Increased nonstationarity of neonatal heart rate before the clinical diagnosis of sepsis.

The clinical diagnosis of neonatal sepsis is preceded by abnormal heart rate (HR) characteristics of transient decelerations and reduced variability, which intuitively appear to be more nonstationary than normal HR variability. Our goals were to investigate stationarity of HR, and to devise measures useful for early diagnosis of neonatal sepsis. In this context, we define non-stationarity to be present when the observed data differ from surrogate data generated by stationary Gaussian noise with arbitrary linear correlations. We devised statistical methods for determining stationarity of HR data based on the two-sample Kolmogorov-Smirnov (KS) test. We compared distributions of KS distances between small sample epochs from clinical data with those of isospectral surrogates and of surrogates generated using the amplitude-adjusted Fourier transform technique, reasoning that they should differ significantly for nonstationary data. We found significant evidence of non-stationarity for records longer than 1 min. We developed new HR measures based on the empirical cumulative distribution function (ECDF) that are highly significantly associated with sepsis, but are not correlated with HR measures such as moments or sample entropy. We conclude that neonatal HR data cannot be assumed to be stationary, and become even less stationary prior to sepsis.