Dynamic and Explainable Mortality Risk Prediction for TBI Patients in the ICU

Dynamic mortality risk prediction in the intensive care unit (ICU) is crucial for supporting clinicians’ decision-making, specifically in traumatic brain injury (TBI) patients. We aim to develop and evaluate a dynamic deep learning (DL) framework that can provide hourly updates of 30-day mortality risk prediction for TBI patients following ICU admission. Using demographics and time-series physiological data, a recurrent neural network-based model was trained on data from 135 TBI patients admitted to the Gold Coast University Hospital (GCUH) in Australia. Model’s performance was evaluated utilizing the area under the receiver operating characteristics (AUC), Matthews correlation coefficient (MCC), accuracy, and other metrics, performed calibration and decision curve analysis to interpret the model’s output and determine its clinical usefulness. The Shapley additive explanation algorithm was utilized to clarify the contribution of features to the predictions. The proposed method showed predictive performance on the cross-validation dataset that improved over time: MCC 0.24 and AUC 0.713 for the prediction at 24 h after admission, 0.451 and 0.756 at 72 h, 0.519 and 0.803 at 120 h, and 0.748 and 0.946 before twelve hours to the outcome (either death or discharge), respectively. The model was further tested with a holdout test dataset with 34 TBI patients, achieving an average prediction accuracy of 0.851, AUC of 0.632, and MCC of 0.403, respectively, in the first 24-h interval. The proposed model demonstrates proof of principle with explainable results in predicting mortality risk, encouraging further development and validation in a clinical setting.