Scholarlink

Background <p>Prediabetes affects approximately 98 million U.S. adults, yet its independent contribution to stroke risk and the underlying potential pathways remain incompletely characterized. We investigated the prediabetes-stroke association, quantified mediation effects, and developed machine learning-based risk prediction models.</p> Methods <p>This cross-sectional study analyzed data from 14,500 adults participating in the National Health and Nutrition Examination Survey (NHANES) from 2005 to 2018. Prediabetes was defined by fasting plasma glucose (100–125 mg/dL) and/or HbA1c (5.7–6.4%) in the absence of diabetes. The association between prediabetes and self-reported stroke was assessed using multivariable logistic regression and propensity score matching (PSM). Associative mediation analyses quantified effects through hypertension and body mass index (BMI). Six machine learning algorithms were developed and validated, with feature importance assessed via SHAP values.</p> Results <p>Among 14,500 participants, 3,318 had prediabetes. After comprehensive adjustment for demographic, lifestyle, and clinical factors, prediabetes was significantly and independently associated with prevalent stroke (OR 1.30, 95% CI: 1.06–1.60). Feature importance analyses ranked prediabetes sixth among 40 predictors (composite score 1.74), surpassing smoking and physical activity. Mediation analysis revealed that hypertension and BMI jointly mediated 35–45% of the prediabetes-stroke association. The ensemble machine learning model demonstrated superior discrimination (AUC 0.848, 95% CI: 0.833–0.863) compared to traditional logistic regression (AUC 0.776, 95% CI: 0.758–0.794) and exhibited excellent calibration. A very high-risk subgroup (1% of the population) was identified with a 27% stroke prevalence.</p> Conclusions <p>Prediabetes represents a clinically significant factor independently associated with stroke in the U.S. adult population. Hypertension and obesity partially mediate this risk, supporting comprehensive metabolic management. Machine learning-based risk stratification effectively identifies high-risk individuals, suggesting potential utility for personalized prevention strategies.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Prediabetes and stroke risk in U.S. adults: a machine learning-based analysis with causal mediation insights

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月23日

查看全文
  • Yaoquan He,
  • Jun Zhang,
  • Shuang Liu,
  • Jianhui Pan,
  • Jinliang Peng,
  • Xiang Li,
  • Zanzhi Wang,
  • Qinghong Liu,
  • Peng Huang

摘要

Background

Prediabetes affects approximately 98 million U.S. adults, yet its independent contribution to stroke risk and the underlying potential pathways remain incompletely characterized. We investigated the prediabetes-stroke association, quantified mediation effects, and developed machine learning-based risk prediction models.

Methods

This cross-sectional study analyzed data from 14,500 adults participating in the National Health and Nutrition Examination Survey (NHANES) from 2005 to 2018. Prediabetes was defined by fasting plasma glucose (100–125 mg/dL) and/or HbA1c (5.7–6.4%) in the absence of diabetes. The association between prediabetes and self-reported stroke was assessed using multivariable logistic regression and propensity score matching (PSM). Associative mediation analyses quantified effects through hypertension and body mass index (BMI). Six machine learning algorithms were developed and validated, with feature importance assessed via SHAP values.

Results

Among 14,500 participants, 3,318 had prediabetes. After comprehensive adjustment for demographic, lifestyle, and clinical factors, prediabetes was significantly and independently associated with prevalent stroke (OR 1.30, 95% CI: 1.06–1.60). Feature importance analyses ranked prediabetes sixth among 40 predictors (composite score 1.74), surpassing smoking and physical activity. Mediation analysis revealed that hypertension and BMI jointly mediated 35–45% of the prediabetes-stroke association. The ensemble machine learning model demonstrated superior discrimination (AUC 0.848, 95% CI: 0.833–0.863) compared to traditional logistic regression (AUC 0.776, 95% CI: 0.758–0.794) and exhibited excellent calibration. A very high-risk subgroup (1% of the population) was identified with a 27% stroke prevalence.

Conclusions

Prediabetes represents a clinically significant factor independently associated with stroke in the U.S. adult population. Hypertension and obesity partially mediate this risk, supporting comprehensive metabolic management. Machine learning-based risk stratification effectively identifies high-risk individuals, suggesting potential utility for personalized prevention strategies.