Development and Internal Validation of an Automated CT-Based Model for Early Mortality Prediction in Traumatic Brain Injury.
Authors
Affiliations (4)
Affiliations (4)
- Department of AI-based Convergence, Dankook University, Yongin, Gyeonggi-do, Republic of Korea.
- Department of Neurosurgery, Dankook University College of Medicine, Cheonan, Republic of Korea.
- Department of Computing, Gachon University, Seongnam, Republic of Korea. Electronic address: [email protected].
- Department of Computer Engineering, Dankook University, Yongin, Gyeonggi-do, Republic of Korea. Electronic address: [email protected].
Abstract
Traumatic brain injury outcomes are strongly influenced by intracranial hemorrhage burden, midline shift, and secondary hypoxic injury. Although traditional prognostic tools such as the Marshall CT score, Rotterdam CT score, and IMPACT model are widely used to estimate outcomes, they require expert manual interpretation and may be difficult to apply consistently when trained specialists are not immediately available. Therefore, we aimed to develop a fully automated multimodal framework for early mortality prediction in TBI, in which CT-derived imaging biomarkers were extracted using deep-learning models and integrated with clinical variables using a shallow MLP-based predictor, and to compare its prognostic performance with that of conventional TBI scoring systems. Hematoma volume and brain midline shift (MLS) were automatically extracted from CT images using deep-learning-based segmentation and landmark detection models. The resulting imaging biomarkers were integrated with demographic and clinical variables and used to train a shallow multilayer perceptron-based multimodal predictor for early mortality prediction. Model interpretability was assessed using an exploratory SHAP analysis to examine how each input feature contributed to the model predictions. Mortality prediction models using automated imaging biomarkers achieved discriminative performance similar to that of conventional scoring systems, and further performance improvement was observed when these biomarkers were integrated with additional clinical variables. Automatically derived hematoma volume and MLS provided prognostic performance comparable to that of conventional TBI scoring systems, and performance improved further when these imaging biomarkers were combined with clinical variables. These findings support the feasibility of a fully automated and reproducible framework for early mortality risk stratification in patients with traumatic brain injury.