Machine Learning-Driven Prognostication in Traumatic Subdural Hematoma: Development of a Predictive Web Application

BACKGROUND AND OBJECTIVES: Our focus was on creating an array of machine learning (ML) models to predict unfavorable in-hospital outcomes after acute traumatic subdural hematoma (atSDH). Our subsequent aim was to deploy these models in an accessible web application, showcasing their practical value. METHODS: Data from the American College of Surgeons Trauma Quality Program database were used to identify patients with atSDH. In-hospital mortality was the primary outcome of interest. Secondary outcomes were (1) nonhome discharges, (2) prolonged length of stay (LOS), (3) prolonged length of stay in the intensive care unit (ICU-LOS), and (4) major complications. Feature selection was performed with least absolute shrinkage and selection operator algorithm. Five ML algorithms, including TabPFN, TabNET, XGBoost, LightGBM, and Random Forest, were used along with the Optuna optimization library for hyperparameter tuning. RESULTS: There were 104 055 patients included in the analysis for the outcome mortality, 82 988 for the outcome nonhome discharges, 104 207 for the outcome prolonged LOS, 62 543 for the outcome prolonged ICU-LOS, and 100 241 for the outcome major complications. The models with the highest area under receiver operating characteristic curve (AUROC) values included TabPFN for mortality and major complications, and LightGBM for nonhome discharges, prolonged LOS, and ICU-LOS. The TabPFN model for the primary outcome of our study, in-hospital mortality, showed an AUROC of 0.934. The models with the highest AUROC values were integrated into an application to predict the outcomes of interest. CONCLUSION: Our findings show that ML tools aid in predicting various outcomes for patients with atSDH. We developed a web application that has the potential to integrate the developed models into clinical practice.