Machine learning of physiological variables to predict diagnose and treat cardiorespiratory instability

University Of Pittsburgh At Pittsburgh, Pittsburgh PA

Investigators

Linked publications & trials

Abstract

Project Summary/Abstract: If one could accurately predict who, when and why patients develop cardiorespiratory instability (CRI), then effective preemptive treatments could be given to improve outcome and better use care resources. However, CRI is often unrecognized until it is well established and patients are more refractory to treatment, or progressed to organ injury. We have shown that an integrated monitoring system alert obtained from continuous noninvasively acquired monitoring parameters and coupled to a care algorithm improved step-down unit (SDU) patient outcomes. We also showed that advanced HR variability analysis (sample entropy) identified SDU patients at CRI risk within 2 minutes, and if monitored for 5 minutes differentiated between patients who would develop CRI or remain stable over the next 48 hours. We also applied machine learning (ML) modeling to our clinically-relevant porcine model of hemorrhagic shock to characterize responses to hypovolemia, hemorrhage, and resuscitation, predict which animals would or would not collapse during hypovolemia, and identify occult bleeding 5 minutes earlier than with traditional monitoring. We now propose to apply our work to vulnerable and invasively monitored ICU patients. We will develop multivariable models through ML data-driven classification techniques such as regression, Fourier and principal component analysis, artificial neural networks, random forest classification, etc. as well as more novel approaches (temporal rule learning developed by our team; Bayesian Aggregation) to predict CRI in ICU patients. We will first use our existing annotated high fidelity waveform MIMIC II clinical data set (4200 patients) to develop predictive models and differential signatures for various CRI drivers. We will also use our high-density data collection and processing platform (Bernoulli) to prospectively collect data from ICUs in three institutions: Univ. Pittsburgh (PITT), Univ. California (UC) Irvine and UC San Diego (initial algorithm development conducted at PITT and validated in the UC systems). We will identify the number and type of independent measures, sampling frequency, and lead time necessary to create robust algorithms to: 1) predict impending CRI, 2) select the most effective treatments, 3) monitor treatment response, and 4) determine when treatment has restored physiologic stability and can be stopped. We will also determine the smallest number and types of parameters coupled to the longest CRI lead time to achieve the above four targets with the best sensitivity and specificity (a concept we call Monitoring Parsimony).We will simultaneously iteratively design and test a graphical user interface (GUI) and clinical decision support system (CDSS) driven by these parsimoniously derived predictive smart alerts and functional hemodynamic monitoring treatment approaches in two human simulation environments (PITT & UC Irvine).We envision a basic monitoring surveillance that identifies patients most likely to develop CRI to apply focused clinician attention and targeted treatments to deliver highly personalized medical care.