130 Prediction of In-Hospital Mortality After Acute Traumatic Brain Injury managed by Tranexamic Acid: Secondary Exploration of CRASH-3 Trial via Automated Machine Learning. (April 2023)

• Record Type: Journal Article
• Authors: Haider, Ali; Ullah, Tauseef; Khan, Inayat Ullah; Zehra, Saiqa; Khawaja, Hashir Fahim; Mehmood Abbas, Syed Mohammad; Mohsin, Rabia; Khan, Ayesha Khalid; Shah, Ali Haider
• Is Part Of: Neurosurgery. Volume 69(2023)Supplement 1
• Page Start: 32
• Page End: 33
• Publication Date: 2023-04
• Subjects: Nervous system -- Surgery -- Periodicals; 617.48005
• DOI: 10.1227/neu.0000000000002375_130
• Languages: English
• ISSNs: 0148-396X

Abstract

Abstract : INTRODUCTION: Acute traumatic brain injury (acTBI) has significant morbidity and mortality (M&M) associated with it— increased intracranial pressure (ICP) among the attributed causative mechanisms. Tranexamic acid, therefore, is adopted to decrease ICP and subsequently the acTBI-associated M&M. METHODS: The study population comprised 12, 743 patients from 175 hospitals in 29 countries experiencing acTBI randomly assigned to receive tranexamic acid or placebo. The current state of the art (SOTA) for automated Machine Learning (aML) was adopted to develop predictive models with the incorporation of the ensemble approach. Logloss and Macro weighted average Area Under the Receiver Operating Curve (mWA-AUROC) assessed the predictive discriminative ability of the developed models. RESULTS: An ensemble of Extreme Gradiant Boosting Machine, Light Gradiant Boosting Machine, CatBoost, Random Forest and Neural Network predicted IHM along with the cause of death in acTBI patients randomly assigned to Tranexamic acid management with an mWA-AUROC of 0.94, an accuracy of 85.8% and a logloss of 0.37. Motor and verbal responses along with pupillary reactions were recognized as the most influential predictors for IHM in such patients with maintenance dose of Tranexamic acid imparting moderate to minuscule influence. Boxplot indicates a relative decrease in logloss by adopting the ensemble approach. CONCLUSIONS: Our novel approach to developing predictive prognosticative models for acTBI patients managed by Tranexamic acid by exploring automated machine learning is the very first attempt of its nature. Adoption of the current SOTA for aML provides optimal predictions which, when incorporated into the respective management and prognosticative protocols, shall translate into a decrease in the M&M associated with TBI by assisting in risk stratification and complication triaging.