Cardiovascular oscillations at the bedside: early diagnosis of neonatal sepsis using heart rate characteristics monitoring

Abstract

We have applied principles of statistical signal processing and nonlinear dynamics to analyze heart rate time series from premature newborn infants in order to assist in the early diagnosis of sepsis, a common and potentially deadly bacterial infection of the bloodstream. We began with the observation of reduced variability and transient decelerations in heart rate interval time series for hours up to days prior to clinical signs of illness. We find that measurements of standard deviation, sample asymmetry and sample entropy are highly related to imminent clinical illness. We developed multivariable statistical predictive models, and an interface to display the real-time results to clinicians. Using this approach, we have observed numerous cases in which incipient neonatal sepsis was diagnosed and treated without any clinical illness at all. This review focuses on the mathematical and statistical time series approaches used to detect these abnormal heart rate characteristics and present predictive monitoring information to the clinician.

Figure 1

Neonatal HR time series. (A) Normal, (B) Reduced variability, (C) Reduced variability and a storm of transient decelerations, (D) Excerpt from (C). All records are from the same infant, and B to D occurred in the hours prior to clinical signs of sepsis.

Figure 2

Time series and histograms of neonatal heart rates.

Figure 3

Sample asymmetry weights deviations from the center of a distribution.

Figure 4

Schema for entropy estimation in HR timer series

Figure 5

Examples of test data sets. (A) random numbers within a possible range of a RR intervals (mean 400, S.D. 25). (B),(C),(D) RR intervals from a same infant when the infant was healthy (B), and within 3 to 6 hours prior to an episode of sepsis when the RR interval time series was (C) abnormal with low variability and (D) abnormal with low variability and transient deceleration.

Figure 6

Histogram of KS distance D for test datasets in Figure 5 to demonstrate stationarity. The smoothed contour is the expected histogram for a stationary dataset theoretically.