Cardiac complications represent a common cause of birth death [1-3]. Fetal well-being evaluation is usually accomplished through cardiotocography, which provides information about the fetal status based on the fetal cardiac rhythm [4]. To reduce the number of operative deliveries possibly due to fetal distress, cardiotocography may be combined to fetal electrocardiography (FECG) [5-7] that, in addition to heart-rate monitoring, allows evaluations strictly related to the FECG morphology. Indeed, both P-wave and QRS-complex durations may be used to assess intrauterine growth restriction [8]; fetal supraventricular extrasystoles may indicate cases of congenital heart diseases to be treated during pregnancy or immediately after birth [9, 10]; eventually, fetal ST changes may indirectly indicate myocardial hypoxia [11, 12]. Direct FECG (DFECG), obtained from recordings acquired by directly positioning an electrode on the fetal scalp, is considered the gold standard but its invasiveness and its restricted applicability to the stage of labor only, have limited its clinical use [13]. Indirect FECG (IFECG), extracted from recordings acquired by positioning electrodes on the maternal abdomen [11, 14-16], is noninvasive and has a wider applicability also extended to the final weeks (approximately from the 38^th, when the vernix caseosa, which almost electrically shields the fetus, start to dissolve) of gestation. Besides IFECG, such recordings typically contain maternal electrocardiogram, maternal and fetal muscular noise, fetal electroencephalography, and other kinds of noise [11, 15, 16]. Denoising these recordings for IFECG extraction is a very challenging task [14]; actually, IFECG clinical use is mainly limited by its poor signal quality. This work, which is the second of a two-paper series on noninvasive fetal electrocardiography [17], proposes a procedure for denoising indirect recordings in order to obtain an IFECG tracing characterized by a signal quality at least comparable to that of DFECG. Availability of such procedure would indeed justify use of IFECG instead of DFECG and could contribute to the clinical spread of FECG.

The Segmented-Beat Modulation Method (SBMM) was recently proposed [18-21] as a denoising technique for electrocardiograms. SBMM works under the hypothesis of knowing R peaks and its theoretic principles make it particularly suitable for FECG applications. We have proposed an adaptation of the Pan-Tompkins algorithm [22] to fetal R-peak detection [17]. Here, we evaluated SBMM suitability to denoise indirect recordings in order to obtain a good-quality IFECG tracings. To this aim, SBMM was applied to direct and indirect recordings simultaneously acquired from pregnant women during labor.

By way of example, SBMM application to record 1 is depicted in Fig. (2), where simultaneously acquired DREC and IREC (channel 1) are represented together with all their components. As it can be seen, DFECG amplitude was much higher than IFECG amplitude (104 µV vs. 18 µV), whereas the amplitude difference between DN and IN was less marked (34 µV vs. 14 µV). Consequently, DSNR was higher than ISNR1 (4.9 dB vs. 1.0 dB). The noise level affecting IFECG at the end of the SBMM procedure was very low so that DSNR was lower than ISNR2 (4.9 dB vs. 11.2 dB).

Generalizing, for all records DFECG was always characterized by amplitude higher than that characterizing IFECG, independently of the channel (Table 1). Consequently, median (over the records) DFECG amplitude was significantly higher than median IFECG amplitude (104 [89;157] µV vs. 22 [16, 28] µV, P=7.66·10^-4). Instead, DN amplitude was higher or equal to IN amplitude (independently of channel) in records 1, 4 and 5, comparable in record 3 and lower in record 2 (Table 1). Consequently, median DN amplitude and median IN amplitude were not significantly different (70 [39;78] µV vs. 49 [25;77] µV, P=0.45). Moreover, in the direct acquisition modality, median DFECG amplitude was significantly higher than median DN amplitude (P=3.17·10^-2), whereas in the indirect acquisition modality, median IFECG amplitude was significantly lower than median IN amplitude (P=1.90·10^-3). Thus, DSNR was greater than ISNR1 in all channels of every record but channels 1 and 2 of record 5 (Table 1). Consequently, median DSNR was significantly greater than median ISNR1 (3.3 [1.6;4.8] dB vs. -2.3 [-7.4;0.6] dB, P=3.90·10^-3). At the end of SBMM processing, however, the noise level affecting IFECG was mostly removed so that median DSNR was significantly lower than median ISNR2 (3.3 [1.5;4.8] dB vs. 9.6 [8.0; 10.9] dB, P=9.84·10^-4).

Correlation between DFECG and IFECG was typically high and significant (ρ=0.78 [0.75;0.83], P<10^-208; Table 1). Only in records 3 and 5, ρ showed lower but still significant values (ρ=0.45 in channel 1 of record 3, and ρ=0.28 in channel 3 of record 5, respectively; P<10^-208) in correspondence of the channel with the lowest ISNR1 (-11.1 dB and -1.8 dB, respectively; Table 1).

Fig. (3) shows the 4 IREC channels of record 5 after subtraction of MECG (i.e. IFECG+IN). As it can be seen, represented signals show a significant amplitude variability among channels so that the IFECG component is more easily visible in some channels than in others. This finding can be generalized to all records. Indeed, ISNR1 variability among channels is a direct consequence of IFECG amplitude variability and IN amplitude variability (Table 1).

Fig. (2). Representation of 4.5 s of simultaneously acquired DREC and IREC (channel 1) relative to record 1, together with all their components individually plotted (DFECG amplitude: 104 µV; DN amplitude: 34 µV; IFECG amplitude: 18 µV; and IN amplitude: 14 µV).

