Uncovering clinical significance of vital sign dynamics in critical care

Vital-sign time series of heart rate (HR) and blood pressure (BP) exhibit complex patterns of fluctuations, reflecting the underlying pathological and physiological states of patients. In this work, we adopt a switching vector auto-regressive process framework to learn a shared global library of “phenotypic” vital-sign dynamical behaviors from HR and BP time series of a patient cohort. Using HR and BP time series of over 450 adult ICU patients, we demonstrate that the fluctuation patterns in HR and BP are significantly correlated with several laboratory measurements routinely monitored in the ICU, and can potentially be used to reveal the underlying patho-physiological states of the patients. We demonstrate that the bivariate dynamics of HR/BP alone achieve similar performance in sepsis detection in comparison to the SAPS I scores, which use age and the most extreme values of 13 physiological variables. Further, combining the bivariate dynamics of HR/MAP time series and SAPS I provides a significantly more accurate (p=0.02) assessment of patients' risks in sepsis than using SAPS I alone (AUC 0.70 [0.63, 0.78] vs 0.78 [0.70, 0.85]), suggesting that the dynamics of the interaction between HR and BP may contain additional predictive value beyond that contained in the SAPS I scores.