- 學位論文
比較國際疾病分類標準碼與主訴在台灣醫院急診症候群偵測類流感與腸胃道流行的偵測效益及其加入每日氣象因子之預測
Comparison between Chief Complaints and ICD-9 in Detecting Influenza-like Illness and Gastro-enteritis Outbreaks through Hospital Emergency Department-based Syndromic Surveillance in Taiwan and Their Prediction Models after Integrating with Meteorological Factors
摘要
傳統的傳染病偵測系統面對生物恐怖攻擊事件及新興傳染病流行時，其時效性明顯不足；因此建立症候群偵測系統偵測「非特異性的症候群」，以及早發現流行、推動公共衛生措施，避免流行擴散。本研究以台北市五家醫院急診病患之電 子資訊，建立「標準化主訴勾選」的症候群偵測系統，期能有最符合台灣現況的急診症候群偵測系統。本研究目的為：(1)建立主訴及國際疾病分類標準碼第九版(International classification disease ninth code, ICD-9 碼)的最佳類流感與急性腸胃道感染症候群，並比較其優缺點與適用性，(2)探討天氣因子及其一週內幾日後與以主訴及ICD-9 碼兩症候群定義而偵測得之類流感及急性腸胃道感病患「異常警訊」的相關性，(3)建立類流感及急性腸胃道感染之時間序列預測模式，同時考慮年齡、性別及天氣因子，以預測醫院此二疾病趨勢，及(4)比較不同統計方法、主訴及ICD-9碼兩篩選標準及考慮天氣對急診類流感及急性腸胃道感染症候群偵測系統的特異度、時效性及敏感度之影響。 做法上，由專科醫師回顧台北市某醫院2004年一周共1,281 筆的急診病患資料進行症候群分類，建立症候群的黃金標準；並收集台北市五家醫院2005年10月1日至2007年2月14日共126,675 位急診病人之主訴症狀、ICD-9 碼、體溫及流行病學相關資料，進行症候群偵測數據分析，以建立波以松時間序列分析預測模式(Poisson regression for counts time series)，且控制重要因子(性別、年齡、醫院別、周六、周日及假日效應、日溫差、每日最低溫、相對濕度及季節)，進行流感及急性腸胃炎個案數之預測。又據此模式的預測值與觀察值之差值，以累積合統計模式(cumulative sum, CUSUM)及指數加權平滑模式(exponential weighted movingaverage, EWMA)等統計法進行警示分析，並比較不同統計方法之偵測效益。 比較主訴及ICD-9 碼定義的類流感及急性腸胃炎病患「異常警訊」之偵測效益，類流感以主訴的敏感度較佳(主訴及ICD-9: 0.86 vs 0.82)，而ICD-9 碼的特異度較佳(0.87 vs 0.94)，且其與黃金標準的一致性也較佳(0.757 vs 0.678)；在急性腸胃道感染症候群偵測之敏感度(0.7 vs 0.83)、特異度(0.97 vs 0.99)及一致性(0.68 vs0.82)均以ICD-9 碼較佳。另依據主訴及ICD-9 分別選出64,970、54,598 位類流感病患及18,061、26,027 位急性腸胃炎病患，發現週日及假日效應在此兩類症候群預測模式均有顯著影響，農曆新年又比其他國定假日的影響為高[以不考慮天氣的ICD-9 類流感模式兩者之危險比(risk ratio, RR)分別為1.139, 1.212]。 在天氣因子上，3 天前日溫差極小(小於當月平均之兩倍標準差)及5 天前相對濕度極大(大於當月平均之兩倍標準差)，會偵測到「ICD-9 定義的類流感」病例數之增加(各為RR=1.134, RR=1.146)；而1 或6 天前的日溫差極小反減少類流感數 (RR=0.854, 0.856)，但「主訴定義的類流感」卻看不到與天氣的顯著相關。「ICD-9定義的急性腸胃感染」病例數會因1 天前的最低溫極高或2 天前的最低溫極低而增加(RR=1.24, RR=1.127)；但卻因7 天前的日溫差極小(RR=0.754)、4-5 天前的相對濕度極低(RR=0.545, 0.654)及當天的相對溼度極大(RR=0.743)而減少。「主訴定義的急性腸胃道感染」病例數受到4 天前最低溫極大及5 天前颱風來襲而增加(RR=1.301, RR=1.282)；但會因當天相對濕度極大(RR=0.742)或極小RR=0.638)而減少。未來若以天氣預測傳染病流行時，須考慮天氣干擾就醫行為。 不同統計方法之偵測效益，在類流感及腸胃炎模式均以指數0.9 加權平滑模式[EWMA(λ=0.9)]之特異度較佳，而以EWMA(λ=0.6)及C3 之敏感度較佳。整體而言，以ICD-9 碼篩選的急性腸胃道感染病例與黃金標準之ㄧ致性較高，且偵測警訊之敏感度及時效性亦較佳；主訴定義的類流感在病例篩選時雖有較佳的敏感度，但與黃金標準的一致性及偵測敏感度卻較ICD-9 差。 未來建立症候群偵測預測模式，應先年齡、性別分層，再加入天氣因子，並增加檢體採集作為佐證症候群偵測系統警訊之依據，明瞭不同病毒或不同流感病毒型別/亞型之異同，研展更完善快速的回饋系統，做好警訊發生時的準備計劃，期能於發現流行的第一時間即推動公共衛生防疫，降低病例數，減少社會損失。
並列摘要
Faced several major outbreaks of emerging/reemerging infectious diseases (EID), the traditional surveillance systems are less timeliness to detect outbreak than syndromic surveillance system. Syndromic surveillance system monitor “non-specfic clinical symptoms” to detect outbreak earlier, and public health intervention can be started earlier to stop the epidemic. This study used daily ED patients from five hospitals in Taipei to set up syndromic surveillance system with predefined chief complaint (CC) list. The aims of this study are: (1) to set up a best syndrome groups of influenza-like-illness (ILI) and acute gastroenteritis (AGE) by CC and International classification disease ninth code (ICD-9), and (2) to investigating the relationship of climate factors and their lag effect versus ILI and AGE, (3) using age, gender and climate factors to establish time series predict models