Iterative Covariance-Based Removal of Time-Synchronous Artifacts: Application to Gastrointestinal Electrical Recordings

Abstract

Objective: The aim of this study was to develop, validate, and apply a fully automated method for reducing large temporally synchronous artifacts present in electrical recordings made from the gastrointestinal (GI) serosa, which are problematic for properly assessing slow wave dynamics. Such artifacts routinely arise in experimental and clinical settings from motion, switching behavior of medical instruments, or electrode array manipulation.

Methods: A novel iterative Covariance-Based Reduction of Artifacts (COBRA) algorithm sequentially reduced artifact waveforms using an updating across-channel median as a noise template, scaled and subtracted from each channel based on their covariance.

Results: Application of COBRA substantially increased the signal-to-artifact ratio (12.8 ± 2.5 dB), while minimally attenuating the energy of the underlying source signal by 7.9% on average ( -11.1 ± 3.9 dB).

Conclusion: COBRA was shown to be highly effective for aiding recovery and accurate marking of slow wave events (sensitivity = 0.90 ± 0.04; positive-predictive value = 0.74 ± 0.08) from large segments of in vivo porcine GI electrical mapping data that would otherwise be lost due to a broad range of contaminating artifact waveforms.

Significance: Strongly reducing artifacts with COBRA ultimately allowed for rapid production of accurate isochronal activation maps detailing the dynamics of slow wave propagation in the porcine intestine. Such mapping studies can help characterize differences between normal and dysrhythmic events, which have been associated with GI abnormalities, such as intestinal ischemia and gastroparesis. The COBRA method may be generally applicable for removing temporally synchronous artifacts in other biosignal processing domains.

Fig. 1

Sample electrograms illustrating various types of artifacts typically encountered during recording sessions. Waveform amplitude was rescaled for clarity of display; vertical scale bars indicate 75 μV. Gray traces: filtered electrograms prior to artifact removal. Sections highlighted in light blue indicate manually identified artifact-corrupted time windows. Black traces: electrogams after application of COBRA artifact removal. The periodic slow wave waveform was sufficiently recovered in most cases. Significant overlap of gray and black traces indicates that the underlying SW waveform was well-preserved in non-artifact corrupted regions following application of COBRA.

Fig. 2

Flowchart summarizing main steps of the artifact removal method. The iterative step returning to compute an updated common reference signal is a key feature which allows artifacts to be removed sequentially. Stopping criteria are described in Section II-F.

Fig. 3

Illustration of synthetic data corrupted by a common-mode artifact extracted from real data set before (red) and after (black) application of COBRA artifact reduction. The width of the light gray box highlights the duration of the added artifact, and the height denotes the peak-to-peak noise in that window after artifact removal. The point of steepest descent was preserved for both SW waveforms, which occur ≈5 s apart.

Fig. 4

Validation performance metrics for in vivo data. The shape and color coded legend at bottom indicates the threshold value η. Data points represent the trial-averaged mean across 6 data segments for SIA and ΔSAR, and all 11 data segments for Sens, PPV, and A_roc. Error bars represent the standard error of the mean. Data points shown are jittered along the abscissa for visual clarity. A_roc results are shown for data segments not cleaned with COBRA (open black circles) to illustrate the net benefit of artifact removal.

Fig. 5

Illustration of removal of multiple artifacts in a 40 s duration data segment. (a) Iterative cleaning on a single electrogram. Red circles indicate FEVT marks. Note the seven consecutive false negatives in the signal prior to artifact removal (top) resulted from the two large artifacts dominating the SW detection process, both of which were initially identified by FEVT as false positives. Following artifact removal (bottom electrogram), 10 out of 10 SWs were properly marked as true positives by the FEVT algorithm. (b) Recordings from 12 electrodes pre- (gray) and post-cleaning (black). For visual clarity, traces have been normalized to have uniform peak-to-peak range. Traces are vertically stacked to emphasize the propagating nature of the SW as illustrated with orange arrow. A copy of an uncleaned signal is provided at bottom for clarity. Blue arrowheads mark common mode artifacts that swamp the underlying SW activity, leading to data loss. (c) Ladder plots before (blue +) and after (red ○) artifact cleaning of data segments shown in (b). The former is offset by +0.5 s for visual clarity. Prior to cleaning, the artifacts swamped correct marking of SWs throughout the data segment.

Fig. 6

Illustration of removal of long time-scale artifact beginning abruptly in the middle of a 40 s duration recording. (a) Successive iterations of COBRA artifact removal improved the ability of the FEVT algorithm to properly mark SWs. Red circles indicate FEVT marks, thickened on the uncleaned trace (top) for clarity. Initially, 5 FPs were marked (also indicated by blue marks in (b)). After artifact removal, 9 TPs and 2 FPs (1st and 3rd arrowheads, dark blue) were marked, with 1 FN (2nd arrowhead, cyan) identified by FEVT. (b) Ladder plots corresponding to signals shown in (a), with formatting as described in Fig. 5(c). Prior to cleaning, the SW ATs could not be identified in the clean or artifact-corrupted segments of the recordings. After cleaning, the SW pattern could be discerned from ATs marked by FEVT.

Fig. 7

Top panel: Array-wide SW rate with (red) and without (blue) COBRA artifact reduction. The presence of large artifacts often led to missed detection of true SW events, obscuring the true SW rate throughout the baseline analysis period. Bottom panel: Isochronal maps for 3 successive SWs measured from in vivo porcine jejunum following artifact removal in a time window (45.3 ≤ t ≤ 63.1 s) indicated by gray rectangle. Dots spaced on a square lattice correspond to electrode sites. Isochrones are drawn at 0.5 s intervals. Red and blue color-code the earliest and latest activation times, respectively, and the white arrow indicates the retrograde direction of SW propagation.