個人化第二型前期及糖尿病風險預測模型

Abstract

2024年全球成人糖尿病盛行率已達11.1%，約影響5.89億人，預估2050年將上升至13%。我國2017至2020年糖尿病盛行率為10.3%，糖尿病前期盛行率則估計達25.5%。透過早期篩檢與生活介入有助於延緩病程與降低併發症風險，然而多數疾病自然史模型未納入糖尿病前期，無法完整了解病程。本研究將糖尿病前期納入分析，以完整描述第二型糖尿病自然史。研究目的為運用社區篩檢資料，結合流行病學與機器學習，建構糖尿病前期與第二型糖尿病之風險預測模型。資料來自2005至2018年彰化社區整合式篩檢計畫 (CHCIS)，共納入123,713位年滿30歲且具完整健康檢查紀錄之參與者。分析共涵蓋34個變項，包括人口學特徵、生理量測、生化指標、生活習慣與家族病史。依據美國糖尿病學會空腹血糖標準，將參與者分為正常血糖、糖尿病前期與第二型糖尿病三類。使用羅吉斯迴歸、Cox比例風險模型與隨機森林辨識風險因子，並以AUC評估其效能。同時建立連續時間馬可夫模型，估算轉移機率並模擬長期血糖變化過程。結果顯示，正常至糖尿病前期、糖尿病前期至無臨床糖尿病、無臨床至臨床糖尿病之轉移速率分別為0.0328、0.1764與0.0988/人年。風險因子的部分，BMI與年齡為所有階段之顯著風險因子，三酸甘油酯主要影響早期階段，而丙胺酸轉胺酶與尿酸則與後期進展相關。模型效能方面，三種模型預測正常血糖進展至糖尿病前期之AUC相對較低，顯示早期風險辨識仍具挑戰。個人化風險分層發現，極低風險族群20年後逾八成仍為正常血糖；中度風險族群中，約有一半維持正常血糖，另有19.2%發展為無臨床症狀第二型糖尿病，17.7%則發展為臨床第二型糖尿病；極高風險族群中，僅有一成維持正常血糖，54.1%則已發展為臨床第二型糖尿病。本研究結果有助於了解糖尿病前期與第二型糖尿病在風險因子與病程表現上的差異，強化早期預測與個人化介入之實務應用，為未來社區健康促進與公共衛生策略提供實證基礎。

Introduction Diabetes affects approximately 589 million adults globally, with a prevalence of 11.1% projected to rise to 13% by 2050. In Taiwan, diabetes prevalence reached 10.3% between 2017 and 2020, mirroring global trends. Prediabetes prevalence estimates vary widely across countries; in Taiwan, it is estimated at 25.5%. Early screening and lifestyle interventions are critical for arresting disease progression and reducing complications. However, prediabetes is often excluded from natural history models, precluding us from elucidating disease trajectories. This study aims to address this gap by explicitly incorporating prediabetes transitions to better characterize the natural history of diabetes. Aim This study aimed to develop risk prediction models for prediabetes and type 2 diabetes mellitus using community-based screening data, integrating epidemiological and machine learning approaches. Methods Data were drawn from the Changhua Community-based Integrated Screening (CHCIS) program (2005–2018), comprising 123,713 participants aged 30 years and older with complete health examination records. Thirty-four variables—including demographic characteristics, anthropometric measures, biochemical markers, health behaviors, and family history—were analyzed using logistic regression, Cox proportional hazards models, and random forest. A continuous-time Markov model was employed to estimate disease progression probabilities and simulate long-term transitions between glycemic states. Participants were classified as normoglycemic, prediabetic, or diabetic based on American Diabetes Association (ADA) fasting plasma glucose criteria. Results Estimated transition rates were 0.0328/person-year from normoglycemia to prediabetes, 0.1764/person-year from prediabetes to asymptomatic diabetes, and 0.0988/person-year from asymptomatic to clinical diabetes. BMI and age were significant risk factors across all stages, with triglycerides prominent in early stages and alanine aminotransferase (ALT) and uric acid in later stages. In terms of model performance, all three models yielded relatively lower AUCs in predicting the progression from normoglycemia to prediabetes, indicating that early risk identification remains a significant challenge. Personalized risk stratification showed over 80% of the very low-risk group remained normoglycemic after 20 years, indicating slow disease progression. In the moderate-risk group, about half maintained normal glucose levels, while 19.2% progressed to preclinical diabetes and 17.7% developed clinical type 2 diabetes. Conversely, only 10% of the very high-risk group remained normoglycemic, with 54.1% progressing to clinical diabetes, indicating a rapid disease trajectory. Conclusion The findings of this study contribute to a better understanding of the differences in risk factors and disease progression between prediabetes and type 2 diabetes. These insights enhance the practical application of early prediction and personalized intervention, providing an evidence-based foundation for future community health promotion and public health strategies.