Skip to Content

Machine Learning of Physiological Variables to Predict Diagnose and Treat Cardiorespiratory Instability

PI: Michael Pinsky, MD

Co-Is: Gilles Clermont, MD, Artur Dubrawski, PhD

Funding: R01GM117622  NIH/NIGMS (4/1/2016 to 3/31/2020)

If one could accurately predict who, when and why patients develop shock then effective preemptive treatments could be given to improve outcome and more effectively use healthcare resources. But signs of shock often occur late once organ injury is already present. We have shown that an integrated monitoring system (IMS) when coupled to a care plan improves outcome in step down unit patients. We also showed that by advanced HR variability analysis (sample entropy), identifies at-risk patients within 2 minutes and if monitored for 5 minutes predicts well those SDU patients who would deteriorate or remain stable over the next 48 hr. In collaboration with the Auton Lab at CMU, we propose to first develop multivariable models through machine learning (ML) data-driven classification techniques [multivariate regression analysis, Fourier analysis, principal component analysis, artificial neural networks, random forest classification, etc.] to parsimoniously predict cardiovascular insufficiency. We will initially use our existing annotated high fidelity waveform MIMIC II clinical data set, (n=4200 patients) as to timing and types of instability. Then we will use our prospective ICU data set using our high-density data collection and processing platform (Bernoulli) to collect all clinically relevant data from three institutions: Univ. of Pittsburgh (PITT), UC Irvine and UC San Diego. The primary rule learning and model development will be done at PITT and validated in the UC systems. We will simultaneously test decision support systems driven by these smart alerts and functional hemodynamic monitoring approaches in two human simulation environments (PITT & UC Irvine), developing predictive models, smart alarms and treatment tools in an iterative fashion. ML modeling of our clinically-relevant porcine model of severe hemorrhagic shock allowed us to characterize the responses to hemorrhage, hypovolemia to decompensation and resuscitation, and predict which animals would or would not collapse during severe hypovolemia. We can identify occult bleeding 10-15 minutes earlier than traditional monitoring methods using featurization of routine hemodynamic monitoring variables. What is not known is the minimal data set relative to the number and type of independent measures, the sample frequency and the lead time which are necessary to create a robust algorithm that will: 1) predict impending shock, 2) select the most effective treatments, 3) monitor the response to those treatments, and 4) determine when resuscitation has effectively restored physiologic wellness and can be stopped. We refer to this concept as Monitoring Parsimony. Also, it is not known if the addition of more monitoring parameters or the lead time before which insufficiency is recognized will improve the ability to achieve the above four targets with the best sensitivity and specificity once an alert is raised. We will test a model-driven graphic user interface tools in our simulation centers. We envision a basic monitoring surveillance that identifies those patients most likely to become unstable for focused monitoring and targeted treatments to deliver highly personalized medical care.

Public Health Relevance Statement: If one could accurately predict who, when and why patients develop shock then effective preemptive treatments could be given to improve outcome and more effectively use healthcare resources. But signs of shock often occur late once organ injury is already present. The purpose of this study is to first develop multivariable models through data-driven classification techniques to parsimoniously predict cardiovascular insufficiency, etiology and response to treatment. We will do this first in our existing MIMIC II clinical data sets of 4200 ICU patients as to timing and types of instability. Then we will prospectively collect real time high- density data on patients admitted to our trauma intensive care units of University of Pittsburgh, UC Irvine and UC San Diego. We will create and test in simulators of ICU care bedside user interfaces to drive recognition and treatment algorithms based on these models in all three medical centers.