Scholarlink

Background <p>The stress hyperglycemia ratio (SHR), which quantifies the degree of acute hyperglycemia relative to a patient’s chronic glycemic background, serves as a biomarker reflecting acute metabolic dysregulation, yet its association with mortality among critical ill COVID-19 patients remains ambiguous. This study aims to elucidate this relationship and integrate SHR into machine learning models to predict mortality risk.</p> Method <p>This study retrospectively analyzed critically ill COVID-19 patients from the Northwestern ICU (NWICU) database. The primary and secondary outcomes were all-cause ICU and in-hospital mortality, respectively. To assess the relationship between SHR and mortality, a comprehensive statistical framework was employed, incorporating multivariable modeling, survival analysis, and nonlinear trend assessment, alongside subgroup analyses for robustness. Mortality risk was predicted using five machine learning (ML) algorithms. After identifying the optimal model, a parsimonious feature set was selected using variable importance ranking and the one-standard-error rule. The final model was deployed as an interactive Shiny web application.</p> Result <p>Our retrospective cohort comprised 4,151 COVID-19 critically ill patients (58.4% male), with multivariable-adjusted analyses revealing SHR as an independent predictor of all-cause ICU mortality (HR = 1.383, 95% CI:1.179–1.622) and in-hospital mortality (HR = 1.266, 95% CI:1.097–1.462; both <i>P</i> &lt; 0.001). The association between SHR and mortality exhibited a nonlinear pattern in restricted cubic spline (RCS) analyses, as indicated by significant nonlinearity for ICU and in-hospital mortality (<i>P</i> = 0.0017 and 0.0315, respectively). Similarly, subgroup analyses indicated that these relationships were attenuated in patients with diabetes or those receiving insulin therapy (P for interaction &lt; 0.05). Among five candidate machine learning models, the extreme gradient boosting (XGBoost) algorithm achieved optimal discrimination for ICU mortality, yielding a mean area under the curve (AUC) of 0.800 ± 0.027.</p> Conclusion <p>SHR is an independent predictor of mortality in COVID-19 critical ill patients. Its incorporation into machine learning models may improve risk stratification and assist in informing bedside decisions.</p> Clinical trial number <p>Not applicable.</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

Stress hyperglycemia ratio improves machine learning-based mortality risk prediction in critically ill COVID-19 patients: a multicenter retrospective study

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月16日

查看全文
  • Jiaxing Du,
  • Keze Ma,
  • Zhiwei Ye,
  • Juanli Song,
  • Sujun Chen,
  • Jie Yu,
  • Bing Liu,
  • Zixuan Jiang,
  • Fen Zhang

摘要

Background

The stress hyperglycemia ratio (SHR), which quantifies the degree of acute hyperglycemia relative to a patient’s chronic glycemic background, serves as a biomarker reflecting acute metabolic dysregulation, yet its association with mortality among critical ill COVID-19 patients remains ambiguous. This study aims to elucidate this relationship and integrate SHR into machine learning models to predict mortality risk.

Method

This study retrospectively analyzed critically ill COVID-19 patients from the Northwestern ICU (NWICU) database. The primary and secondary outcomes were all-cause ICU and in-hospital mortality, respectively. To assess the relationship between SHR and mortality, a comprehensive statistical framework was employed, incorporating multivariable modeling, survival analysis, and nonlinear trend assessment, alongside subgroup analyses for robustness. Mortality risk was predicted using five machine learning (ML) algorithms. After identifying the optimal model, a parsimonious feature set was selected using variable importance ranking and the one-standard-error rule. The final model was deployed as an interactive Shiny web application.

Result

Our retrospective cohort comprised 4,151 COVID-19 critically ill patients (58.4% male), with multivariable-adjusted analyses revealing SHR as an independent predictor of all-cause ICU mortality (HR = 1.383, 95% CI:1.179–1.622) and in-hospital mortality (HR = 1.266, 95% CI:1.097–1.462; both P < 0.001). The association between SHR and mortality exhibited a nonlinear pattern in restricted cubic spline (RCS) analyses, as indicated by significant nonlinearity for ICU and in-hospital mortality (P = 0.0017 and 0.0315, respectively). Similarly, subgroup analyses indicated that these relationships were attenuated in patients with diabetes or those receiving insulin therapy (P for interaction < 0.05). Among five candidate machine learning models, the extreme gradient boosting (XGBoost) algorithm achieved optimal discrimination for ICU mortality, yielding a mean area under the curve (AUC) of 0.800 ± 0.027.

Conclusion

SHR is an independent predictor of mortality in COVID-19 critical ill patients. Its incorporation into machine learning models may improve risk stratification and assist in informing bedside decisions.

Clinical trial number

Not applicable.