Learning Cyber-Physical Models of Resuscitation

Abstract

Ability to predict outcomes and forecast trajectories of recovery from resuscitated intensive care patients could guide treatment decisions and improve outcomes of care in both clinical and field settings. We develop a machine learning driven cyber-physical model to provide such predictive capabilities by leveraging arterial blood pressure (ABP) waveforms, one of the routinely collected vital signs. A cohort of 51 Yorkshire pigs was subjected to induced slow rate hemorrhage followed by fluid resuscitation. To represent physics of the arterial system and emulate blood pressure dynamics, we combine a two-element Windkessel model with an Unscented Kalman Filter (UKF) to track the instantaneously estimated Windkessel parameters over time. As the arterial pressure waveform exponentially decays during diastole after each pump, we use UKF-tracked Windkessel parameter estimates to identify time windows of ABP waveforms taken from other subjects in the cohort to reconstruct the shapes of the test subject’s ABP signal and its moving average. We allow UKF covariance to temporarily increase to account for the effects of treatment such as administering norepinephrine. When evaluated under leave-one-subject-out cross-validation protocol, the model stays within 14+/-5% (mean+/-standard deviation) of mean absolute percentage error when reconstructing the current 250Hz ABP waveforms, and 19+/-6%, 24+/-6%, and 25+/-6% when forecasting at 5, 15 and 30 minute horizons, respectively. Our results demonstrate feasibility of using cyber-physical modeling of hemodynamic waveform data to predict trajectories of resuscitation and therefore timely inform treatment of hemorrhagic patients in both clinical and prolonged field care settings. We also provide a few thoughts for future work, including potential improvements attainable by calibrating neighbor selections and model predictions to the individual subject baselines to improve handling heterogeneity of the subjects, on-line tracking of model performance to estimate confidence in its predictions in real-time, and forecasting eventual outcomes of resuscitation as well as time-to-recovery for those subjects who are likely to recover, to name a few such ideas.