In this original study, we investigate the performances of machine
learning algorithms on a neonatal sepsis detection task. We consider
this work to be of great interest to both engineers and clinicians, as
it uses non-invasive, already existing, vital signs monitoring signals
in a population of very low birth weight infants at high risk of sepsis.
Vital sign variability may indeed represent a general indicator of
health and wellbeing and be helpful in the early detection of systematic
inflammation such as sepsis. We used state of the art feature extraction
technics and evaluate a large variety of binary classification models
among which a neural network based generative model. The models were
chosen from two main families: discriminative and generative. This
enables a comprehensive study of different kinds of traditional and
advanced binary classification algorithms.
Our study reveals that advanced machine learning models are more robust
to changes in the feature extraction pipeline, although linear
classifiers have a comparable performance when the feature extraction is
tuned. The advanced model performing the best is a neural network based
generative model which is a hybrid generative and discriminative model.
A large window length when computing the features is beneficial to
almost all algorithms, indicating the relevance of frequency domain
related features for the neonatal sepsis detection task.
Overall we obtain a classification AUROC above 0.85, which makes our
prediction models potentially relevant in clinical practice. This will
enable earlier therapeutic interventions and thereby reduce morbidity
and mortality in infants.