Biosignal-Based Machine Learning Predictors of Sepsis: A Mini-Review

Abstract

This work aims to provide insight into the most recent machine learning approaches to biosignal-based early sepsis prediction in the intensive care unit environment. A systematic search of the PubMed database revealed 29 original research papers. These works present sepsis prognosis and detection models that employ vital signs or densely sampled physiological waveforms (or their derivatives) acquired at the bedside or retrieved from electronic medical records. The papers were reviewed for the methods, predictors, datasets, number of participants, and performance achieved in the test set. Even though the sepsis prediction landscape is dominated by models that employ parameters derived from sparsely sampled biosignals, there are notable approaches built around densely sampled data, which speaks in favor of more synergistic solutions that benefit from both signal types. Given the already good quality of the models demonstrated using offline data, future research should prioritize achieving the promised performance in real-world intensive care unit operating conditions.