Augmentation of Dispersion Entropy for Handling Missing and Outlier Samples in Physiological Signal Monitoring

Abstract

Entropy quantification algorithms are becoming a prominent tool for the physiological monitoring of individuals through the effective measurement of irregularity in biological signals. However, to ensure their effective adaptation in monitoring applications, the performance of these algorithms needs to be robust when analysing time-series containing missing and outlier samples, which are common occurrence in physiological monitoring setups such as wearable devices and intensive care units. This paper focuses on augmenting Dispersion Entropy (DisEn) by introducing novel variations of the algorithm for improved performance in such applications. The original algorithm and its variations are tested under different experimental setups that are replicated across heart rate interval, electroencephalogram, and respiratory impedance time-series. Our results indicate that the algorithmic variations of DisEn achieve considerable improvements in performance while our analysis signifies that, in consensus with previous research, outlier samples can have a major impact in the performance of entropy quantification algorithms. Consequently, the presented variations can aid the implementation of DisEn to physiological monitoring applications through the mitigation of the disruptive effect of missing and outlier samples.

Assessment of Outliers and Detection of Artifactual Network Segments Using Univariate and Multivariate Dispersion Entropy on Physiological Signals

“…The disruption is more significant when outliers are present in the EEG and ECG signals, with slightly smaller disruption observed when outliers are present in the BP signal and significantly smaller for outliers in the RESP signal. A similar pattern was observed in a previous study, focusing on the development of univariate DisEn variations that would be robust to outliers [ 25 ]. When outliers are present in a signal segment, they tend to disrupt the process of allocating classes across the amplitude range of the segment [ 10 , 12 ] by significantly expanding it based on the outliers’ amplitudes.…”

“…No major deviations, in the values of , are observed among the same multivariate features across different experimental setups where the signal containing the outliers, changes. These findings indicate that the potential combination of univariate DisEn algorithms modified to be robust to the effect of outliers [ 25 ] with the current multivariate DisEn variation [ 12 ], could provide an effective entropy quantification interface for the extraction of features from network segments.…”

“…Concerning the distribution mean ( ), there are two choices to be considered. The first choice, which is in alignment with the simulation processes followed by previous studies testing the effect of outlier samples in the performance of ApEn, SampEn, and univariate DisEn [ 23 , 24 , 25 ], would be to ensure that outlier values are outside the physiological range of each recorded signal while maintaining the amplitude boundaries of respective sensing equipment, leading to a distribution equal to . The second choice would be to set a lower distribution to ensure that a minority of outliers remain within physiological range to also cover certain scenarios where sensor miscalibration and recording interferences could produce outliers of the respective magnitude, and for that purpose, a equal to would be set.…”

“…However, while their utilisation within a network physiology paradigm could aid in the monitoring of physiological systems, the challenges introduced by the existence of artifactual outlier samples, which are a common occurrence in physiological recordings due to electromagnetic interference, loose equipment attachment, and user movement, have to be addressed [ 20 , 21 , 22 ]. When tested, the performance of the univariate ApEn, SampEn, and DisEn algorithms was significantly disrupted during the analysis of signal segments containing outlier samples [ 23 , 24 , 25 ].…”

Entropy-Based Estimation of Event-Related De/Synchronization in Motor Imagery Using Vector-Quantized Patterns

“…Thus, Dispersion Entropy appraises the frequency of a symbolic space that is built in mapping each sample through a class pattern set across epochs [ 20 ], retaining higher sensitivity to amplitude differences and accepting adjacent instances of the same class [ 21 ]. Further improvements can be achieved by introducing information about amplitudes and distances [ 22 , 23 ]. Besides entering more free parameters to tune, the sample-based estimators face additional restrictions in the extraction of ERD/S dynamics like the fact that motor imagery activity reduces the EEG signal complexity [ 24 ].…”