Developing practical machine learning survival models to identify high-risk patients for in-hospital mortality following traumatic brain injury

Machine learning (ML) offers precise predictions and could improve patient care, potentially replacing traditional scoring systems. A retrospective study at Emtiaz Hospital analyzed 3,180 traumatic brain injury (TBI) patients. Nineteen variables were assessed using ML algorithms to predict outcomes. Data preparation addressed missing values and balancing methods corrected imbalances. Model building involved training-test splits, survival analysis, and ML algorithms like Random Survival Forest (RSF) and Gradient Boosting. Feature importance was examined, with patient risk stratification guiding survival analysis. The best-performing model, RSF with ROS resampling, achieved the highest mean AUC of 0.80, the lowest IBS of 0.11, and IPCW c-index of 0.79, maintaining strong predictive ability over time. Top predictors for in-hospital mortality included age, GCS, pupil condition, PTT, IPH, and Rotterdam score, with high variations in predictive abilities over time. A risk stratification cut-off value of 63.34 separated patients into low and high-risk categories, with Kaplan-Meier curves showing significant survival differences. Our high-performing predictive model, built on first-day features, enables time-dependent risk assessment for tailored interventions and monitoring. Our study highlights the feasibility of AI tools in clinical settings, offering superior predictive accuracy and enhancing patient care for TBI cases.