Sparse Coding of Cardiac Signals for Automated Component Selection after Blind Source Separation

Abstract

Wearable sensor technology like textile electrodes provides novel ambulatory health monitoring solutions but most often goes along with low signal quality. Blind Source Separation (BSS) is capable of extracting the Electrocardiogram (ECG) out of heavily distorted multichannel recordings. However, permutation indeterminacy has to be solved, i.e. the automated selection of the desired BSS output. To that end we propose to exploit the sparsity of the ECG modeled as a spike train of successive heartbeats. A binary code derived from a two-item dictionary {peak, no peak} and physiological a-priori information temporally represents every BSS output component. The (best) ECG component is automatically selected based on a modified Hamming distance comparing the components' code with the expected code behavior.Non-standard ECG recordings from ten healthy subjects performing common motions while wearing a sensor garment were subsequently processed in 10 s segments with spatio-temporal BSS. Our sparsity-based selection RCODE achieved 98.1% heart beat detection accuracy (ACC) by selecting a single component each after BSS. Traditional component selection based on higherorder statistics (e.g. skewness) achieved only 67.6% ACC. IntroductionAmbulatory vital sign recording supplements the standard clinical data acquisition by long-term measurements for early diagnosis of diseases or health and stress monitoring of people performing potentially dangerous tasks. Using measurement techniques like textile electrodes for wearable sensing, the recorded electrocardiogram (ECG) is of non-standard nature compared to its clinical counterpart. Moreover, the minimal-conductive measurement principle which allows for a flexible health monitoring is also strongly affected by movement artifacts [1].Blind Source Separation (BSS) is a signal processing technique capable of separating signal mixtures (e.g. mixtures of ECG and distortions) into its constituting components [2]. Spatio-temporal Independent Component Analysis (ICA) based on the FastICA algorithm [2] is one realization of BSS which has shown a superior performance on wearable data compared to the standard ICA [1]. Despite its ability to separate ECG from distortions, ICA is typically only solved up to a permutation (i.e. separated components are available but the output is mostly unordered). Accordingly, a desired output component (e.g. the one best representing the ECG) has to be automatically selected. This selection gains special interest while processing a large number of channels. This is typical for ambulatory multi-channel health recordings and particularly spatio-temporal ICA which adds extra channels during the processing [3].Two principles for handling permutation indeterminacy have been proposed in the context of ECG processing. The first principle identifies and discards the undesired components (i.e. artifacts) thus indirectly obtaining the ECG component. A combination of second-order and higherorder statistics was used for that purpose in [4] whereas auto-correlat...

Robust Methods for Automated Selection of Cardiac Signals After Blind Source Separation

“…In order to interpret the peak detections from the above procedure, we applied two different algorithms. The first (RCODE) was initially proposed by our group in [25]. It delivers a quasi-continuous measure between the expected behavior of a cardiac component consisting of peak detections followed by a reasonable time between subsequent peaks and differently pronounced deviations from this behavior up to a lack of multiple detections.…”

“…One major problem of automatically selecting a single (desired) component after BSS application is given by undesired components (e.g., artifact components) resembling features initially chosen to characterize the desired component. In particular, this is relevant for higher order moments, which are heavily affected by outliers or time-/frequency based features [25]. Moreover, both types of features may vary in absolute and relative values between datasets of different origin [37], which renders an according feature selection to be even more complicated.…”

“…A simple periodicity criterion was also applied to identify the ECG component inside the BSS output of an electroencephalogram (EEG) by using QRS detections in [24]. A novel approach for evaluation of the temporal behavior of QRS detections by utilization of sparse coding was proposed by our group in [25].…”

“…Identifying a robust component selector would allow powerful techniques as spatio-temporal BSS to become applicable for multichannel biosignal processing. Accordingly, this study compares the performance of two novel automated BSS output selection approaches that are based on the temporal assessment of heartbeat (QRS) detections (proposed in its initial form in [24] and [25]) with the use of higher order statistics and time/frequency domain features in order to identify a single optimal ECG component. The novel approaches to solve permutation indeterminacy aim at directly identifying ECG components in BSS outputs based on a sparse representation of each component.…”