Development and validation of an interpretable machine learning model for predicting intravenous contrast media extravasation risk in CT imaging
- 1Department of Radiology, Tongji Hospital, Tongji Medical College, Huazhong University of Science and Technology
- 2Medical Affairs Team, GE Healthcare
Abstract
Objectives Intravenous contrast media extravasation (ICME) is a frequent and potentially severe complication during contrast-enhanced computed tomography (CECT). Although prevention of ICME is critical in radiology practice, valid and clinically practical prediction models remain limited. This study aimed to develop and validate an interpretable machine learning (ML) model for predicting the risk of ICME in patients undergoing CECT. Patients and methods This multicenter, retrospective case–control study analyzed 500 patients (252 ICME cases, 248 controls) from three institutions. Predictors encompassed patient demographics, clinical history, procedural details, venous access characteristics, and the diurnal temperature difference. From 22 initial candidates, 14 features were selected using least absolute shrinkage and selection operator (LASSO) regression. Five machine learning models—logistic regression, support vector machine (SVM), random forest, XGBoost, and K-nearest neighbors (KNN)—were trained and evaluated on an independent test set. The optimal model was assessed based on discrimination, calibration, and clinical utility, with its predictions explained using SHapley Additive exPlanations (SHAP). Results Among the five models, Logistic Regression demonstrated superior and the most generalizable performance on the test set, achieving an area under the receiver operating characteristic curve (AUC) of 0.82 (95% CI 0.72, 0.89), with a sensitivity of 78%, a precision of 75%, and an F1-score of 0.76. It also showed excellent calibration (Hosmer–Lemeshow p = 0.067). The other models' test AUCs were 0.81 (95% CI 0.71, 0.89) for SVM, 0.80 (95% CI 0.71, 0.88) for Random Forest, 0.79 (95% CI 0.68, 0.87) for XGBoost, and 0.74 (95% CI 0.64, 0.83) for KNN. Decision curve analysis confirmed the clinical utility of the Logistic Regression model across a broad range of threshold probabilities, with a clinically optimal threshold of 0.25. SHAP analysis identified needle gauge (18G, protective), poor vein quality (risk factor), and needle type (Y-type, risk factor) as the most influential predictors. The diurnal temperature difference was established as a novel, independent environmental risk factor. Conclusions The developed Logistic Regression model can effectively predict the occurrence of ICME, with a validated AUC of 0.82, indicating good discriminatory ability. It provides a practical and interpretable tool for risk stratification, highlighting key modifiable factors (e.g., needle choice and vein assessment) and a novel environmental cue (diurnal temperature difference), which can assist in guiding targeted preventive strategies to enhance patient safety during CECT. Clinical trial number. Not applicable.
Related
- Development and validation of an interpretable machine learning model for predicting in-hospital mortality among critically ill patients with liver cirrhosis
- Development and Validation of an Interpretable Machine‐Learning Model for Predicting Suboptimal Clinical Response After Sleeve Gastrectomy: A Multicenter Study
- An interpretable machine learning model for predicting intensive care unit mortality risk in elderly patients with concomitant cardiogenic shock and sepsis: Development and external validation using two large critical care databases
- Development and validation of interpretable machine learning models for predicting cumulative live birth rates in assisted reproductive technology: a SHAP-based analysis.
- Development and internal validation of an interpretable machine learning model for predicting vancomycin-induced nephrotoxicity in hospitalized children
BibTeX
@article{zhang2026development,
title = {{Development and validation of an interpretable machine learning model for predicting intravenous contrast media extravasation risk in CT imaging}},
author = {Wenyan Zhang and Qin Li and Xianrong Kong and Basen Li and Hanxiong Guan and Beibei Shao},
journal = {European Journal of Medical Research},
year = {2026},
month = jul,
doi = {10.1186/s40001-026-04925-5},
url = {https://doi.org/10.1186/s40001-026-04925-5},
note = {CC BY},
}
RIS
TY - JOUR
TI - Development and validation of an interpretable machine learning model for predicting intravenous contrast media extravasation risk in CT imaging
AU - Wenyan Zhang
AU - Qin Li
AU - Xianrong Kong
AU - Basen Li
AU - Hanxiong Guan
AU - Beibei Shao
PY - 2026
DA - 2026/07/01
JO - European Journal of Medical Research
DO - 10.1186/s40001-026-04925-5
UR - https://doi.org/10.1186/s40001-026-04925-5
AB - Objectives Intravenous contrast media extravasation (ICME) is a frequent and potentially severe complication during contrast-enhanced computed tomography (CECT). Although prevention of ICME is critical in radiology practice, valid and clinically practical prediction models remain limited. This study aimed to develop and validate an interpretable machine learning (ML) model for predicting the risk of ICME in patients undergoing CECT. Patients and methods This multicenter, retrospective case–control study analyzed 500 patients (252 ICME cases, 248 controls) from three institutions. Predictors encompassed patient demographics, clinical history, procedural details, venous access characteristics, and the diurnal temperature difference. From 22 initial candidates, 14 features were selected using least absolute shrinkage and selection operator (LASSO) regression. Five machine learning models—logistic regression, support vector machine (SVM), random forest, XGBoost, and K-nearest neighbors (KNN)—were trained and evaluated on an independent test set. The optimal model was assessed based on discrimination, calibration, and clinical utility, with its predictions explained using SHapley Additive exPlanations (SHAP). Results Among the five models, Logistic Regression demonstrated superior and the most generalizable performance on the test set, achieving an area under the receiver operating characteristic curve (AUC) of 0.82 (95% CI 0.72, 0.89), with a sensitivity of 78%, a precision of 75%, and an F1-score of 0.76. It also showed excellent calibration (Hosmer–Lemeshow p = 0.067). The other models' test AUCs were 0.81 (95% CI 0.71, 0.89) for SVM, 0.80 (95% CI 0.71, 0.88) for Random Forest, 0.79 (95% CI 0.68, 0.87) for XGBoost, and 0.74 (95% CI 0.64, 0.83) for KNN. Decision curve analysis confirmed the clinical utility of the Logistic Regression model across a broad range of threshold probabilities, with a clinically optimal threshold of 0.25. SHAP analysis identified needle gauge (18G, protective), poor vein quality (risk factor), and needle type (Y-type, risk factor) as the most influential predictors. The diurnal temperature difference was established as a novel, independent environmental risk factor. Conclusions The developed Logistic Regression model can effectively predict the occurrence of ICME, with a validated AUC of 0.82, indicating good discriminatory ability. It provides a practical and interpretable tool for risk stratification, highlighting key modifiable factors (e.g., needle choice and vein assessment) and a novel environmental cue (diurnal temperature difference), which can assist in guiding targeted preventive strategies to enhance patient safety during CECT. Clinical trial number. Not applicable.
SN - 2047-783X
ER -