Machine learning-based prediction of non-ionic iodinated contrast media-induced acute adverse reactions following contrast-enhanced CT.
Authors
Affiliations (7)
Affiliations (7)
- School of Clinical Medicine, Beijing Tsinghua Changgung Hospital, Tsinghua University, Changping District, Beijing 102218, China.
- Orthopedic and Sports Medicne Center, Beijing Tsinghua Changgung Hospital, School of Clinical Medicine, Tsinghua University, Beijing 102218, China.
- Department of Radiology and Nuclear Medicine, Xuanwu Hospital, Capital Medical University, No. 45 Changchun St, Xicheng District, Beijing 100053, China.
- Department of Obstetrics and Gynecology, Daping Hospital, Army Medical University, No. 10 Changjiang Road, Yuzhong District, Chongqing 400042, China.
- Department of Radiology and Nuclear Medicine, Xuanwu Hospital, Capital Medical University, No. 45 Changchun St, Xicheng District, Beijing 100053, China. Electronic address: [email protected].
- Department of Radiology, PLA Rocket Force Characteristic Medical Center, No. 16 Xinjiekou Outer Street, Beijing 100088, China; Department of Radiology, PLA 96603 Army Hospital, No. 65 Huaidong Road, Hecheng District, Huaihua City, Hunan Province 418099, China. Electronic address: [email protected].
- Department of Radiology, Daping Hospital, Army Medical University, No. 10 Changjiang Road, Yuzhong District, Chongqing 400042, China; Department of Radiology, the Third Affiliated Hospital of Chongqing Medical University, No. 1 Shuanghu Branch Road, Huixing Street, Yubei District, Chongqing 400042, China. Electronic address: [email protected].
Abstract
Iodinated contrast media (ICM) used in contrast-enhanced CT (CECT) examinations may induce acute adverse reactions (AAR) with different severity. To improve upon traditional linear risk assessments, this study aims to develop and validate a machine learning (ML)-based predictive model for ICM-AAR using routinely available clinical factors from a large-scale real-world cohort. Five ML models including Logistic Regression, Random Forest, XGBoost, CatBoost and LightGBM were trained and validated on a retrospective cohort of 332,090 patients who underwent CECT scans between 2014 and 2020. The final cohort was split into train, test and external validation datasets. Synthetic minority over-sampling technique (SMOTE) and under-sampling strategies were used to balance the data. Model training was performed using the GridsearchCV algorithm with 5-fold cross-validation. Model performance was evaluated using the area under the receiver operating characteristic curve (AUROC). The SHapley Additive exPlanations (SHAP) algorithm was performed to identify the most significant features. The train set included 132,102 patients (median age, 60 years [IQR, 49-69 years]; 72,666 males). The CatBoost model exhibited the best performance, with an AUROC of 0.6916 (95%CI = [0.6555, 0.7272]) in test set (n = 56,616; median age, 60 years [IQR, 49-69 years]; 31,400 males) and 0.6565 (95%CI = [0.6206, 0.6913]) in external validation set (n = 111,334; median age, 58 years [IQR, 49-68 years]; 61,447 males). Age, Injection rate, Type of contrast media, Injection dose, and Examination site were identified as the five most significant features by SHAP analysis. A ML model based on clinical factors was capable of predicting the occurrence of ICM-AAR, demonstrating improved ability in tackling extreme class imbalance. The model could be used to help clinical decision-making.