Scholarlink

Background <p>Heart failure (HF) remains a major cause of morbidity and mortality worldwide, and acute decompensation frequently necessitates intensive care. Early identification of high-risk patients is essential, yet traditional HF risk scores were developed largely in chronic or ward-based cohorts and often fail to capture early physiologic deterioration in the ICU. Explainable machine-learning (ML) models may improve early risk stratification.</p> Methods <p>Using the MIMIC-IV v3.1 database, we conducted a retrospective study of adult ICU patients with heart failure identified by ICD-10 codes (I50.x) between October 1, 2015 and 2022. Predictors included demographics, comorbidities, pre-ICU cardiovascular medications, and vital signs and laboratory tests obtained within 0–6&#xa0;h of ICU admission. Implausible values were removed; missing data were imputed using medians or modes. An XGBoost classifier was trained with an 80/20 stratified split and class weighting. Performance in the test cohort was assessed using AUC, average precision, sensitivity, specificity, F1-score, and Brier score. SHapley Additive exPlanations (SHAP) were used for global and feature-level interpretability.</p> Results <p>A total of 12,110 ICU patients with HF were included, of whom 2,041 (16.9%) died in hospital. In the test cohort (<i>n</i> = 2,422), the model achieved an AUC of 0.797 (95% CI, 0.772–0.822) and an average precision of 0.491. Sensitivity and specificity were 0.632 and 0.793 at the default threshold; using the Youden-optimal threshold increased sensitivity to 0.708. Kaplan–Meier analysis demonstrated significant separation of survival curves across predicted risk quartiles (log-rank <i>P</i> &lt; 0.001), supporting the survival stratification of early risk predictions. SHAP analysis identified FiO₂, age, lactate, Charlson comorbidity index, BUN, and systolic blood pressure as the strongest predictors, revealing clinically coherent non-linear relationships and interactions.</p> Conclusion <p>These findings suggest that an interpretable model based on routinely available early ICU data may facilitate structured and quantitative risk stratification within the first hours of admission, thereby assisting clinicians in identifying high-risk patients who may benefit from closer monitoring and timely escalation of supportive care, rather than replacing clinical judgment.</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

Early prediction of in-hospital mortality in ICU patients with heart failure using an interpretable machine learning model: a retrospective study based on MIMIC-IV

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月7日

查看全文
  • Meiqi Sun,
  • Ping Lu,
  • Jing Cheng,
  • Linzhi Zhao,
  • Junyang Ma

摘要

Background

Heart failure (HF) remains a major cause of morbidity and mortality worldwide, and acute decompensation frequently necessitates intensive care. Early identification of high-risk patients is essential, yet traditional HF risk scores were developed largely in chronic or ward-based cohorts and often fail to capture early physiologic deterioration in the ICU. Explainable machine-learning (ML) models may improve early risk stratification.

Methods

Using the MIMIC-IV v3.1 database, we conducted a retrospective study of adult ICU patients with heart failure identified by ICD-10 codes (I50.x) between October 1, 2015 and 2022. Predictors included demographics, comorbidities, pre-ICU cardiovascular medications, and vital signs and laboratory tests obtained within 0–6 h of ICU admission. Implausible values were removed; missing data were imputed using medians or modes. An XGBoost classifier was trained with an 80/20 stratified split and class weighting. Performance in the test cohort was assessed using AUC, average precision, sensitivity, specificity, F1-score, and Brier score. SHapley Additive exPlanations (SHAP) were used for global and feature-level interpretability.

Results

A total of 12,110 ICU patients with HF were included, of whom 2,041 (16.9%) died in hospital. In the test cohort (n = 2,422), the model achieved an AUC of 0.797 (95% CI, 0.772–0.822) and an average precision of 0.491. Sensitivity and specificity were 0.632 and 0.793 at the default threshold; using the Youden-optimal threshold increased sensitivity to 0.708. Kaplan–Meier analysis demonstrated significant separation of survival curves across predicted risk quartiles (log-rank P < 0.001), supporting the survival stratification of early risk predictions. SHAP analysis identified FiO₂, age, lactate, Charlson comorbidity index, BUN, and systolic blood pressure as the strongest predictors, revealing clinically coherent non-linear relationships and interactions.

Conclusion

These findings suggest that an interpretable model based on routinely available early ICU data may facilitate structured and quantitative risk stratification within the first hours of admission, thereby assisting clinicians in identifying high-risk patients who may benefit from closer monitoring and timely escalation of supportive care, rather than replacing clinical judgment.