Fetal electrocardiography (FECG) provides important information on fetal well-being and allows diagnosis of cardiac abnormalities, especially in early stages of heart development [1N. Barnes, and N. Archer, "Understanding congenital heart disease", Curr. Paediatr., vol. 15, pp. 421-428, 2005.
[http://dx.doi.org/10.1016/j.cupe.2005.06.003] -3L.K. Hornberger, and D.J. Sahn, "Rhythm abnormalities of the fetus", Heart, vol. 93, no. 10, pp. 1294-1300, 2007.
[http://dx.doi.org/10.1136/hrt.2005.069369] [PMID: 17890709] ]. FECG can be acquired in invasive (direct) or noninvasive (indirect) modalities by applying the electrodes directly on the fetal scalp or on the maternal abdomen, respectively [4A. Agostinelli, M. Grillo, A. Biagini, C. Giuliani, L. Burattini, S. Fioretti, F. Di Nardo, S.R. Giannubilo, A. Ciavattini, and L. Burattini, "Noninvasive fetal electrocardiography: an overview of the signal electrophysiological meaning, recording procedures, and processing techniques", Ann. Noninvasive Electrocardiol., vol. 20, no. 4, pp. 303-313, 2015.
[http://dx.doi.org/10.1111/anec.12259] [PMID: 25640061] ]. Direct FECG (DFECG) is typically characterized by a good signal quality but its applicability is limited to labour. On the other hand, indirect FECG (IFECG) is typically corrupted by a significant amount of noise, but the procedure can be applied from the 37^th-38^th week of pregnancy. Given its noninvasiveness and the wider period of applicability, IFECG is considered preferable over DFECG. However, its use is limited by the filtering difficulties for having a sufficiently clean (and thus clinically usefull) FECG [4A. Agostinelli, M. Grillo, A. Biagini, C. Giuliani, L. Burattini, S. Fioretti, F. Di Nardo, S.R. Giannubilo, A. Ciavattini, and L. Burattini, "Noninvasive fetal electrocardiography: an overview of the signal electrophysiological meaning, recording procedures, and processing techniques", Ann. Noninvasive Electrocardiol., vol. 20, no. 4, pp. 303-313, 2015.
[http://dx.doi.org/10.1111/anec.12259] [PMID: 25640061] -6R. Sameni, and G.D. Clifford, "A review of fetal ECG signal processing issues and promising directions", Open Pacing Electrophysiol. Ther. J., vol. 3, pp. 4-20, 2010.
[PMID: 21614148] ].

The main interference of IFECG is the maternal electrocardiographic component that typically has a much higher amplitude (5-10 times) so that can be estimated and subtracted [4A. Agostinelli, M. Grillo, A. Biagini, C. Giuliani, L. Burattini, S. Fioretti, F. Di Nardo, S.R. Giannubilo, A. Ciavattini, and L. Burattini, "Noninvasive fetal electrocardiography: an overview of the signal electrophysiological meaning, recording procedures, and processing techniques", Ann. Noninvasive Electrocardiol., vol. 20, no. 4, pp. 303-313, 2015.
[http://dx.doi.org/10.1111/anec.12259] [PMID: 25640061] , 7A. Agostinelli, A. Sbrollini, L. Burattini, S. Fioretti, F. Di Nardo, and L. Burattini, "Noninvasive fetal electrocardiography part II: Segmented-Beat Modulation Method for signal denoising", Open Biomed. Eng. J., vol. 11, 2017. (Ahead of print), 8R. Vullings, C. Peters, M. Mischi, R. Sluijter, G. Oei, and J. Bergmans, "Artifact reduction in maternal abdominal ECG recordings for fetal ECG estimation", In: 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 22-26 Aug. 2007.
[http://dx.doi.org/10.1109/IEMBS.2007.4352218] ]. Other corrupting noise kinds may have physiological (maternal and fetal electromyograms, fetal electroencephalogram, respiration, etc.) or nonphysiological (instrumentation noise, sampling noise and noise from the electrode/skin interface) origin. The amplitudes of these interferences are typically comparable to that of IFECG, making fetal R-peak detection quite challenging. Fetal R-peak detection, however, is an essential step for getting fetal heart-rate (HR, bpm) information and for extracting a clean IFECG from abdominal recordings when using template-based techniques [8R. Vullings, C. Peters, M. Mischi, R. Sluijter, G. Oei, and J. Bergmans, "Artifact reduction in maternal abdominal ECG recordings for fetal ECG estimation", In: 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 22-26 Aug. 2007.
[http://dx.doi.org/10.1109/IEMBS.2007.4352218] ]. Pan-Tompkins’ algorithm (PTA) [9J. Pan, and W.J. Tompkins, "A real-time QRS detection algorithm", IEEE Trans. Biomed. Eng., vol. 32, no. 3, pp. 230-236, 1985.
[http://dx.doi.org/10.1109/TBME.1985.325532] [PMID: 3997178] ] is a popular and traditional method for R-peak detection, originally designed for adult applications [10I.E. Sigurdardottir, "R-wave detection algorithms using adult and fetal ECG signals", M. Thesis, Chalmers University of Technology, Gothenburg, Sweden, 2013.]. Some studies [10I.E. Sigurdardottir, "R-wave detection algorithms using adult and fetal ECG signals", M. Thesis, Chalmers University of Technology, Gothenburg, Sweden, 2013., 11P. Rajput, B. Singh, and P. Karan, "Fetal heart monitor", Int. J. Sci. Eng. Res., vol. 3, pp. 1358-1364, 2012.] suggested its use also for fetal applications without, however, addressing the issue relative to its adaptation to the fetal conditions. Eventually, fetal R-peak detection by PTA proved to be superior to that based on zero-crossing counting and filter banks, respectively [10I.E. Sigurdardottir, "R-wave detection algorithms using adult and fetal ECG signals", M. Thesis, Chalmers University of Technology, Gothenburg, Sweden, 2013.]. Thus, this work, which is the first of a two-paper series on noninvasive fetal electrocardiography [7A. Agostinelli, A. Sbrollini, L. Burattini, S. Fioretti, F. Di Nardo, and L. Burattini, "Noninvasive fetal electrocardiography part II: Segmented-Beat Modulation Method for signal denoising", Open Biomed. Eng. J., vol. 11, 2017. (Ahead of print)], aimed to evaluate the suitability of PTA to FECG applications, and propose some adjustments and optimizations to improve fetal R-peak detection from FECG, especially IFECG. To evaluate and compare PTA and improved fetal PTA performances, both methods were applied to DFECG and IFECG. Clearly, R-peaks detection from IFECG is much more challenging than from DFECG and is also strongly dependent on the goodness of the procedure used to get it from the abdominal recording.

