대한신장학회

논문분류	춘계학술대회 초록집
제목	Prediction model for postoperative acute kidney injury after noncardiac major surgery using machine learning.
저자	Seokwoo Park* 1, Kwangsoo Kim2, Seongkeun Park3, Dong Ki Kim1, Kwon Wook Joo1, Yon Su Kim1, Hyung-Jin Yoon4, Hajeong Lee1
출판정보	2016; 2016(1):
키워드	Acute kidney injury
초록	Background: Postoperative acute kidney injury (AKI) is a serious adverse event which leads to higher mortality. Prediction for postoperative AKI can be helpful to identify high risk patients and to elicit early intervention. In this study, we incorporated preoperative and operative factors to build a prediction model for postoperative AKI using various machine learning methods. Methods: We included adults (age ≥ 18 years) who received noncardiac major (more than 1 hour duration) surgery in Seoul National University Hospital from 2004 to 2013. Baseline serum creatinine (Cr) was defined as an average of measurements for 3 months before surgery. We excluded patients with baseline Cr over 4 mg/dl or on renal replacement therapy (RRT). AKI was diagnosed according to both KDIGO and AKIN based on Cr measurements and initiation of RRT in 14 days after surgery. Variables considered in models were age, gender, department, diabetes mellitus, hypertension, chronic obstructive pulmonary disease, American Society of Anesthesiologists class, blood pressure, anesthesia type, emergent surgery, and surgery duration. Preoperative laboratory measurements were also utilized in analyses, including serum white blood cell count, hemoglobin, platelet, albumin, cholesterol, bilirubin, uric acid, BUN, creatinine, calcium, phosphorus, sodium, potassium and urinalysis. We developed prediction models with Naïve Bayes (NB), logistic regression (LR), voted perception (VP), bagging (BAG), JRip (JR) and random forest (RF) method, and calculated sensitivities and specificities of each method using 10-fold cross-validation. Results: For all 58,919 cases, RF among individual classifiers showed outstanding sensitivity for each of AKI definition (sensitivities under AKIN vs. KDIGO definition; NB, 0.492 vs. 0.428; LR, 0.693 vs. 0.647; VP, 0.719 vs. 0.684; BAG, 0.718 vs. 0.686; JR, 0.715 vs. 0.707; RF, 0.743 vs. 0.720; specificities; NB, 0.892 vs. 0.888; LR, 0.772 vs. 0.742; VP, 0.659 vs. 0.615; BAG, 0.726 vs. 0.687; JR, 0.717 vs. 0.657; RF, 0.736 vs. 0.695). Most learning methods were more sensitive and specific in detecting AKIN than KDIGO AKI. Subgroup analyses according to surgical departments revealed highest performance for gynecologic surgeries (sensitivities; NB, 0.764 vs. 0.704; LR, 0.655 vs. 0.722; VP, 0.764 vs. 0.722; BAG, 0.745 vs. 0.731; JR, 0.745 vs. 0.676; RF, 0.818 vs. 0.750) and lowest for urologic surgeries (NB, 0.764 vs. 0.704; LR, 0.655 vs. 0.722; VP, 0.764 vs. 0.722; BAG, 0.745 vs. 0.731; JR, 0.745 vs. 0.676; RF, 0.818 vs. 0.750). RF was a best prediction tool for most departments including general (sensitivities; 0.788 vs. 0.759), gynecologic (0.818 vs. 0.750), orthopedic (0.745 vs. 0.709) and urologic (0.668 vs. 0.728) surgeries, except neurologic (0.548 vs. 0.605) surgeries in which VP (0.710 vs. 0.685) was the most sensitive. Conclusion: Postoperative AKI after noncardiac major surgery can be predicted sensitively with preoperative and operative variables using machine learning methods.
원문(PDF)	PDF 원문보기

간행물 검색

대한신장학회 회원관리 규정