Prof. Vladimir Zadorozhny of DINS is Co-PI on a recently funded NIH grant titled “Optimizing Recovery Prediction After Cardiac Arrest”. The project PI is Dr. Jonathan Elmer, from the Department of Critical Care Medicine at Pitt. This is a collaborative project between the University of Pittsburgh and Carnegie Mellon University. Dr. Zadorozhny's group is responsible for developing and implementing methods and techniques for continuous accumulation, storage, aggregation and interpretation of large-scale medical data, including highly multivariate health time-series, to facilitate efficient medical decision-making. More information follows:
Project Summary
Predicting recovery from anoxic brain injury and coma after cardiac arrest is challenging. Although patients resuscitated from cardiac arrest are intensively monitored in critical care units, clinicians use only a tiny subset of available data to predict potential for recovery, making neurological prognostication both slow and imprecise. This is a specific example of a ubiquitous problem in modern medicine: routine clinical monitoring generates vast quantities of rich information, but tools to transform these data to useful knowledge are lacking. This project will leverage expertise in post-arrest critical care, information science, statistical modeling and machine learning to make a system that rapidly delivers actionable prognostic knowledge. We have cleaned, organized and aggregated a large, highly multivariate time series database with physiological and clinical information with over 170,000 hours of quantitative electroencephalographic (EEG) features for >1,850 post-arrest patients. We will refine and optimize analytical tools that predict recovery in this patient population more rapidly and accurately than clinical experts. We will use innovative approaches to minimize risk of bias during training of models introduced by outcome labels created by fallible human providers.
In Aim 1 of this proposal, we will use novel approaches to create informative and interpretable features from heterogeneous clinical data including EEG waveforms, vital signs, medications and laboratory test results. We will use deep learning to identify interpretable and parsimonious sets of these features that predict outcome. We will train, test and compare the performance of multiple analytical tools. In Aim 2, we will prospectively compare the best performing model(s) against a panel of expert clinicians. Models that confidently identify patients with near-zero prospect of recovery with greater sensitivity or faster than expert clinicians can serve as decision support systems. Improving the speed and accuracy of post-arrest prognostication will save lives, allow appropriate resources to be directed to patients who are likely to benefit, avoid long and difficult care for patients who cannot recover, and spare families the agony of uncertainty.