Can unguided machine learning re-discover physiologic signatures of neonatal sepsis?

BackgroundCardiorespiratory deterioration due to sepsis is a leading cause of morbidity and mortality for extremely premature infants with very low birth weight (VLBW, birthweight <1500g). Abnormal heart rate (HR) patterns precede the clinical diagnosis of late-onset sepsis in this population. Decades ago, clinicians recognized a pattern of reduced HR variability and increased HR decelerations in electrocardiogram tracings of septic preterm infants. A predictive logistic regression model was developed from this finding using mathematical algorithms that detect this signature of illness. Display of this model as the fold increase in risk of imminent sepsis reduced mortality in a large randomized trial. Here, we sought to determine if machine learning or deep learning approaches would identify this uncommon but distinctive signature of sepsis in VLBW infants. MethodsWe studied VLBW infants admitted from 2012 to 2021 to a regional Level IV NICU. We collected one-hour HR time series data from bedside monitoring sampled at 0.5 Hz (n=300 HR values per series) throughout the NICU admission. First, we applied the principles of highly comparative time series analysis (HCTSA) to generate many mathematical time series features and combined them in a machine learning model. Next, we used deep learning in the form of a convolutional neural network on the raw data to learn the HR features. The output was a set of HR records determined by HCTSA or deep learning to be at high risk for imminent sepsis. ResultsWe analyzed data from 566 infants with 61 episodes of sepsis. HCTSA and deep learning models predicted sepsis with high out-of-sample validation metrics. The riskiest records determined by both approaches demonstrated the previously identified HR signatures-reduced variability and increased decelerations. ConclusionsWe tested the ability of unguided machine learning approaches to detect the novel HR signature of sepsis in VLBW infants previously identified by human experts. Our main finding is that the computerized approach returned the same result - it identified heart rate characteristics of reduced variability and transient decelerations. We conclude that unguided machine learning can be as effective as human experts in identifying even a very rare phenotype in clinical data.