Spatio-Temporal Fuzzy Coding of Video Motion Time Series for Late Onset Sepsis Detection in Neonatalogy

Late-onset sepsis affects between 10 to 25% of very premature babies, standing as a primary contributor to neonatal mortality. In this paper, we present an innovative approach that relies solely on the analysis of newborns' motion, derived from the processing of videos acquired in neonatal intensive care units, for several consecutive days. Among a group of 16 very premature newborns, consisting of 8 sepsis and 8 control subjects, we initially extract 14 distinct motion features, in a total of 2636 hours of video recordings. These features are processed through a fusion of fuzzy spatio-temporal coding and smoothed principal component analysis. It leads to 16 trajectories, one per newborn, positioned differently in a 2D plane. They are then compared across parameters such as position, speed, and acceleration. Our findings underscore significant differences in 8 trajectory parameters between sepsis and control newborns. Finally, unsupervised classification was used to separate infected and healthy newborns using these parameters, achieving an accuracy of 81.3%.Clinical relevance- This approach offers a non invasive video-based motion method for early sepsis detection in premature newborns, enabling timely interventions and potentially improving neonatal outcomes.