Cited 0 times in Scipus Cited Count

Mortality Risk Factors in Myocardial Injury after Noncardiac Surgery: Registry extracted via Clinical Data Warehouse (CDW)

DC Field Value Language
dc.contributor.advisor박, 래웅-
dc.contributor.author양, 광모-
dc.date.accessioned2021-11-10T05:52:56Z-
dc.date.available2021-11-10T05:52:56Z-
dc.date.issued2021-
dc.identifier.urihttp://repository.ajou.ac.kr/handle/201003/19863-
dc.description.abstractGlobally, 1.5% of adults who undergo inpatient noncardiac surgery die during the subsequent 30 days, and cardiac complications are the leading cause of postoperative deaths. Myocardial injury after noncardiac surgery (MINS) can be diagnosed by the level of cardiac troponin, which is estimated from the serum of the patient, and is related to the increased mortality after the operation. However, factors such as a high level of C-reactive protein at discharge, statins prescription at discharge, antiplatelet prescription at discharge, direct oral anticoagulants prescription at discharge, and antihypertensive drugs prescription at discharge that contribute to the mortality of MINS patients have not been comprehensively evaluated. Clinical data for MINS patients was extracted from the Clinical Data Warehouse (CDW) of the Samsung Medical Center (i.e., DARWIN-C). For mortality statistics at institutions other than ours, this system is consistently validated by the National Population Registry of the Korea National Statistical Office. To investigate the optimal model for predicting mortality, we used several widely used machine learning algorithms: a generalized boosted regression model (GBM), random forest (RF), support vector machines (SVM) with a radial basis function, classification and regression trees (CART), linear discriminant analysis (LDA), lasso/ridge/elastic net (GLMNET), k-nearest neighbors (KNN), extreme gradient boosting (XGB). Each model was evaluated for accuracy, precision, recall, F1 score, AUROC (area under the receiver operating characteristic) curve, and area under precision-recall curve (AUPRC) (area under the precision recall curve). We finally selected a model that maintains its performance even on external datasets, as assessed using a separately prepared external dataset. In this study we explored the role of clinical factors, by developing a mortality prediction model with all variables associated with MINS patients, and analyzing the impact of each factor on the performance of the model using the SHAP value. XGB outperformed the other models, producing the highest values of the performance indicators (AUROC 0.925, AUPRC 0.750, F1 score 0.642, and balanced accuracy 0.759). It was therefore selected to be used to analyze the impact of each clinical factor on mortality in MINS patients, by maintaining significant performance even on the test datasets prepared for performance validation. Factors such as a high level of CRP, antiplatelet agents, peak cardiac troponin, insulin, calcium channel blockers, operation duration, and statins were found to have a relatively large impact on 30-day mortality. Finally, in this observational cohort study, we confirmed that machine learning methods can be used to predict mortality in MINS patients. By identifying which factors in the model were most important for the prediction of mortality, we were able to select clinical markers that should be monitored, to reduce mortality.-
