A Novel Business Intelligence Model for Monitoring Nosocomial Infection
院內感染 ； 視覺化 ； 支持向量機 ； 決策樹 ； healthcare-associated infection ； visualization ； support vector machines ； decision tree
|Volume or Term/Year and Month of Publication||
|Academic Degree Category||
背景：在一次 20mL生理食鹽水注射液污染 Ralstonia pickettii菌種，造成的院內感染案件中，我們發現類似的案例有可能因為臨床微生物醫檢師對報告資訊的不連貫，或是人員的工作輪替而遺漏了發現案件的機會，因此本研究以建立系統的方式，提供警示服務。 方法：系統收集2013年 9月至 2015年 3月的細菌培養結果，期間共有 260,779份報告，陽性結果佔66,446份，平均每日118 ± 30份報告。統計每日菌種培養數平均值及標準差，設立各菌種的警示門檻為平均值加上 1.28標準差，然後排程程式每日自動統計培養數量與判斷是否超出門檻。對於超出警示值的菌種即以email寄發通知簡訊給細菌室及管染管控人員。資料使用Google chart做視覺化的網頁呈現，包括菌種每日培養數量趨勢圖、檢體分佈圖及病房分佈圖等。本研究也利用此系統建置的資料中，以 SAS Enterprise Miner High-Performance Data Mining 的 Support Vector Machines, SVM及 Decision tree 模型測試 6種院內感染常見菌種，評估是否可以預測院內感染的發生。 結果：視覺化介面經問卷統計，評估使用者認為系統的實用性，在滿分 5分的標準中，得到平均 4.1分以上的成績。預測模型的測試結果在不同菌種間差異性很大，在 SVM模型中驗證組的靈敏度從20.4%~96.2%；Decision tree模型靈敏度則從25.0%~82.1%之間。 結論：藉由程式自動化收集、彙整資料與視覺化的圖表趨勢呈現，使得醫院醫療照護相關感染的作業更有效率。彙整後的資料，也有機會使用於院內感染的預測。
BACKGROUND: Recently, there was a healthcare-associated infection (HAI) event which induced by 20 ml injection normal saline contamination with Ralstonia pickettii in Taiwan. This event probably could not be discovered because laboratory staff were young or the staff were unfamiliar with the random bacterial appear trend. Therefore, we propose to develop a detection system that could efficiently summarize daily bacterial culture positive rate, and raise warning signals if there is any unusual high incidence. METHODS: Our system processes analytic data and proposes warnings in several steps. First, the baseline values and warning ranges were calculated by the average and standard deviation (S.D.), respectively, of daily bacteria culture positive counts from September 2013 to March 2015. The total, positive and daily report numbers were 260,779, 66,446 and 118±30, respectively. The warning threshold was set as average + 1.28 S.D. for each bacterium. The detection system was scheduled to run every day. It counts each bacterial report, aggregates a summary and determines which bacterial report numbers are over the threshold. Once the report numbers are higher than the threshold, the system will send an email to the laboratory and infection control staff. The monitoring user interface was designed with the Google chart web application. The designed interface presents visualization of the bacterial daily culture count trend, specimen distribution, and ward distribution. Two machine learning algorithms, support vector machines (SVM) and decision trees (DT), from SAS Enterprise Miner software were used to predict the occurrences of HAI. Six most commonly HAI related bacteria were discussed in the evaluation. RESULTS: The practicality of this visualization system was evaluated by an online questionnaire. The proposed infection detection system achieves an average score of 4.1 or more (in the score of 5 points). The sensitivity of the prediction models varies among different bacteria. In SVM model, the sensitivity ranges from 20.4% to 96.2%, and DT model ranges from 25.0% to 82.1%. CONCLUSION: This infection detection system provides automated data collection and summarization. It also provides data visualization as a chart view, which makes the management and prevention of the infection control more conveniently and efficiently.
醫藥衛生 > 醫藥總論
醫學科技學院 > 醫學資訊研究所