Entropy Analysis of Total Respiratory Time Series for Sepsis Detection

Neonatal sepsis is characterized by the system's extreme response to an infection and persists as one of the biggest life-threatening diseases. The gold standard treatment is administrating an antibiotic, which, unfortunately, is often made too late. The diagnosis should be easier, faster, and achieved through non-invasive methods. Recently, entropy, a non-linear feature, has been applied to different physiological signals to detect diseases having very promising results. In this study, several entropy measures were applied to the breathing cycle duration (TTot) of the respiratory signals for 20 neonates. In total, 18 distinct methods of entropy were initially applied to 30-minute segments. Using Spearman's correlation, it was detected strong correlation similarities between some of the measures. On the other hand, bubble, attention, phase, and spectral entropies were negatively correlated with all the other measures. To detect the presence of Sepsis, the slope of the multiscale entropy index was analyzed. Also, a changing point in the slope was probed, when possible, and then was applied linear regression to two subsets of data, before and after the changing point. Effectively, the Wilcoxon Sign Rank Test showed that the results for the total slope of the Sample, Corrected Conditional, Distribution, Permutation, Fuzzy, Gridded Distribution, Incremental, and Entropy of Entropy were statistically significant to infer that entropy decreases with time. Nonetheless, further work should confirm these results with a larger dataset that includes healthy and pathological neonates.