of ILI and AGE, (4) sensitivity, specfifcity, and timeliness comparison of models by different statistical methods, CC and ICD-9, and with or without climate. First, a specialist in Emergency medical review the 1,281 charts during June 24-July 8, 2004 to decide syndrome group of each patient, and using it as the golden standard. This study collected CC, ICD-9 code, body temperature, and epidemiological data of 126,675 emergency department patients in five hospitals in Taipei during Oct. 1, 2005-Feb. 14, 2007 to set up Poisson regression for counts time series model and use cumulative sum (CUSUM) and exponential weighted moving average (EWMA) for alerting. The Poisson regression for counts time series model controlled age, gender, hospitals, Saturday, Sunday, holiday, daily temperature difference, daily lowest temperature, relative humidity and seasonal factors to predict ILI and AGE cases. Using the value difference between observation and prediction establish alert system and compare difference statistical methods. Comparison of the sensitivity and specificity of ILI and AGE cases defined by CC and ICD-9, ILI defined by CC has higher sensitivity than ICD-9 (CC vs ICD-9: 0.86 vs 0.82), but ILI by ICD-9 has higher specificity and kappa value than CC (0.87 vs 0.94and 0.678 vs 0.757). AGE by ICD-9 has higher sensitivity (0.7 vs 0.83), specificity (0.97 vs 0.99) and kappa value (0.68 vs 0.82) than CC. ILI by CC and ICD-9 are 64970 and 54598 cases respectively, and AGE by CC and ICD-9 are 18061 and 26027 cases during Oct. 1, 2005-Feb. 14, 2007. Those syndrome groups have significant weekend and holiday effects, and the cases on Chinese lunar New Year are more than other national holidays. ILI by ICD-9 are highly related to extreme daily temperature difference of lag 1, 3, 6 days and extreme relative humidity of lag 5 days, but ILI by CC are not significantly related to metrological factors. AGE cases by ICD-9 are highly related to excessive daily temperature difference of lag 7 days, daily lowest temperature of lag 1 and 2 days and relative humidity of lag 0, 4, 5 days. AGE cases by CC are significantly related to lowest temperature of lag 4 days, relative humidity and typhoon of lag 5 days. When we explained the relationships between metrological factors and diseases, we should consider the other factors that related to diseases and metrological factors. When analyzing the different statistical methods, EWMA(λ=0.9) has higher specificity and EWMA(λ=0.6) and C3 have loftier sensitivity. Overall, cases by ICD-9 have higher kappa value with golden standard and can detect alert more sensitivity and more timeliness. According to this study, we can set up a prediction model stratified by the age and gender and investigate the relationship of metrological factors and diseases in the future. Increasing the sample collection and establish the feedback systems are very important to determine whether the alerts are true or not and well done the public health actions to stop outbreaks. If we can discover the signals as early as possible, take epidemic prevention and control immediately. It will be efficiency to lower down the infection cases and social cost.