請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/83280
標題: | 以整合機器學習設計探索與心血管疾病相關糖尿病動態路徑 Synthetic Design of Machine Learning for Cardiovascular Disease Process with Selected Pathways on Dynamics of Type 2 Diabetes Mellitus |
其他標題: | Synthetic Design of Machine Learning for Cardiovascular Disease Process with Selected Pathways on Dynamics of Type 2 Diabetes Mellitus |
作者: | 許容綺 Jung-Chi Hsu |
指導教授: | 陳秀熙 Hsiu-Hsi Chen |
關鍵字: | 糖尿病,糖尿病前期,心房顫動,馬可夫鏈分析,機器學習,隨機存活森林,貝氏網路,因果分析, diabetes mellitus,prediabetes,atrial fibrillation,multi-state Markov analysis,machine learning,random survival forest,Bayesian network,counterfactual analysis, |
出版年 : | 2023 |
學位: | 博士 |
摘要: | 背景
第二型糖尿病與胰島素阻抗性與心血管疾病與心律不整有關,同時共病因子也因疾病負荷增加。傳統的統計方法因只評估單一時間點的風險而忽略糖尿病在進展過程前的糖尿病前期所造成的風險。目前沒有相關研究顯示糖尿病前期是否會因代謝症候群增加心房顫動的風險。本篇論文我們使用各種不同的機器學習、多階段馬可夫鏈分析和貝氏網路分析,對心房顫動的風險建立統計模型,以期藉由介入措施達到初級、二級和三級預防的效益。 材料與方法 本篇論文涵蓋三部分。第一部分為分析台大醫院的資料庫,我們使用Cox存活分析、機器學習及隨機存活森林分析等方法,找出對心房顫動高風險的糖尿病前期族群及其相關重要特徵。第二部分分析台灣基隆地區長期社區篩檢追蹤的資料,我們使用馬可夫鏈探索從正常血糖到糖尿病前期到糖尿病的自然病史,並運用貝氏網路探討糖尿病前期患者在代謝症候群隨之進展時對心房顫動的影響。第三部分則進一步使用反事實因果分析模型來探討危險因子。 結果 在台大醫院的資料庫中,我們從Cox模型與Kaplan-Meier分析可知,糖尿病前期的人比正常血糖的人較易罹患心房顫動 (風險比為1.26, 95%信賴區間為1.13-1.40, p值<0.001)。我們比較了各種不同的機器學習方法來預測糖尿病前期的病患發生心房顫動,隨機森林有最高的接收者操作特徵曲線下面積,而當考慮時間因素時,隨機存活森林分析亦比傳統Cox模型有較高的預測能力,並且我們發現心臟衰竭指數與心臟超音波左心房與左心室參數的變化最為重要,其次為腎功能和年齡。非監督性群集分析則將糖尿病前期患者分成三個特徵群,年齡大與心臟衰竭有最高的心房顫動發生率,其次為腎功能較差的族群,最後是肥胖與高血脂的族群。 在社區研究的資料裡,我們使用馬可夫分析並計算從正常血糖發展到糖尿病前期、糖尿病前期到糖尿病、與糖尿病前期返回到正常血糖的每年發生率。每年大約一半的糖尿病前期可回復為正常血糖的人。空腹血糖異常的糖尿病前期當合併葡萄糖耐受不良時有較高的風險從糖尿病前期進展到第二型糖尿病。 貝氏網路分析顯示年紀、尿酸高、男性的糖尿病前期族群較易從無代謝症候群轉變成輕微代謝症候群或代謝症候群,而男性與較差的腎功能較易從代謝症候群階段發生心房顫動;而在調整多重共變項因子後,腎功能差則有正向因果貢獻從代謝症候群到心房顫動。在反事實假說下,若無糖尿病前期與異常腎功能時,相較於有糖尿病前期與異常腎功能的病人,發生心房顫動的風險五年內可以從0.071%降到0.023%,十年內從0.142%到0.047%,十五年內可以從0.212%降到0.070%。 結論 我們發現糖尿病前期有較高的心房顫動的發生率,並且高風險的特徵為心臟衰竭,腎功能差,與年老。在糖尿病前期的階段,有可能心臟結構的變化已經開始。而在社區長期追蹤下,大約有一半的糖尿病前期患者是可逆回正常血糖狀態。同為糖尿病前期的空腹血糖異常與葡萄糖耐受不良在疾病表現上各有不同的特性。當糖尿病前期合併代謝症候群時,男性和異常的腎功能伴隨著較高的風險發生心房顫動。 Background: The incidence and prevalence of diabetes mellitus have increased worldwide. Diabetes is known to raise the risk of atrial fibrillation (AF), however it is unclear if prediabetes, the intermediate stage before diabetes, also increases this risk. The purpose of this research is to determine risk factors for AF in individuals with prediabetes, to understand the natural course of prediabetes and diabetes, and to identify the high-risk prediabetes for AF. Methods: The research has three parts. In Part I, we evaluated individuals with prediabetes by using a medically-integrated database from a Taiwanese hospital and implementing supervised and unsupervised machine learning algorithms for AF prediction as well as time-to-event analysis. The machine learning techniques coverage logistic regression, decision tree, random forest, polynomial support vector machine (SVM), and artificial neural networks (ANN). In Part II, we examined the dynamic process of diabetes utilizing community data and developed a Bayesian network model from free of metabolic disorder (FMD), mild metabolic disorder (MMD), and metabolic syndrome (MetS) to atrial fibrillation (AF) utilizing the Keelung community-based integrated screening program. In Part III, we investigated the risk distribution for AF using Bayesian network counterfactual analysis. Results: We evaluated data from 174,835 National Taiwan University Hospital patients over a median follow-up period of 47.6 months. Prediabetes is independently associated with a significantly higher incidence of AF compared to normal glucose test (NGT) individuals (Hazard ratio [HR] 1.26, 95% confidence interval [CI) 1.13-1.40, p<0.001). This association was particularly pronounced in individuals with HbA1c levels greater than 5.5%. Elevated N-terminal natriuretic peptide (NT-proBNP) level and changed cardiac structure were the most significant factors, followed by worse renal function and advanced age. Hierarchical cluster analysis indicated three phenotypes of prediabetes with significantly different risks of atrial fibrillation (log-rank p<0.001). Between 1999 and 2009, 107,501 individuals were enrolled in the Keelung community-based screening program. Using a continuous-time Markov model, the diabetes transition rate was 14.4% from impaired fasting glucose (IFG) with impaired glucost tolerance (IGT) and 8.6% from IFG without IGT in males, and 13.0% from IFG with IGT and 7.9% from IFG without IGT in women. Approximately fifty percent of IFG participants reverted to NGT annually. Bayesian analysis revealed that older age, male gender, and abnormal renal function increased AF in prediabetes, whereas smoking and alcohol intake contributed to the transition from MMD to MetS. When hyperglycemia and elevated BUN were absent, the risk of AF decreased from 0.071% to 0.023% in five years, 0.142% to 0.047% in ten years, and 0.212% to 0.070% in fifteen years. Conclusions: Prediabetes increases the risk of atrial fibrillation. Important contributors to this increased risk include alterations in left heart anatomy, which may already be present in the prediabetes stage. Distinct characteristics were observed between IFG and IGT in a community-based cohort study. Prediabetes and metabolic syndrome, in conjunction with being male and having impaired renal function, have been associated with a higher incidence of AF. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/83280 |
DOI: | 10.6342/NTU202300287 |
全文授權: | 同意授權(限校園內公開) |
顯示於系所單位: | 流行病學與預防醫學研究所 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-111-1.pdf 授權僅限NTU校內IP使用(校園外請利用VPN校外連線服務) | 3.1 MB | Adobe PDF | 檢視/開啟 |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。