Signal Information Prediction of Mortality Identifies Unique Patient Subsets after Severe Traumatic Brain Injury: A Decision-Tree Analysis Approach.

Nonlinear physiological signal features that reveal information content and causal flow have recently been shown to be predictors of mortality after severe traumatic brain injury (TBI). The extent to which these features interact together, and with traditional measures to describe patients in a clinically meaningful way remains unclear. In this study, we incorporated basic demographics (age and initial Glasgow Coma Scale [GCS]) with linear and non-linear signal information based features (approximate entropy [ApEn], and multivariate conditional Granger causality [GC]) to evaluate their relative contributions to mortality using cardio-cerebral monitoring data from 171 severe TBI patients admitted to a single neurocritical care center over a 10 year period. Beyond linear modelling, we employed a decision tree analysis approach to define a predictive hierarchy of features. We found ApEn (p = 0.009) and GC (p = 0.004) based features to be independent predictors of mortality at a time when mean intracranial pressure (ICP) was not. Our combined model with both signal information-based features performed the strongest (area under curve = 0.86 vs. 0.77 for linear features only). Although low "intracranial" complexity (ApEn-ICP) outranked both age and GCS as crucial drivers of mortality (fivefold increase in mortality where ApEn-ICP <1.56, 36.2% vs. 7.8%), decision tree analysis revealed clear subsets of patient populations using all three predictors. Patients with lower ApEn-ICP who were >60 years of age died, whereas those with higher ApEn-ICP and GCS ≥5 all survived. Yet, even with low initial intracranial complexity, as long as patients maintained robust GC and "extracranial" complexity (ApEn of mean arterial pressure), they all survived. Incorporating traditional linear and novel, non-linear signal information features, particularly in a framework such as decision trees, may provide better insight into "health" status. However, caution is required when interpreting these results in a clinical setting prior to external validation.