Data-Driven Phenotyping of Severe Traumatic Brain Injury

项目总结

Project Summary This predoctoral fellowship will provide the applicant (Hayley Falk), a doctoral candidate in the Department of Computational Medicine & Bioinformatics at the University of Michigan, with the skills necessary to become an independent research investigator with expertise in novel applications of machine learning for TBI. The limited accuracy of current models for early prediction of GCS 3-8 TBI (commonly referred to as severe TBI) outcomes (prognostic models) is a major barrier to improving the clinical care of patients with GCS 3-8 TBI. Less than 20% of patients with GCS 3-8 TBI experience a good neurologic recovery and currently there are no therapeutic agents that improve long-term outcomes. In GCS 3-8 TBI clinical trials of promising therapeutic agents, a favorable outcome is typically defined as a better outcome than would be expected, taking into account the predicted prognosis for each individual patient. Therefore, accurate estimation of predicted prognosis is critical to assessing the efficacy of novel therapeutic agents. The leading prognostic models for GCS 3-8 TBI, IMPACT (International Mission for Prognosis and Analysis of Clinical Trials in TBI) and CRASH (Corticosteroid Randomization After Significant Head Injury), have undergone extensive external validation, however, the discriminative accuracy is highly cohort dependent with AUCs as low as 0.60 in some patient groups. The two major limitations of the IMPACT and CRASH models include one-time measurements of clinical predictor variables and regression-based methods, which are not designed to handle complex, multidimensional datasets. Our objective is to derive a dynamic prognostic model which provides updated outcome predictions as new data becomes available. We will then develop a clustering algorithm to identify physiologically distinct subtypes (clusters) of GCS 3-8 TBI derived from continuous, high frequency data streams. Our proposed goals will be achieved by the following specific aims: 1) we will derive a dynamic prognostic model using a RNN (recurrent neural network)-based framework and data collected during the first two weeks post-injury from BOOST-2 (Brain Oxygen Optimization in Severe Traumatic Brain Injury: Phase 2) which provides updated 6-month outcome predictions every 24 hours; and 2) using time series hierarchical clustering and continuous measures of physiologic parameters collected from subjects enrolled in BOOST-2 during the first two weeks post-injury, we will identify distinct subtypes of GCS 3-8 TBI and examine the association between subtype and 6-month outcome. In Aim 1, we hypothesize that our dynamic prognostic model derived from time-varying data will have a higher discriminative accuracy (AUC) than a static prognostic model (similar to IMPACT and CRASH) derived using single timepoint data collected on the day of injury. In Aim 2, we hypothesize that subtypes of GCS 3-8 TBI characterized by continuous physiologic parameters such as increasing ICP (intracranial pressure), decreasing CPP (cerebral perfusion pressure), and decreasing PbtO2 (brain tissue oxygenation) will be associated with poor 6-month outcomes.