A Transformer-Based Risk Prediction Model for Contrast Media Extravasation During High-Pressure Injection: A Dual-Center Study with External Validation
Related Experiment Videos
Xi Zhu1,2, Zhuqing Bao3, Zhen Hu4
1School of Biological Sciences and Medical Engineering, Southeast University, Nanjing, Jiangsu, People's Republic of China.
Purpose:
To assess whether a Transformer-based deep learning model can predict contrast media extravasation (CME) risk before high-pressure intravenous iodinated contrast injection, and to provide an automated screening tool for high-risk patients before contrast-enhanced CT.
Patients And Methods:
We retrospectively analysed 35,537 patients who underwent contrast-enhanced CT between November 2023 and June 2025 at Northern Jiangsu People's Hospital (internal cohort, n = 26,126; 0.88% CME-positive) and its South Branch (external validation, n = 9,411; 0.99% CME-positive), an overall CME-positive rate of 0.91%. Sixteen demographic, clinical, injection and nursing variables were used. A Transformer encoder model (three stacked encoder blocks) was trained with a batch-adaptive hard negative sample-corrected Focal Loss for extreme class imbalance, with class weights recalculated per mini-batch and an extra penalty on hard negative samples; perturbation-based feature attribution was also performed. Internal five-fold cross-validation was followed by independent external validation. Metrics included accuracy, recall, specificity, positive and negative predictive value, F1-score, Cohen's Kappa, Matthews correlation coefficient (MCC) and area under the ROC curve (AUC).
Results:
Internal five-fold cross-validation gave an accuracy of 93.07% ± 0.76%, recall of 66.80% ± 2.16%, specificity of 93.31% ± 0.79% and AUC 0.907; external validation gave 92.71%, 63.44%, 93.00% and 0.862. Cohen's Kappa was 0.12-0.15 (internal mean 0.13) and MCC 0.21-0.23 (internal mean 0.22). Puncture site, contrast agent viscosity, total injection volume and venous thrombosis were high-contribution features (>10%); injection rate, diabetes, sex, age, BMI, VIA scale, malignant tumor history and patient source contributed 1-10%.
Conclusion:
The model showed acceptable discriminative ability and generalizability across two centers, but remains a retrospective research prototype: because a rare-event model with high specificity still generates many false-positive alerts, its outputs require local recalibration and the alert threshold has not been clinically optimized. Prospective multicenter validation is therefore needed before clinical deployment.
