Risk stratification and data-driven phenotyping of peripherally inserted central catheter-related bloodstream infection: an explainable machine learning analysis of a large critical care cohort.

BACKGROUND

Peripherally inserted central catheter-related bloodstream infections (PICC-CRBSIs) are significant causes of morbidity and mortality in critically ill patients. Therefore, improved risk stratification is required for targeted prevention. This study aimed to identify key predictors and novel patient phenotypes of PICC-CRBSI using explainable machine learning.

METHODS

We conducted a retrospective cohort study using data from the MIMIC-IV Critical Care database. Patients undergoing PICC insertion were categorized into PICC-CRBSI (n = 237) and control (n = 8,052) groups. A machine learning pipeline was developed incorporating LASSO regression for feature selection and SHAP analysis for model interpretation. Dimensionality reduction was performed using supervised UMAP, followed by HDBSCAN clustering to identify distinct patient phenotypes. The dataset was split into training (70%) and test (30%) sets to ensure robust validation.

RESULTS

Of the 8,289 patients included in the study, 237 (2.86%) developed PICC-CRBSIs. Unsupervised clustering revealed two distinct patient phenotypes, with Cluster 0 representing a high-risk subgroup (1.2% of cohort) showing significantly higher PICC-CRBSI incidence (100% vs. 0.00% in training set, 5.63% vs. 2.16% in test set, all p < 0.001). The clustering demonstrated excellent performance on the training set (Silhouette Coefficient: 0.914) and maintained good generalizability on the test set (Silhouette Coefficient: 0.765). These research results are used for experimental use by clinicians through an interactive clinical profiling tool for dynamic risk exploration.

CONCLUSION

The application of an explainable ML framework successfully identified key predictors and a novel high-risk patient phenotype for PICC-CRBSI. Our data-driven approach provides a promising pathway for enhancing risk stratification and developing targeted preventive strategies in critically ill patients, with potential for clinical implementation through interactive decision support tools.