Prediction of Early TBI Mortality Using a Machine Learning Approach in a LMIC Population

To develop and validate practical prognostic models for death at 14 days and for death or severe disability six months after traumatic brain injury. Multivariable logistic regression to select variables that were independently associated with two patient outcomes. Two models designed: "basic" model (demographic and clinical variables only) and "CT" model (basic model plus results of computed tomography). The models were subsequently developed for high and low-middle income countries separately. Medical Research Council (MRC) CRASH Trial. 10,008 patients with traumatic brain injury. Models externally validated in a cohort of 8509. The basic model included four predictors: age, Glasgow coma scale, pupil reactivity, and the presence of major extracranial injury. The CT model also included the presence of petechial haemorrhages, obliteration of the third ventricle or basal cisterns, subarachnoid bleeding, midline shift, and non-evacuated haematoma. In the derivation sample the models showed excellent discrimination (C statistic above 0.80). The models showed good calibration graphically. The Hosmer-Lemeshow test also indicated good calibration, except for the CT model in low-middle income countries. External validation for unfavourable outcome at six months in high income countries showed that basic and CT models had good discrimination (C statistic 0.77 for both models) but poorer calibration. Simple prognostic models can be used to obtain valid predictions of relevant outcomes in patients with traumatic brain injury.

Intracranial-pressure monitoring is considered the standard of care for severe traumatic brain injury and is used frequently, but the efficacy of treatment based on monitoring in improving the outcome has not been rigorously assessed. We conducted a multicenter, controlled trial in which 324 patients 13 years of age or older who had severe traumatic brain injury and were being treated in intensive care units (ICUs) in Bolivia or Ecuador were randomly assigned to one of two specific protocols: guidelines-based management in which a protocol for monitoring intraparenchymal intracranial pressure was used (pressure-monitoring group) or a protocol in which treatment was based on imaging and clinical examination (imaging-clinical examination group). The primary outcome was a composite of survival time, impaired consciousness, and functional status at 3 months and 6 months and neuropsychological status at 6 months; neuropsychological status was assessed by an examiner who was unaware of protocol assignment. This composite measure was based on performance across 21 measures of functional and cognitive status and calculated as a percentile (with 0 indicating the worst performance, and 100 the best performance). There was no significant between-group difference in the primary outcome, a composite measure based on percentile performance across 21 measures of functional and cognitive status (score, 56 in the pressure-monitoring group vs. 53 in the imaging-clinical examination group; P=0.49). Six-month mortality was 39% in the pressure-monitoring group and 41% in the imaging-clinical examination group (P=0.60). The median length of stay in the ICU was similar in the two groups (12 days in the pressure-monitoring group and 9 days in the imaging-clinical examination group; P=0.25), although the number of days of brain-specific treatments (e.g., administration of hyperosmolar fluids and the use of hyperventilation) in the ICU was higher in the imaging-clinical examination group than in the pressure-monitoring group (4.8 vs. 3.4, P=0.002). The distribution of serious adverse events was similar in the two groups. For patients with severe traumatic brain injury, care focused on maintaining monitored intracranial pressure at 20 mm Hg or less was not shown to be superior to care based on imaging and clinical examination. (Funded by the National Institutes of Health and others; ClinicalTrials.gov number, NCT01068522.).

Traumatic brain injury (TBI) is a heterogeneous condition that encompasses a broad spectrum of disorders. Outcome can be highly variable, particularly in more severely injured patients. Despite the association of many variables with outcome, prognostic predictions are notoriously difficult to make. Multivariable analysis has identified age, clinical severity, CT abnormalities, systemic insults (hypoxia and hypotension), and laboratory variables as relevant factors to include in models to predict outcome in individual patients. Advances in statistical modelling and the availability of large datasets have facilitated the development of prognostic models that have greater performance and generalisability. Two prediction models are currently available, both of which have been developed on large datasets with state-of-the-art methods, and offer new opportunities. We see great potential for their use in clinical practice, research, and policy making, as well as for assessment of the quality of health-care delivery. Continued development, refinement, and validation is advocated, together with assessment of the clinical impact of prediction models, including treatment response. Copyright 2010 Elsevier Ltd. All rights reserved.