Machine learning-based prediction model for postoperative delirium in non-cardiac surgery
- 주제(키워드) delirium , machine learning
- 주제(DDC) 610
- 발행기관 아주대학교
- 지도교수 박래웅
- 발행년도 2023
- 학위수여년월 2023. 8
- 학위명 박사
- 학과 및 전공 일반대학원 의학과
- 실제URI http://www.dcollection.net/handler/ajou/000000032836
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Delirium is a reversible and preventable condition that causes temporary confusion and poses a significant burden on patients, families, and healthcare providers. It is associated with longer hospital stays, increased healthcare costs, complications, readmission rates, and even in-hospital deaths. Delirium is a common complication among surgical patients, with its occurrence varying between 5% to 40% depending on the type of surgery. Previous studies have demonstrated that screening for delirium can lead to a higher rate of diagnosis and early intervention, resulting in a reduction in the duration and complications of delirium. Although effective perioperative interventions based on predicting delirium have been described, accurately predicting postoperative delirium remains challenging, and a universally accepted prediction tool is not yet available. Various factors beyond the type of surgery contribute to the occurrence of delirium, including factors indicating brain vulnerability and external stressors that affect cognitive function. Furthermore, there is an overlap between factors that predispose individuals to delirium and those that trigger it, making prevention during postoperative care a complex process. In order to overcome the challenge of accurately predicting postoperative delirium, the authors of the study turned to machine learning techniques, which have gained popularity in recent studies focused on identifying predictors of delirium. These models are capable of handling large numbers of variables in nonlinear and interactive ways. A recent study demonstrated that a machine learning model outperformed traditional clinician-based regression models in predicting postoperative delirium, but this model has not yet been validated in other patient populations. Another previous model was based on a small sample size and limited to an older age group. To determine the most effective model for predicting delirium, the authors of this study utilized clinical data from the Clinical Data Warehouse (CDW) of Samsung Medical Center (known as DARWIN-C), which included data from surgical patients. The authors trained several commonly used machine learning algorithms, including XGB (extreme gradient boosting), RF (random forest), LR (logistic regression), and NB (Naive Bayes), and compared their performance based on metrics such as accuracy, precision, recall, F1 score, AUROC (area under the receiver operating characteristic) curve, and AUPRC (area under the precision recall curve). Furthermore, the authors of the study aimed to develop a prediction model that could perform consistently well even when applied to external datasets. To achieve this, they created a separate dataset from another center and tested the model's performance on it. The study's main objective was to investigate the impact of various clinical factors on delirium prediction. The authors used machine learning techniques and developed a delirium prediction model that included all variables associated with delirium. They then analyzed the influence of each factor on the model's performance using SHAP value analysis. Among the four commonly used machine learning algorithms (XGB, RF, LR, and NB), the XGB algorithm demonstrated the best performance, with higher values in the performance indicators (AUROC 0.902, AUPRC 0.170, F1 score 0.136, and balanced accuracy 0.855), and was ultimately selected for analyzing the effect of each clinical factor on delirium. The authors selected the top five variables that had a significant impact on delirium and analyzed the influence of each clinical factor while maintaining a consistent level of performance, even on the test datasets prepared for performance validation. The results of the study showed that age, operation duration, physical status classification, male sex, and surgical risk were significant factors that had a relatively large impact on delirium. In conclusion, he authors of this cohort study found that machine learning methods can accurately predict delirium. Additionally, by analyzing the factors used by the model, they identified specific clinical markers that should be closely monitored to prevent delirium.
more목차
I. Introduction 1
A. Background and necessity of the study 1
1. Importance of Delirium after Surgery 1
2. Understanding Mechanism of Postoperative Delirium 5
3. Previous studies on Postoperative Delirium 11
B. Rational and Necessity of the Study 22
1. Pharmaco-economic Reasons 22
2. Clinical implication of prediction of postoperative delirium 23
C. Purpose of the study 27
II. Methods 28
A. Study population & Data Curation 30
B. Values of Da Features of Dataset ta Frame 33
C. Definition and Study Endpoints 37
D. Prediction model development 39
E. Model evaluation 50
F. Sub-analysis of internal validation dataset and external validation 54
G. Statistical Analysis 57
III. Result 61
A. Patient Characteristics 61
B. Development of prediction model 68
C. Development of prediction model with the entire variables 71
D. Development of prediction model using selected variables 74
E. Publicly accessible delirium prediction model 78
F. External Validation of Prediction Model 81
G. Clinical Usefulness of Prediction Models 83
H. Calibration of Prediction Model 89
IV. Discussion 92
A. The main finding of the study 92
B. Methodologic considerations to build models 95
C. The clinical implication of the study 98
D. The clinical strength of the study 100
E. Clinical suitability of the results 102
F. Clinical usefulness of the prediction model 105
G. Analysis of model performance considering the asymmetry of data 107
H. Limitations of the study 110
V. Conclusion 115
References 116
