Scholarlink

Background <p>Urinary tract infections (UTIs) commonly have complications in the form of sepsis. We used machine learning (ML) via MIMIC-IV clinical data coupled with bi-directional Mendelian randomization to (i) diagnose UTI inpatients with sepsis and to (ii) test causal inference between UTI and sepsis.</p> Methods <p>We internally validated seven ML classifiers (including Random Forest [RF]) using 10-fold cross-validation on 13,567 UTI inpatients. The feature selection strategy used a stepwise logistic regression approach. To summarize model performance, the area under the curve (AUC), accuracy, sensitivity, and specificity were calculated. We then used two-sample MR (with inverse-variance weighted [IVW] acting as the primary method), leveraging European GWAS for UTI and sepsis with sensitivity analysis provided using MR Egger, weighted median, MR PRESSO, heterogeneity tests, pleiotropy assessments, and leave-one-out analyses.</p> Results <p>The multivariable model offered predictors of age, sex, congestive heart failure, hypertension, liver disease, obesity, urolithiasis, fluid–electrolyte abnormalities, diabetes, white blood cell count, hemoglobin, creatinine, proteinuria, hematuria, red blood cell distribution width (RDW), the first pathogen culture result, along with the Charlson Comorbidity Index. Regarding methods of machine learning, the Random Forest model performed best (test AUC: 0.948; accuracy: 0.873; sensitivity: 0.842; and specificity: 0.905). The MR analysis supported the hypothesis that genetic liability to UTI increased the risk of sepsis (IVW OR: 1.271, 95% CI: 1.158–1.395; for heterogeneity or horizontal pleiotropy, no evidence of either), whereas reverse Mendelian randomization indicated a modest effect of sepsis liability on UTI risk (IVW OR: 1.044, 95% CI: 1.002–1.089; however, very low power.)</p> Conclusions <p>This research merges machine learning for diagnostic classification with Mendelian Randomization for causal inference, addressing a gap in the literature by providing a complete approach to investigating the relationship between UTIs and sepsis. We included a high-performing diagnostic RF classifier that distinguishes UTI inpatients with versus without sepsis using readily available clinical variables, and our MR supports the notion that there is a causal component of liability from UTI to sepsis. These results are more consistent in support of previous recognition of in-hospital events than forthcoming risk analysis of prediction, and more external validation is needed.</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

The relationship between urinary tract infections and sepsis: results from Mendelian randomization and MIMIC data study

  • 期刊论文

  • 发表日期：

    2026年3月6日

查看全文
  • Aji Kaisaier,
  • Maimaijiang Munila,
  • Guanglu Song

摘要

Background

Urinary tract infections (UTIs) commonly have complications in the form of sepsis. We used machine learning (ML) via MIMIC-IV clinical data coupled with bi-directional Mendelian randomization to (i) diagnose UTI inpatients with sepsis and to (ii) test causal inference between UTI and sepsis.

Methods

We internally validated seven ML classifiers (including Random Forest [RF]) using 10-fold cross-validation on 13,567 UTI inpatients. The feature selection strategy used a stepwise logistic regression approach. To summarize model performance, the area under the curve (AUC), accuracy, sensitivity, and specificity were calculated. We then used two-sample MR (with inverse-variance weighted [IVW] acting as the primary method), leveraging European GWAS for UTI and sepsis with sensitivity analysis provided using MR Egger, weighted median, MR PRESSO, heterogeneity tests, pleiotropy assessments, and leave-one-out analyses.

Results

The multivariable model offered predictors of age, sex, congestive heart failure, hypertension, liver disease, obesity, urolithiasis, fluid–electrolyte abnormalities, diabetes, white blood cell count, hemoglobin, creatinine, proteinuria, hematuria, red blood cell distribution width (RDW), the first pathogen culture result, along with the Charlson Comorbidity Index. Regarding methods of machine learning, the Random Forest model performed best (test AUC: 0.948; accuracy: 0.873; sensitivity: 0.842; and specificity: 0.905). The MR analysis supported the hypothesis that genetic liability to UTI increased the risk of sepsis (IVW OR: 1.271, 95% CI: 1.158–1.395; for heterogeneity or horizontal pleiotropy, no evidence of either), whereas reverse Mendelian randomization indicated a modest effect of sepsis liability on UTI risk (IVW OR: 1.044, 95% CI: 1.002–1.089; however, very low power.)

Conclusions

This research merges machine learning for diagnostic classification with Mendelian Randomization for causal inference, addressing a gap in the literature by providing a complete approach to investigating the relationship between UTIs and sepsis. We included a high-performing diagnostic RF classifier that distinguishes UTI inpatients with versus without sepsis using readily available clinical variables, and our MR supports the notion that there is a causal component of liability from UTI to sepsis. These results are more consistent in support of previous recognition of in-hospital events than forthcoming risk analysis of prediction, and more external validation is needed.