What question did this study set out to answer?

The aim is to develop models for predicting thrombotic complications associated with PICCs in critically ill patients.

June 1, 2026Open Access

Machine learning prediction of PICC-associated thrombotic complications in critically ill patients

Key Points

The aim is to develop models for predicting thrombotic complications associated with PICCs in critically ill patients.
Retrospective cohort study utilizing data from the MIMIC-IV database with 8,145 PICC insertions.
Five machine learning models were developed, addressing class imbalance with SMOTE and optimizing hyperparameters.
Feature importance analysis was performed with 29 predictor variables including demographics, coagulation, and treatment variables.
Random Forest model achieved an AUC-ROC of 0.809 (95% CI: 0.780–0.838), significantly outperforming other models (all P < 0.01).
Feature importance identified prior thrombosis history as the most critical predictor affecting outcomes.
Sensitivity analyses confirmed robustness of findings across various patient subgroup analyses.

Abstract

Abstract Background Peripherally inserted central catheters (PICCs) are widely used vascular access devices in intensive care, yet thrombotic complications remain a significant clinical concern. Traditional risk assessment tools fail to capture the complex, non-linear interactions among thrombosis risk factors in critically ill patients. Methods This retrospective cohort study, conducted in adherence to TRIPOD + AI guidelines, developed and compared five machine learning models for predicting PICC-associated thrombotic complications using data from the Medical Information Mart for Intensive Care IV (MIMIC-IV) database. A total of 8,145 PICC insertions from 6,376 patients were analyzed, with 29 predictor variables spanning demographics, coagulation parameters, hematologic indices, renal function, vital signs, comorbidities, and treatment variables. Synthetic Minority Over-sampling Technique (SMOTE) addressed class imbalance (16.88% event rate), applied exclusively to the training set after the train-test split. Hyperparameters were optimized via five-fold cross-validated grid search. Results Random Forest achieved the best discriminative performance (AUC-ROC 0.809, 95% CI: 0.780–0.838, bootstrap n = 2,000), significantly outperforming all other models on DeLong’s test (all P < 0.01); its calibration slope of 1.369 indicated mild underconfidence. Decision curve analysis confirmed net clinical benefit for the three tree-based ensembles across threshold probabilities of approximately 1–50%. Feature importance analysis identified prior thrombosis history as the dominant predictor, followed by coagulation parameters and hematologic indices. Sensitivity analyses using patient-clustered data splitting, multiple imputation (m = 5, Rubin’s rules), first-PICC-only subsets, and a no-prior-thrombosis subgroup confirmed the robustness of the primary findings. Conclusions These findings support the development of machine learning-based clinical decision support tools for individualized thrombosis risk stratification in critical care settings. The completed TRIPOD + AI checklist is provided as Additional file 1.

Bookmark

View Full Paper

Cite This Study

Xia et al. (Sat,) studied this question.

synapsesocial.com/papers/6a1d22f702fbce9130638ade https://doi.org/https://doi.org/10.1186/s12911-026-03607-w

Bookmark

View Full Paper