dc.description.abstract전 세계적으로 입원 비심장 수술을 받은 성인의 1.5%가 이후 30일 동안 사망하며, 수술 후 사망 원인은 주로 심장 합병증이다. 비심장 수술 후 심근 손상 (Myocardial injury after noncardiac surgery: MINS)은 Serum에서 측정한 Cardiac Troponin 수치를 기준으로 진단할 수 있으며 수술 후 사망률 증가와 관련이 있다. 이전의 연구들은 MINS 환자들의 임상적 요인들(e.g. High Level of C-reactive Protein, Statin, Antiplatelet, DOAC, HTN Drugs 등)을 독립적으로 평가했지만, 그 중 어떤 변수가 더 큰 영향을 미치는지에 대해서는 평가가 없었기에 이번 연구를 계획했다. MINS 환자들의 임상 데이터는 삼성서울병원 임상 데이터베이스 DARWIN-C에서 추출되었다. 본 시스템은 본원 외에서 사망한 경우 통계청 국가인구등록부를 통해 확인 및 자동으로 갱신했다. 사망률 예측에 있어 최적의 모델을 조사하기 위해, 널리 사용되는 기계 학습 알고리즘들을 훈련시켰다. 각 모델은 정확도, 정밀도, 재현율, F1 점수 및 AUROC(수신기 작동 특성 아래의 영역) 및 AUPRC (정밀도-재현율 곡선 아래 영역) 곡선에 따라 평가되었다. 그리고 별도로 준비된 데이터 세트를 통해 모델이 외부 데이터셋에서도 성능을 유지하는 모델을 최종적으로 선택했다. 본 연구에서는 MINS 환자와 관련된 모든 변수를 가지고 사망 예측 모델을 개발하고, 그 모델의 성능에 각 요인들이 미치는 영향을 SHAP value를 통해 분석함으로써 임상적 요인들의 역할을 탐구했다. XGB는 성능 지표(AUROC 0.925, AUPRC 0.750, F1 점수 0.642 및 균형 정확도 0.759)에서 높은 값을 보여 다른 모델을 능가함을 확인하였다. 성능 검증을 위해 준비된 외부 데이터 세트에 대해서도 상당한 성능을 유지함으로써 MINS 환자의 사망률에 대한 각 임상 요인이 미치는 영향을 분석하기 위해 최종적으로 선택되었다. 높은 수준의 CRP, 항혈소판제제, 피크 심장 트로포닌, 인슐린, 칼슘채널 차단제, 수술 시간, 스타틴 등의 요인이 30일 사망률에 상대적으로 큰 영향을 미치는 것으로 나타났다. 결론적으로 우리는 기계 학습 방법을 사용하여 MINS 환자의 30일 사망률을 성공적으로 예측할 수 있음을 확인했다. 또한 우리가 사용한 모델이 사망을 예측하기 위해 어떤 요인들을 더 중요하게 사용했는지 확인함으로써 환자의 사망률을 낮추기 위해 보다 주의 깊게 관찰해야 하는 임상적 변수들을 추려낼 수 있었다.-
dc.description.tableofcontentsI. Introduction 1
A. Background and necessity of the study 1
1. Importance of Myocardial Injury after Noncardiac Surgery 1
2. Understanding of Cardiac Troponin Test 2
3. Characteristics of High Sensitivity Cardiac Troponin Assay 4
4. Previous MINS study 7
B. Rational and Necessity of the study 9
1. Pharmacoeconomic Reasons 9
2. Clinical implication 10
3. Prediction of postoperative mortality 11
C. Purpose of the study 13
II. Methods 14
A. Study population & Data Curation 17
B. Value of Data Frame 21
C. Definition and Study Endpoints 25
D. Perioperative Management and cTn measurements 27
E. Prediction Models 29
F. Model Hyperparameter setting 32
G. Model evaluation 37
H. Statistical Analysis 39
III. Result 40
A. Patient Charateristics 40
B. Predictability (30-days mortality) 47
C. Predictability (1-year mortality) 64
D. Feature Importance of XGB (30-days) 67
E. SHAP Summary plot of XGB (30-days) 69
F. Light Prediction Model of MINS using RFE 80
1. Light Prediction Model using RFE (Variables = 27) 80
2. Light Prediction Model using RFE (Variables = 10) 96
3. Ultra-Light Prediction Model using RFE (Variables = 5) 110
G. Models created by selected variables 124
1.Light Prediction Model of MINS created by chosen 10 variables 124
2. Prediction models for MINS prior to surgery 127
IV. Discussion 132
A. Considerations on the study method and results 132
1. Analysis of model performance considering the asymmetry of data 132
2.Comparison of 30-days and 1-year preference model performance 134
3. Clinical Implication 135
4. Difference of SHAP value in Charlson and Updated Charlson score 136
5. Reduction of features 138
6. XGB is a better algorithm than the others 138
B. Limitations of the study 139
V. Conclusion 140
References 141
-
dc.language.isoen-
dc.titleMortality Risk Factors in Myocardial Injury after Noncardiac Surgery: Registry extracted via Clinical Data Warehouse (CDW)-
dc.title.alternative비심장수술 후 심근손상에 따른 사망위험 요인들: 임상데이터 웨어하우스(CDW)에서 추출한 레지스트리-
dc.typeThesis-
dc.identifier.urlhttp://dcoll.ajou.ac.kr:9080/dcollection/jsp/common/DcLoOrgPer.jsp?sItemId=000000031248-
dc.subject.keywordMINS-
dc.subject.keywordhigh sensitivity troponin-
dc.subject.keywordmachine learning-
dc.subject.keywordpredicting model-
dc.subject.keywordXGB-
dc.description.degreeDoctor-
dc.contributor.department대학원 의학과-
dc.contributor.affiliatedAuthor양, 광모-
dc.date.awarded2021-
dc.type.localTheses-
dc.citation.date2021-
dc.embargo.liftdate9999-12-31-
dc.embargo.terms9999-12-31-
Appears in Collections:
Theses > School of Medicine / Graduate School of Medicine > Doctor
Files in This Item:
There are no files associated with this item.

qrcode

해당 아이템을 이메일로 공유하기 원하시면 인증을 거치시기 바랍니다.

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.

Browse