Machine Learning to Predict In-Hospital Morbidity and Mortality after Traumatic Brain Injury

• Published in: Journal of neurotrauma Vol. 37; no. 1; pp. 202 - 210
• Main Authors: Matsuo, Kazuya, Aihara, Hideo, Nakai, Tomoaki, Morishita, Akitsugu, Tohma, Yoshiki, Kohmura, Eiji
• Format: Journal Article
• Language: English
• Published: Los Angeles, CA SAGE Publications 01.01.2020; Mary Ann Liebert, Inc
• Subjects: Adolescent; Adult; Aged; Aged, 80 and over; Algorithms; Artificial intelligence; Bayesian analysis; Blood pressure; Brain Injuries, Traumatic - mortality; C-reactive protein; Child; Coma; Computed tomography; Degradation products; Emergency medical care; Female; Fibrin; Fibrinogen; Hospital Mortality; Humans; Learning algorithms; Machine Learning; Male; Middle Aged; Morbidity; Mortality; Prognosis; Traumatic brain injury; Young Adult; traumatic brain injury; outcome predictor; machine learning; artificial intelligence
• ISSN: 0897-7151; 1557-9042; 1557-9042
• DOI: 10.1089/neu.2018.6276

Recently, successful predictions using machine learning (ML) algorithms have been reported in various fields. However, in traumatic brain injury (TBI) cohorts, few studies have examined modern ML algorithms. To develop a simple ML model for TBI outcome prediction, we conducted a performance comparison of nine algorithms: ridge regression, least absolute shrinkage and selection operator (LASSO) regression, random forest, gradient boosting, extra trees, decision tree, Gaussian naïve Bayes, multi-nomial naïve Bayes, and support vector machine. Fourteen feasible parameters were introduced in the ML models, including age, Glasgow Coma Scale (GCS), systolic blood pressure (SBP), abnormal pupillary response, major extracranial injury, computed tomography (CT) findings, and routinely collected laboratory values (glucose, C-reactive protein [CRP], and fibrin/fibrinogen degradation products [FDP]). Data from 232 patients with TBI were randomly divided into a training sample (80%) for hyperparameter tuning and validation sample (20%). The bootstrap method was used for validation. Random forest demonstrated the best performance for in-hospital poor outcome prediction and ridge regression for in-hospital mortality prediction: the mean statistical measures were 100% sensitivity, 72.3% specificity, 91.7% accuracy, and 0.895 area under the receiver operating characteristic curve (AUC); and 88.4% sensitivity, 88.2% specificity, 88.6% accuracy, and 0.875 AUC, respectively. Based on the feature selection method using the tree-based ensemble algorithm, age, Glasgow Coma Scale, fibrin/fibrinogen degradation products, and glucose were identified as the most important prognostic factors for poor outcome and mortality. Our results indicate the relatively good predictive performance of modern ML for TBI outcome. Further external validation is required for more heterogeneous samples to confirm our results.