念珠菌菌血症預測模型之建構：病例對照研究

Abstract

念珠菌是一種伺機性病原菌，是造成醫療照護相關感染的重要病原菌，在免疫系統受損的宿主會引發危及生命的侵襲性感染，而且會造成高致死率、高併發症 及額外的醫療費用和住院天數。現行診斷工具有其限制，以菌血症為例，血液培養為念珠菌菌血症的唯一診斷方法，然而敏感度(21-71%)不理想，而且檢驗時程有其限制。然而醫療照護相關感染中，侵襲性念珠菌感染與多重抗藥性細菌引起的感染臨床表現不易區分，且危險因子大多相同。對侵襲性念珠菌感染的推測性治療或經驗治療的決定是有困難的，其治療策略難以獲得驗證。加上高危險病人反覆感染，藥物選擇決策困難，然而延遲治療可能影響病人預後。此外，目前採取經驗性治療策略常常導致過度使用藥物。念珠菌菌血症等侵襲性念珠菌感染及早適當的藥物治療很重要，而即時且正確診斷是醫療尚待解決之挑戰，牽涉病人安全以及醫療經濟。醫療資訊利用巨量資料探勘以發掘深層價值是全世界強調的新趨勢，也是國家發展重點。建立預測模型客觀的系統性風險評估，可以主動反覆評估，而不會增加資源消耗。此研究針對念珠菌菌血症建立預測模型，作為建構念珠菌菌血症醫療決策輔助系統的基礎，以達到血液培養確診之前評估念珠菌菌血症之風險，其預測結果可讓醫師在決定治療藥物選擇時作參考。 本研究以台大醫院2016年01月01日至2017年12月31日病人資料，針對初次發生菌血症且單菌菌血症的病人，比較個案(念珠菌菌血症)與對照病人(細菌菌血症)，以多變數分析之危險因子建立預測模型並進行內部驗證。病人資料包含基本資料、住院資料、手術等處置資料、血液培養等檢驗資料、抗生素等藥物使用資料等，使用五種資料探勘算法建立模型，包括支援向量機、決策樹、隨機森林、k最近鄰居法和單純貝氏分類器，同時以廣義線性模型作為比較基準。由於念珠菌菌血症病人數遠少於細菌菌血症，我們使用兩種抽樣技術，包括隨機下採樣和隨機上採樣，去處理病例對照研究的不平衡資料。隨機下採樣樣本建立的單純貝氏分類器，敏感度和平衡準確率最高而特異性和準確率在可接受範圍內，是最佳的預測模型。此研究主要貢獻為應用資料探勘算法處理複雜的醫療資料，針對菌血症病人預測念珠菌菌血症的初步結果。

Candida species are among top four pathogens causing healthcare-associated infections in critically ill patients which are associated with considerable morbidity and mortality, additional length of hospital stay and extra costs. Culture is currently the only diagnostic method. The sensitivity of blood cultures is only 21 to 71%. Blood cultures are further limited by slow turn-around times. Nosocomial colonization and infection with multiple drug resistant organism and Candida species share many common clinical signs and risk factors. Presumptive or empirical treatment of patients with invasive candidiasis is difficult, as such strategies have not been validated. Although delayed initiation of antifungal therapy is associated with increased mortality, overuse of antifungals may lead to the emergence of drug-resistant Candida species. Therefore, prediction rules have been developed and validated in order to identify those patients at high risk for invasive candidiasis and likely to benefit from early treatment. This case-control study analyzed comprehensive data of hospitalized patients with candidemia (case) or bacteremia (control) to identify risk factors independently associated with candidemia and implemented data mining techniques to develop and internal validate a candidemia prediction model for patients with first episode of bloodstream infections. Among complicated patient data, 5 variables were used as input of the algorithm. We used five data mining algorithms including support vector machine (SVM), decision tree (DT), random forest, k-Nearest Neighbors (k-NN) and Naive Bayes while a generalized linear model (GLM) was set as the baseline. We used two sampling techniques, i.e., random undersampling (RUS) and random oversampling (ROS), to handle the imbalanced datasets. The most parsimonious model was Naïve Bayes with RUS sample with sensitivity (71.9%), specificity (65.1%), positive predictive value (8.9%), negative predictive value (98.0%), accuracy (65.4%), balanced accuracy (68.5%) and F1 score (15.8%), and achieves an area under the receiver operating characteristic curve of 0.817. In conclusion, this data mining-based algorithms using only 5 variables generated in routine practice from heterogeneous patient population to discriminate patients with a high risk of candidemia among hospitalized patients with positive blood cultures, illustrating the potential value and feasibility of personalizing candidemia treatment.