Integration of early physiological responses predicts later illness severity in preterm infants

Abstract

Physiological data are routinely recorded in intensive care, but their use for rapid assessment of illness severity or long-term morbidity prediction has been limited. We developed a physiological assessment score for preterm newborns, akin to an electronic Apgar score, based on standard signals recorded noninvasively on admission to a neonatal intensive care unit. We were able to accurately and reliably estimate the probability of an individual preterm infant's risk of severe morbidity on the basis of noninvasive measurements. This prediction algorithm was developed with electronically captured physiological time series data from the first 3 hours of life in preterm infants (< or =34 weeks gestation, birth weight < or =2000 g). Extraction and integration of the data with state-of-the-art machine learning methods produced a probability score for illness severity, the PhysiScore. PhysiScore was validated on 138 infants with the leave-one-out method to prospectively identify infants at risk of short- and long-term morbidity. PhysiScore provided higher accuracy prediction of overall morbidity (86% sensitive at 96% specificity) than other neonatal scoring systems, including the standard Apgar score. PhysiScore was particularly accurate at identifying infants with high morbidity related to specific complications (infection: 90% at 100%; cardiopulmonary: 96% at 100%). Physiological parameters, particularly short-term variability in respiratory and heart rates, contributed more to morbidity prediction than invasive laboratory studies. Our flexible methodology of individual risk prediction based on automated, rapid, noninvasive measurements can be easily applied to a range of prediction tasks to improve patient care and resource allocation.

Fig. 1. Receiver-operating-characteristic curves demonstrating PhysiScore’s performance in predicting high morbidity as it relates to

(A) conventional scoring systems (B) PhysiScore’s performance with laboratory studies (C) predictions for infants with infection-related complications (D) predictions for infants with major cardiopulmonary complications.

Fig. 2. Receiver-Operating-Characteristic (ROC) curve demonstrating the limited sensitivity of PhysiScore in predicting morbidity for infants with IVH

Each circle represents the IVH grade of a preterm neonate overlaid on their respective score.

Fig. 3. The significance of different physiological parameters in predicting high morbidity

(A) The learned weight (w_i in Eq. 1) for each physiological parameter incorporated in PhysiScore; error bars indicate variation in the weight over the different folds of the cross-validation. (B) The nonlinear function associating the parameter with the risk of high versus low morbidity.

Fig. 4. Processing signal subcomponents

Differing heart rate variability in two neonates matched for gestational age (29 weeks) and weight (1.15 kg ± 0.5 kg). Original and base signals are used to compute the residual signal. Differences in variability can be appreciated between the neonate predicted to have HM (right) versus LM (left) by PhysiScore.

Fig. 5. Distribution of residual heart rate variability (HRvarS) in all infants

Learned parametric distributions overlaid on the data distributions for HRvarS displayed for the HM versus LM categorization.