Early prediction of progressive cerebral contusion using a deep transfer learning–enhanced multimodal nomogram
Wanqin Yang, Shanshan Qian, Ping ZhaoAbstract
Objectives
We developed a multimodal fusion model combining clinical data and deep transfer learning for early progressive cerebral contusion (PCC) prediction, providing precise clinical support for treatment decisions.
Methods
Using a single-center retrospective cohort design, we analyzed 196 cerebral contusion patients between January 2022 and June 2024. PCC was characterized by a contusion volume increase of at least 30 % on CT scans within 24 h. Patients were categorized into a progression group (n=98) and a non-progression group (n=98). The dataset was split into a training coh59 participants, maintaining a 7:3 ratort of 137 participants and a validation cohort of io. A nomogram was developed by combining ResNet-50-based deep transfer learning features with clinical variables. Model performance was assessed through ROC curves, calibration plots, and decision curve analysis, while Grad-CAM was used to evaluate interpretability.
Results
The integrated nomogram demonstrated superior performance with AUC values of 0.999 (95 % CI: 0.998–1.000) in the training cohort and 0.972 (95 % CI: 0.939–1.000) in the validation cohort, surpassing the standalone DTL and clinical models. Grad-CAM demonstrated accurate lesion localization.
Conclusions
The multimodal fusion model integrating DTL and clinical features shows excellent predictive performance and significant clinical value in early PCC prediction.