Prediction of Early TBI Mortality Using a Machine Learning Approach in a LMIC Population
Robson Luís Oliveira de Amorim, Louise Makarem Oliveira, Luíz Marcelo Sá Malbouisson, Márcia Mitie Nagumo, Marcela Simoes, Leandro Miranda, Edson Bor‐Seng‐Shu, André Beer‐Furlan, Almir Ferreira de Andrade, Andrés M. Rubiano, Manoel Jacobsen Teixeira, Angelos G. Kolias, Wellingson Silva Paiva
Abstract
Background: In a time where severe traumatic brain injury (TBI) is decreasing in developed countries and increasing in low-to-middle-income countries (LMIC’s), it’s important to understand the behavior of predictive variables in such population. There are few previous attempts to generate prediction models for TBI outcomes from local data in developing countries. Our study aim is to design and compare a series of predictive models for mortality on a new cohort in TBI patients in Brazil using Machine Learning. Methods: A prospective registry was set in São Paulo, Brazil, enrolling all patients with a diagnosis of TBI requiring admission to the intensive care unit. We evaluated the following predictors: gender, age, pupil reactivity at admission, Glasgow Coma Scale (GCS), presence of hypoxia and hypotension, computed tomography (CT) findings, trauma severity score, and laboratory results. Results: Overall mortality at 14 days was 22.8%. Models had a high prediction performance, with the best prediction for overall mortality achieved through Naive Bayes (area under the curve = 0.906). The most significant predictors were the GCS at admission and prehospital GCS, age, and pupil reaction. When predicting the length of stay at the intensive care unit, the Conditional Inference Tree model had the best performance (root-mean-square error = 1.011), with the most important variable across all models being the GCS at scene. Conclusions: Models for early mortality and hospital length of stay using Machine Learning can achieve high performance when based on registry data even in developing countries, with potential to enhance quality of care.