Options
利用機器學習模型預測糖尿病患兩年內中風之風險
Other Title
Machine Learning Algorithms for Predicting 12 months Risk of Stroke in Type 2 Diabetes Patients
Type
thesis
Date Issued
2022-06-07
Author(s)
ADAMA NS BAH
Advisor
SYED ABDUL SHABBIR
Subjects
系所名稱：醫學資訊研究所碩士班
Publisher
醫學資訊研究所碩士班
Description
口試委員：SYED ABDUL SHABBIR SYED ABDUL SHABBIR;YU WEI WU YU WEI WU;KEVIN CHANG KEVIN CHANG
開放校內, 開放日期為2022-07-19;校外, 開放日期為2023-06-10
開放校內, 開放日期為2022-07-19;校外, 開放日期為2023-06-10
Abstract
Background: Type 2 diabetes mellitus is most common in adults, but an increasing number of children and adolescents are also affected. Individuals with diabetes are at a twofold to fivefold increased risk of stroke compared to non-diabetes. The purpose of this study was to develop machine learning-based models for predicting the personalized risk of developing stroke in diabetes patients at 12 months.
Method: A retrospective cohort design using the MJ database on 1,894 individuals aged ≥35 years with type 2 diabetes mellitus, between 2002 to 2017. Using a propensity score pair matching procedure of 10:1 ratio, 825 T2DM were selected, and the data was used to run our prediction models. Random Forest, K-Nearest Neighbor, and XGBoost classifier were developed to perform our predictions.
Result: A sample of 1,894 participants were included in this study in which 75(3.9%) had first-time stroke. A sub-sample of 825 recruited and analyzed after propensity matching of 10:1 ratio, with an overall median age of 37 years. The AUC was (XGB: 90%, RF: 88%, KNN: 85%)respectively.
Conclusion: BMI, uricemia, TyG-Index, and diastolic blood pressure were found to be the best predictors of stroke in our study owing to the high-performance accuracy of our machine learning models.
Method: A retrospective cohort design using the MJ database on 1,894 individuals aged ≥35 years with type 2 diabetes mellitus, between 2002 to 2017. Using a propensity score pair matching procedure of 10:1 ratio, 825 T2DM were selected, and the data was used to run our prediction models. Random Forest, K-Nearest Neighbor, and XGBoost classifier were developed to perform our predictions.
Result: A sample of 1,894 participants were included in this study in which 75(3.9%) had first-time stroke. A sub-sample of 825 recruited and analyzed after propensity matching of 10:1 ratio, with an overall median age of 37 years. The AUC was (XGB: 90%, RF: 88%, KNN: 85%)respectively.
Conclusion: BMI, uricemia, TyG-Index, and diastolic blood pressure were found to be the best predictors of stroke in our study owing to the high-performance accuracy of our machine learning models.