Fetal QRS Detection by Multiple Channel Temporal Pattern Search Applied to the NInFEA Database

Abstract

Background: In some cases, the fetal electrocardiogram (ECG) has a relatively low signal to noise ratio. Much of the literature pertaining to fetal ECG processing focuses on methods for separating the fetal ECG from the maternal ECG rather than detecting the fetal QRS in high noise conditions. This paper describes the application of a previously described pattern search methodology (TEPS) for detecting the fetal QRS in noisy ECGs. Algorithm Summary: Signals are processed in non-overlapping 5 s segments. The maternal QRS peaks are detected in a segment from a reliable lead. Next, for each of a potentially large number of leads, after removing the maternal QRS peaks from the set of searchable peaks, single channel TEPS is applied to search for provisional fetal peak sequences. The provisional sequences are scored according to previously described single channel TEPS quality measures: temporal regularity, peak pair amplitude ratios, and number of skips. The provisional sequences across all leads are aggregated according to their average RR intervals and scores. The most likely average RR interval for the segment is estimated from this aggregation by forming a score weighted average of the RR intervals. Sequences with RR intervals close to the chosen RR interval are selected to form a set of high quality sequences. To account for the possibility of peak time offsets between different channels, these sequences are time aligned by time lagged cross correlation. An optimal sequence is formed from these time aligned sequences, with optimality criteria based on peak timing coherence, temporal regularity, and skips. Data: The above procedure was applied to 60 ECG signals (avg. duration approximately 31 s) from the Non-Invasive Multimodal Foetal ECG-Doppler Dataset for Antenatal Cardiology Research (NInFEA). The NInFEA dataset also includes ultrasound recordings from which RR intervals may be extracted. Results: Out of 59 records, RR intervals were well tracked in 57 records. (One of the 60 NInFEA records was excluded due to the inability to obtain a sinus rhythm peak sequence from the ultrasound recording). For these 57 records, 84% and 95% of the ECG RR intervals were within 5 ms and 10 ms, respectively, of the ultrasound derived RR intervals. The mean and median average absolute value RR interval difference (between ECG and ultrasound) over 5 second segments were 3.2 ms and 2.4 ms respectively, with 93% of segments having a mean average absolute value RR interval difference less than 7ms. Conclusion: The TEPS methodology shows promise for ECG based fetal sinus rhythm monitoring.

Segment Based Pattern Analysis Reveals a Persistent Regular Rhythm in the Motion Artifact Record of the MIT-BIH Noise Stress Test Database

“…Pan-Tompkins (from Hooman Sedghamiz, available at https://www.mathworks.com/matlabcentral/fileexchange/45840-complete-pan-tompkins-implementation-ecg-qrsdetector) is applied to 20 s long signal segments centered 10 s apart from one another. A triangular peak pulse train signal is generated from these peaks [6], with each triangle centered on a corresponding peak time. The triangles are of unit amplitude and 200 ms wide at the base, wide enough to allow reasonable RR interval changes to be captured by autocorrelation; Figure 2 shows an exemplary peak pulse train.…”

“…An RR interval time series (10 s increments) is derived from these autocorrelation peaks, and outliers are removed. Additionally, pulse trains (for the same 20 s segment) corresponding to two different channels are cross correlated to provide a measure of peak time coherence across channels' peaks [6].…”

“…In addition, the two channels were processed by multi-channel TEPS, which takes advantage of peak time coherence across channels and merges all peaks across channels. Both single and multi-channel TEPS have been previously described in [4,5,6]. Briefly, according to single channel TEPS, ECG sequences are generated from candidate peaks and scored for quality according to: 1) prominence of the peaks within a sequence; 2) temporal regularity; and 3) the number of skipped beats.…”

“…Motion artifact spikes can mimic the shape and size of QRS complexes [7]. Methods for avoiding false positives for such QRS mimicking noise include: 1) tagging them as noise by correlating them to an exogenous noise-related signal (such as an accelerometer) [5], 2) determining whether such peaks temporally cohere across different ECG channels [4,6]; and 3) relying on a priori aspects of heart rhythm that aren't mimicked by sequences of random motion artifact noise [3,4,5,6]. In a single channel recording, option 2 is not available, and option 3 can be confounded if the motion artifact is rhythmic, in which case, only option 1 is viable.…”

“…To overcome challenges associated with serial detection, neural networks have been applied to segments of ECGs to generate a sequence of QRS complexes in parallel [2]. Alternatively, a methodology known as TEPS (for Temporal Pattern Search) involves performing a combinatorial search for an optimal peak sequence within a data segment [3,4,5,6]. One challenge in benchmarking these alternatives is the relative paucity of publicly available databases that include very noisy ECGs with known QRS times.…”