Scholarlink

Background <p>Precise subtyping is crucial for enabling personalized treatments in sepsis patients. This study aims to develop an analytical framework for sepsis intervention, integrating variable selection and phenotype discovery for diverse settings.</p> Methods <p>Using MIMIC-IV database, study included sepsis patients (ICD-9/10 codes), excluding those with &lt; 24-hour stay or missing data. Data included demographics, labs, comorbidities, and follow-up status. The primary endpoint was 90-day all-cause mortality.First, variables significantly associated with 90-day mortality were preliminarily screened using multivariate Cox regression analysis.Subsequently, the optimal machine learning model was selected based on the C-index evaluation, and this model was used to further identify core factors from the preliminarily screened variables. Finally, patients were subgrouped using K-means clustering based on the characteristics of the core factors, and the subgroups were validated through baseline and survival analyses.The optimal number of clusters (K value) was determined using both the elbow method and the gap statistic.</p> Results <p>6,086 patients included (90-day mortality 23.87%). Cox regression identified 24 independent predictors (all <i>p</i> &lt; 0.001), and a high-performance Ridge risk prediction model was developed (average C-index = 0.761). Multidimensional clustering based on 9 core variables(|coefficient|&gt;0.1) revealed distinct subtypes (K = 4, 8,14, respectively). Increasing cluster granularity (K from 4 to 14) revealed converging characteristics across subgroups but refined risk stratification. Current ICU admission was identified as a key protective factor, while past ICU history, high age, poor cardiopulmonary function, hypermagnesemia, and PTT/RDW abnormalities were the three core high-risk drivers. The low-risk group (younger age+current ICU admission + no ICU history+better cardiopulmonary function) had a lowest mortality rate of 9.7% in 4-subtype system and 3.4% in the 14-subtype system. Patients without current ICU protection or with the other risk factors showed higher mortality risk than the low-risk group. Notably, extremely high level of RDW (21.6%, subgroup 14, mortality rate 45.42%) and PTT values (130.35s, subgroup 10, mortality rate 33.78%) were independent of age and cardiopulmonary issues, suggesting they define two distinct high-risk subtypes.</p> Conclusions <p>This study developed a clinically universal and interpretable sepsis subtyping framework using routine clinical variables, identifying subgroups with significant mortality differences, providing a pragmatic tool for risk stratification and prognosis assessment.</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

Machine learning-based precision subtyping and risk prediction in sepsis: a retrospective analysis using MIMIC-IV database

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月21日

查看全文
  • Yu Li,
  • Wenjian Luo,
  • Qian qian Zhang,
  • Fuhai Bai,
  • Jing Wei,
  • Ling Tang,
  • Youliang Deng,
  • Dukun Zuo,
  • Taotao Peng,
  • Hong Li,
  • Zonghong Long

摘要

Background

Precise subtyping is crucial for enabling personalized treatments in sepsis patients. This study aims to develop an analytical framework for sepsis intervention, integrating variable selection and phenotype discovery for diverse settings.

Methods

Using MIMIC-IV database, study included sepsis patients (ICD-9/10 codes), excluding those with < 24-hour stay or missing data. Data included demographics, labs, comorbidities, and follow-up status. The primary endpoint was 90-day all-cause mortality.First, variables significantly associated with 90-day mortality were preliminarily screened using multivariate Cox regression analysis.Subsequently, the optimal machine learning model was selected based on the C-index evaluation, and this model was used to further identify core factors from the preliminarily screened variables. Finally, patients were subgrouped using K-means clustering based on the characteristics of the core factors, and the subgroups were validated through baseline and survival analyses.The optimal number of clusters (K value) was determined using both the elbow method and the gap statistic.

Results

6,086 patients included (90-day mortality 23.87%). Cox regression identified 24 independent predictors (all p < 0.001), and a high-performance Ridge risk prediction model was developed (average C-index = 0.761). Multidimensional clustering based on 9 core variables(|coefficient|>0.1) revealed distinct subtypes (K = 4, 8,14, respectively). Increasing cluster granularity (K from 4 to 14) revealed converging characteristics across subgroups but refined risk stratification. Current ICU admission was identified as a key protective factor, while past ICU history, high age, poor cardiopulmonary function, hypermagnesemia, and PTT/RDW abnormalities were the three core high-risk drivers. The low-risk group (younger age+current ICU admission + no ICU history+better cardiopulmonary function) had a lowest mortality rate of 9.7% in 4-subtype system and 3.4% in the 14-subtype system. Patients without current ICU protection or with the other risk factors showed higher mortality risk than the low-risk group. Notably, extremely high level of RDW (21.6%, subgroup 14, mortality rate 45.42%) and PTT values (130.35s, subgroup 10, mortality rate 33.78%) were independent of age and cardiopulmonary issues, suggesting they define two distinct high-risk subtypes.

Conclusions

This study developed a clinically universal and interpretable sepsis subtyping framework using routine clinical variables, identifying subgroups with significant mortality differences, providing a pragmatic tool for risk stratification and prognosis assessment.