通流量為急診類流感就診率的顯著預測因子。將這些資訊納入統計預測模型，可以改善未來對流感疫情的預測能力，以協助各大醫院急診或門診流感藥物配置和醫護人力調度。

Background and purpose: At the end of January 2016, influenza outbreaks in Taiwan, reaching a peak in February during the Chinese Festival, and so led to the emergency room at hospitals have been completely packed. On 24 January 2016, the temperature was dropping to 4-Celsius degree in Taipei. We use the line chart to find out the correlation between the outbreak of flu, index of weather traffic and the numbers of holidays. Using time series to build the model and predict the future of the proportion of influenza in the emergency department within eight weeks. The prediction could help each hospital emergency room or outpatient clinic in demand of drugs and human resource. Methods: The source of the study is the Taiwan National Infectious Disease Statistics System data of Center for Disease Control, observation data query system of Central Weather Bureau (CWB), Taiwan Area National Freeway Bureau and year calendar from 2007 to 2016. Database obtained from the CDC's Taiwan National Infectious Disease Statistics System data in the first week of 2007 to the fifty-two week of 2016 Taipei City weekly proportion of Influenza-like illness (ILI) and add the data from CWB and calendar days to build the time series model. Using the developed model to predict the proportion of influenza-like illness in Taipei City from the first week of 2017 to the eighth week of 2017 and compare it with the actual value. The proportion is the number of ILI divided the total number of people visited the emergency department. Using time series to analyze the data from Taiwan Area National Freeway Bureau and weekly proportion of Influenza-like illness in Taiwan to prove that the critical mass may cause the outbreak of flu. Compare the prediction of the model which doesn't add input variables (temperature and holidays) with the model add input variables (temperature and holidays). Results The ARMA (1,1) model of the emergency diagnosis rate of the emergency in Taipei district (model 1), the moving average parameter is estimated to be 0.11688 (p-value= 0.0169), and the autoregression parameter is estimated to be 0.91729 (p-value <.0001). The ARMA (1,1) model of the emergency diagnosis rate of the emergency in Taipei district added the input variables temperature and holidays (model 2), the moving average parameter is estimated to be -0.13852 (p -value = 0.0051), the autoregressive coefficient parameter is estimated to be 0.90485 (p-value <.0001), coefficient of low temperature is -0.04536 (p-value = 0.0387, delay 1) and coefficient of holiday is 0.52171 (p-value <.0001, 0). The model 2 is better than them model 1, D^2 = 25.72 for model 1 and D^2= 18.85 for model 2. Obviously, the difference between model 2 and actual value is small.In addition, the ARMA (2,2) model of emergency diagnosis rate of the emergency in Taiwan added input variable the max of million vehicles kilometer, the moving average parameter MA1,1 is estimated to be 0.73192 (p -value<.0001), the moving average parameter MA1,2 is estimated to be -0.19545 (p -value=0.0011), the autoregression parameter AR1,1 is estimated to be 1.66721 (p-value <.0001), the autoregression parameter AR1,2 is estimated to be -0.70858 (p-value <.0001) and coefficient of the max of million vehicles kilometer is - 0.07306 (p-value <.0001, delay 0). Conclusion: This study found that the best model was the regression of ARMA(1,1) for the lag a week of the Taipei City weekly proportion of Influenza-like illness at the lowest temperature of the week.