Scholarlink

Objective <p>Acute kidney injury (AKI) is a serious complication in critically ill patients, contributing to high morbidity and mortality. The stress hyperglycemia ratio (SHR), defined as the ratio of admission blood glucose to estimated average glucose derived from HbA1c, has been previously linked to outcomes in cardiovascular and cerebrovascular conditions, but its role in predicting short-term mortality in ICU patients with AKI remains unclear. This study explores the association between SHR and short-term mortality, and developed machine learning models to improve prognostic accuracy. </p> <p><b>Methods:</b> We retrospectively analyzed 5,555 ICU patients with AKI from the MIMIC-IV (v3.1) database. Patients were divided into SHR quartiles. Primary outcomes were 30- and 90-day all-cause mortality. The association between SHR and mortality was assessed using restricted cubic spline (RCS) modeling, Cox regression, and Kaplan–Meier analysis. Feature selection was performed using the Boruta algorithm and LASSO regression. Twelve machine learning models were developed and systematically compared for 30-day mortality prediction. Model performance was evaluated using multiple metrics, and the best-performing model was further assessed using decision curve analysis and calibration analysis. SHAP analysis was applied to interpret the contributions of individual features. </p> <p><b>Results:</b> A U-shaped relationship was found between SHR and mortality, with higher SHR levels significantly increasing the risk of both 30-day and 90-day death (HR &gt; 1, p &lt; 0.01). Subgroup analyses confirmed SHR's predictive reliability across multiple populations. Among the models, LightGBM demonstrated the best overall predictive performance (AUC = 0.864), outperforming traditional ICU scoring systems. The model showed good calibration and a favorable net clinical benefit. </p> <p><b>Conclusion:</b> SHR is an independent, nonlinear predictor of short-term mortality in ICU patients with AKI. Machine learning models incorporating SHR, especially LightGBM, significantly improve risk stratification, offering valuable support for early identification and personalized ICU management.</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

Prognostic value of stress hyperglycemia ratio in critically Ill patients with acute kidney injury: a machine learning-driven retrospective cohort analysis

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月18日

查看全文
  • Yichun Shuai,
  • Yan Liu,
  • Xiahong Yang,
  • Zhe Chen,
  • Qiaoqian Wan,
  • Jie Zhao,
  • Xin Wang

摘要

Objective

Acute kidney injury (AKI) is a serious complication in critically ill patients, contributing to high morbidity and mortality. The stress hyperglycemia ratio (SHR), defined as the ratio of admission blood glucose to estimated average glucose derived from HbA1c, has been previously linked to outcomes in cardiovascular and cerebrovascular conditions, but its role in predicting short-term mortality in ICU patients with AKI remains unclear. This study explores the association between SHR and short-term mortality, and developed machine learning models to improve prognostic accuracy.

Methods: We retrospectively analyzed 5,555 ICU patients with AKI from the MIMIC-IV (v3.1) database. Patients were divided into SHR quartiles. Primary outcomes were 30- and 90-day all-cause mortality. The association between SHR and mortality was assessed using restricted cubic spline (RCS) modeling, Cox regression, and Kaplan–Meier analysis. Feature selection was performed using the Boruta algorithm and LASSO regression. Twelve machine learning models were developed and systematically compared for 30-day mortality prediction. Model performance was evaluated using multiple metrics, and the best-performing model was further assessed using decision curve analysis and calibration analysis. SHAP analysis was applied to interpret the contributions of individual features.

Results: A U-shaped relationship was found between SHR and mortality, with higher SHR levels significantly increasing the risk of both 30-day and 90-day death (HR > 1, p < 0.01). Subgroup analyses confirmed SHR's predictive reliability across multiple populations. Among the models, LightGBM demonstrated the best overall predictive performance (AUC = 0.864), outperforming traditional ICU scoring systems. The model showed good calibration and a favorable net clinical benefit.

Conclusion: SHR is an independent, nonlinear predictor of short-term mortality in ICU patients with AKI. Machine learning models incorporating SHR, especially LightGBM, significantly improve risk stratification, offering valuable support for early identification and personalized ICU management.