This study evaluated SBMM suitability to denoise IREC in order to obtain an IFECG characterized by a signal quality at least comparable to that of DFECG, the latter being considered as the gold standard for FECG. Goodness of SBMM performance was assessed by correlating IFECG against DFECG. A good correlation between the two above-mentioned signals would justify the application of SBMM to IREC only, in future studies.

SBMM belongs to the class of template-based methods [11] for getting FECG. However, differently from the other proposed techniques, it introduces a modulation procedure to adjust for repolarization-length changes due to physiological heart-rate variability [29]. Thanks to this peculiar feature, SBMM strongly improves the accuracy of FECG estimation. A quantitative comparison of SBMM performance against other template-based methods with no modulation procedure is beyond the scope of the present work and was previously performed [28, 30]. However, a qualitative comparison may help understanding why SBMM represents an improvement with respect to the other template methods. IFECG estimation by template-based methods is obtained after MECG estimation and subtraction from IREC. If the modulation procedure is not applied, MECG is reconstructed as a tracing with fixed heart rate, so that in correspondence of the repolarization segment some misalignments may occur, causing significant artifacts in the resulting fetal tracing obtained by subtraction (Fig. (4), panel a). If the modulation procedure is introduced, maternal repolarization variability is tracked, and the artefacts are strongly reduced (Fig. (4), panel b). Analogously, when SBMM is applied for filtering fetal tracings (direct or indirect) from noise, it performs better than the other template-based techniques, since it is the only one able to track variations in fetal repolarization variability.

As every other template-based technique, SBMM works under the hypothesis of knowing R-peak positions. Here, to avoid confounding factors due to fetal R-peak misplacements, fetal R-peak detection was manually performed on DREC to ensure correct localization. Still, in order to evaluate the possible use of SBMM in real clinical cases in which DREC and IREC are typically not simultaneously available, it is important to observe that localization of the fetal R peaks (which must necessarily be automatic and not manual) is quite straightforward from DREC, where the DFECG component is dominant, whereas may become very challenging from IREC [31], where besides IFECG, are present other high amplitude components. Fetal R-peak localization is an interesting issue which, however, was beyond the scope of the present work but systematically treated by ourselves [17].

Fig. (3). Representation of 4.5 s of the 4 IREC channels (Ch) of record 5 after subtraction of MECG.

Fig. (4). Example of a beat of the IFECG + IN (solid line) overlapped to estimated MECG (bold dotted line) and IREC (dotted line) obtained without (panel a) and with (panel b) the modulation/demodulation process.

In our study, SBMM was applied to 60 s long recordings because the aim was the evaluation of a denoising technique for FECG applications and not the fetal parameters monitoring during labor, which typically last for hours. In studies interested in FECG-parameters monitoring during labor, 60 s windows of FECG should be recursively SBMM-filtered in order to get clean FECG tracings from which to perform measurements. Recursive application of SBMM allows almost real-time (one-minute delay) evaluation of such parameters and adaptation of the SBMM procedure to the physiological variability of FECG.

According to our results, DFECG amplitude was about an order of magnitude higher than IFECG amplitude (few hundred of µV the former, and tens of µV the latter) whereas the noise level was very variable over the records but, on average, comparable between the two acquisition modalities (direct vs. indirect). As a consequence, DSNR was usually higher than ISNR1. This finding matches expectations. DFECG is acquired with electrodes, which are in contact with the fetus, and, thus, its amplitude is expected to be quite large and DN is likely represented by fetal physiological signals other than FECG (such as fetal electroencephalography). Instead, IFECG is acquired by positioning the electrodes on the maternal abdomen, so that its amplitude is expected to be quite low and IN (which, according to our definitions does not include MECG) may incorporate both fetal and maternal physiological signals (such as fetal electroencephalography and maternal uterine contractions, etc). Despite the different signal amplitudes and the different levels of noise affecting them, DFECG and IFECG were strongly correlated (ρ=0.78), confirming that they are two different representations of the electrical activity of the fetal heart. The variability characterizing IFECG and IN amplitudes in the different channels is usually not reflected in the values of ρ, being correlation independent from signals amplitude. In addition, at the end of the SBMM procedure, ISNR2 was lower than DSNR, indicating that extracted IFECG was characterized by a better signal quality than DFECG.

A limitation of this study is the small number of records on which statistics were performed, due to the fact that because simultaneously recorded DREC and IREC are very rare. Nevertheless, the ρ values were statistically very significant in all cases (P < 10^-208), so that the SBMM ability to correctly extract FECG from both DREC and IREC was clearly demonstrated in spite of the limited number of application cases. Only occasionally lower values of ρ were observed. These may indicate that electrodes have not correctly acquired the signal or that the transformation from DFECG to IFECG might not be perfectly linear, as assumed when computing ρ. This latter hypothesis is physiologically sustainable, since there is no determined geometrical relationship among the locations of the electrodes and the fetal heart.

In conclusion, the high correlation observed between DFECG and IFECG indicates that these signals have equivalent morphological content; the finding that ISNR2 was higher than DSNR indicates that IFECG, obtained using SBMM, has better quality than DFECG. Thus, SBMM can be used to obtain clean, potentially clinically useful IFECG also when DFECG is not available. Thus, SBMM application to IFECG may contribute to the spread of this technique in the clinical practice, since able to provide good quality fetal tracings in a noninvasive, safe, simple and economic way. Future studies will compare clinical FECG parameters measured in DFECG and IFECG to confirm clinical utility of SBMM in clinical settings.