Scholarlink

Background <p>Sepsis remains a leading cause of death in critically ill patients. Host mRNA biomarkers may provide complementary biological information for prognostic assessment in sepsis. This study aimed to integrate host mRNA biomarkers and clinical parameters to develop and compare predictive models for short-term mortality using both conventional and machine learning (ML) approaches.</p> Methods <p>In this prospective ICU study (Dec 2022–Mar 2024), 249 patients meeting Sepsis-3 criteria were initially enrolled, and 198 patients with available host mRNA measurements were included in the final analysis. 35 candidate mRNAs were quantified using droplet digital PCR (ddPCR). After initial transcriptomic screening and LASSO-based feature selection, one immune transcript (PAX5) and four clinical variables (hematologic disease, sex, SOFA score, positive blood culture) were retained for model development. Five models were evaluated for predicting 7-, 14-, and 28-day mortality, including Cox proportional hazards regression, random survival forest (RSF), Deep Survival Analysis (DeepSurv), support vector machine (SVM), and extreme gradient boosting (XGBoost). Model discrimination was assessed using AUCs and compared using DeLong’s test. Internal validation was performed using the bootstrap 0.632 method.</p> Results <p>Among the 198 patients included in the final analysis, mortality was 18.7% at 7 days, 28.3% at 14 days, and 30.3% at 28 days. RSF achieved the highest AUCs for 7-, 14-, and 28-day mortality (0.834, 0.854, and 0.847, respectively). DeLong’s test showed that RSF significantly outperformed Cox regression, DeepSurv, and survival SVM at all time points, while its performance was comparable to XGBoost. Bootstrap 0.632 internal validation confirmed favorable overall performance of RSF. In variable importance analysis, hematologic disease, sex, SOFA score, PAX5 and positive blood culture were among the main contributors to model performance. Incremental analysis showed that clinical variables contributed the largest improvement beyond SOFA, whereas inclusion of PAX5 provided complementary prognostic value.</p> Conclusion <p>The integration of host mRNA biomarkers with clinical parameters supported the development of a ML–based prognostic framework for short-term mortality in sepsis. RSF showed favorable discriminative performance in this cohort and performed comparably to XGBoost. PAX5 may serve as a complementary prognostic biomarker, although clinical variables remained the dominant contributors to model discrimination.</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

Integrating host mRNA signatures and clinical trajectories with machine learning for risk stratification and survival prediction in sepsis: an ICU-based prospective cohort study

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月22日

查看全文
  • Zhenping Wu,
  • Liming Zhao,
  • Yinan Yao,
  • Haoran Ma,
  • Xia Zheng,
  • Jiang Xia,
  • Hua Bai,
  • Hao Yu,
  • Yunsong Yu,
  • Hongliu Cai,
  • Jifang Zhou,
  • Hua Zhou

摘要

Background

Sepsis remains a leading cause of death in critically ill patients. Host mRNA biomarkers may provide complementary biological information for prognostic assessment in sepsis. This study aimed to integrate host mRNA biomarkers and clinical parameters to develop and compare predictive models for short-term mortality using both conventional and machine learning (ML) approaches.

Methods

In this prospective ICU study (Dec 2022–Mar 2024), 249 patients meeting Sepsis-3 criteria were initially enrolled, and 198 patients with available host mRNA measurements were included in the final analysis. 35 candidate mRNAs were quantified using droplet digital PCR (ddPCR). After initial transcriptomic screening and LASSO-based feature selection, one immune transcript (PAX5) and four clinical variables (hematologic disease, sex, SOFA score, positive blood culture) were retained for model development. Five models were evaluated for predicting 7-, 14-, and 28-day mortality, including Cox proportional hazards regression, random survival forest (RSF), Deep Survival Analysis (DeepSurv), support vector machine (SVM), and extreme gradient boosting (XGBoost). Model discrimination was assessed using AUCs and compared using DeLong’s test. Internal validation was performed using the bootstrap 0.632 method.

Results

Among the 198 patients included in the final analysis, mortality was 18.7% at 7 days, 28.3% at 14 days, and 30.3% at 28 days. RSF achieved the highest AUCs for 7-, 14-, and 28-day mortality (0.834, 0.854, and 0.847, respectively). DeLong’s test showed that RSF significantly outperformed Cox regression, DeepSurv, and survival SVM at all time points, while its performance was comparable to XGBoost. Bootstrap 0.632 internal validation confirmed favorable overall performance of RSF. In variable importance analysis, hematologic disease, sex, SOFA score, PAX5 and positive blood culture were among the main contributors to model performance. Incremental analysis showed that clinical variables contributed the largest improvement beyond SOFA, whereas inclusion of PAX5 provided complementary prognostic value.

Conclusion

The integration of host mRNA biomarkers with clinical parameters supported the development of a ML–based prognostic framework for short-term mortality in sepsis. RSF showed favorable discriminative performance in this cohort and performed comparably to XGBoost. PAX5 may serve as a complementary prognostic biomarker, although clinical variables remained the dominant contributors to model discrimination.