Outcome Prediction for Patients with Traumatic Brain Injury with Dynamic Features from Intracranial Pressure and Arterial Blood Pressure Signals: A Gaussian Process Approach

Abstract

Previous work has been demonstrated that tracking features describing the dynamic and time-varying patterns in brain monitoring signals provide additional predictive information beyond that derived from static features based on snapshot measurements. To achieve more accurate predictions of outcomes of patients with traumatic brain injury (TBI), we proposed a statistical framework to extract dynamic features from brain monitoring signals based on the framework of Gaussian processes (GPs). GPs provide an explicit probabilistic, nonparametric Bayesian approach to metric regression problems. This not only provides probabilistic predictions, but also gives the ability to cope with missing data and infer model parameters such as those that control the function’s shape, noise level and dynamics of the signal. Through experimental evaluation, we have demonstrated that dynamic features extracted from GPs provide additional predictive information in addition to the features based on the pressure reactivity index (PRx). Significant improvements in patient outcome prediction were achieved by combining GP-based and PRx-based dynamic features. In particular, compared with the a baseline PRx-based model, the combined model achieved over 30 % improvement in prediction accuracy and sensitivity and over 20 % improvement in specificity and the area under the receiver operating characteristic curve.

Intracranial Pressure Monitoring Signals After Traumatic Brain Injury: A Narrative Overview and Conceptual Data Science Framework

“…The PRx threshold had the highest performance to predict the GOS (AUC 0.81, 95% CI 0.74-0.87) over ICP > 20 mmHg (AUC 0.75, 95% CI 0.68-0.81) and ICP > 25 mmHg (AUC 0.77, 95% CI 0.70-0.83) (85). Others have fit logistic regression models to the features extracted from Gaussian process-based models and PRx to predict mortality with model performance of 0.74 accuracy, 0.65 specificity, and 0.83 sensitivity, although this modeling was limited in its clinical utility (84). Of note, most studies do not consider the effects of clinical decisions to withdraw life-sustaining therapies, which may result in differential classification bias.…”

A review of big data applications of physiological signal data

“…An approach often utilized for providing such alerts is that of novelty detection (Pimentel et al 2014) which is a machine learning approach used in situations where a lot of data is available of the Bnormal^state and very few of the Babnormal.^Often modeled as a one-class classification approach, novelty detection techniques have been applied in the clinical context for providing early warning of deterioration in the emergency department (Wilson et al 2016), the ICU (Ghassemi et al 2015), and for post-operative patients (Clifton et al 2014;Pimentel et al 2013). Gaussian processes have also been extensively used for dealing with noisy and unreliable physiological data (Dunitz et al 2015) including in the ICU (Pimentel et al 2016). As reviewed in Leff and Yang (2015), in addition to applications in the early warning of deterioration in continuous monitoring, medical decision support systems have also been proposed for many other applications including to assess and improve protocol adherence (Klann et al 2013), for medication reminders (Nair et al 2010), to improve screening (Wagholikar et al 2012), and to predict hospital readmission (Futoma et al 2015).…”