The results of the application of PTA and IFPTA to “Abdominal and Direct Fetal Electrocardiogram Database” are reported in Table 1. Overall, IFPTA performed better than PTA in all IFECG recordings, whereas the two methods performances were compared when applied to DFECG (Table 1). This is mainly due to the fact that SNR associated to DFECG was significantly grater than that associated to IFECG ([3.3 [1.6;4.8] dB vs. -2.3 [-7.4;0.6] dB, P=3.9∙10^-3, as previously found in [7A. Agostinelli, A. Sbrollini, L. Burattini, S. Fioretti, F. Di Nardo, and L. Burattini, "Noninvasive fetal electrocardiography part II: Segmented-Beat Modulation Method for signal denoising", Open Biomed. Eng. J., vol. 11, 2017. (Ahead of print)]). Specifically, HR and HRV errors computed by PTA were significantly higher than those computed by IFPTA (HR: 1.32 [1.00;2.74] bpm vs. 0.00 [0.00;1.30] bpm, P=4.22∙10^-2; HRV= 37.26 [8.10;52.91] bpm vs. 1.38 [0.00;4.70] bpm; P=9.40∙10^-4). Consequently, PPV, SE and F1 associated to PTA were significantly lower than those associated to IFPTA (PPV: 0.60 [0.60;0.82] vs. 0.90 [0.80;0.97], P=1.04∙10^-2; SE: 0.50 [0.34;0.84] vs. 0.89 [0.72;0.96], P=3.23∙10^-2, F1: 0.54 [0.34;0.83] vs. 0.89 [0.75;0.96], P=2.00∙10^-2). R-peak detection in DFECG was accurately carried out by both the methods (PPV, SE and F1 close to 1 in all cases; (Table 1)) whereas, performances over IFECG was channel dependent (Table 1). In particular, PPV SE and F1 significantly correlated with SNR (which is channel dependent), when using PTA and IFPTA (ρ= 0.75÷0.86, P<10^-4; (Table 2)) indicating that in channels with higher SNR, R-peak detection is more accurate. SNR, however, inversely correlated with HRV, especially for IFTPA (|ρ|= 0.45÷0.65, P<10^-2; (Table 2)), indicating that when SNR decreases, detected HRV tends to increase. Consequently, both PPV and SE significantly inversely correlated with HRV, both when using PTA and IFPTA (|ρ|=0.76÷0.91, P<10^-4; (Table 2)) indicating that in channels with lower HRV, R-peak detection tends to be more accurate. This finding is particularly important because it allows to identify the optimal channels for R-peak detection. According to this criterion, IFECG4 in RCD1, IFECG3 in RCD2, IFECG3 in RCD3, IFECG4 in RCD4 and IFECG1 in RCD5 represent the optimal IFECG channels when using PTA, whereas IFECG4 in RCD1, IFECG2 in RCD2, IFECG4 in RCD3, IFECG4 in RCD4 and IFECG2 in RCD5 represent the optimal channels when using IFPTA (Table 1). Using these optimal IFECG channels only, improvements in R-peak detection accuracy by IFPTA (PPV: 0.94 [0.88;0.96]; SE: 0.95

Table 1
Automatic R-peak detection accuracy.

[0.85;0.96]; F1=0.94 [0.87;0.96]) significantly overcame improvements by PTA (PPV: 0.79 [0.54;0.82], P= 2.38∙10^-2; SE: 0.80 [0.38;0.84], P=4.76∙10^-2; F1=0.79 [0.43;0,83], P=2.00∙10^-2), and approaches that on DFECG (PPV: 0.99 [0.98;1.00], P=7.90∙10^-3; SE:1.00 [1.00;1.00], P=7.90∙10^-3; F1=0.99 [0.99;1.00], P=1.00∙10^-2). IFPTA was found to be superior to PTA also when applied to the “Set A of the Noninvasive Fetal ECG”, as shown by the results obtained when using the optimal IFECG channels only (PPV: 0.68 [0.51;0.91] vs. 0.55 [0.36;0.68], P=4.00∙10^-2; SE: 0.56 [0.39;0.88] vs. 0.46 [0.32;0.72], P=2.30∙10^-1; F1: 0.60 [0.41;0.89] vs. 0.47 [0.33;0.70], P=1.10∙10^-1).

Table 2
Association between noise, heart-rate variability and R-peak detection accuracy.

As for standard electrocardiography, automatic R-peak detection represents a fundamental step in the computerized analysis of FECG. R-peak detection in FECG, however, may become very challenging, especially in IFECG applications, since tracings are often corrupted by physiologic interferences and noise that may completely hide the fetal R-peaks. Some studies claim to perform R-peak identification by means of PTA [9J. Pan, and W.J. Tompkins, "A real-time QRS detection algorithm", IEEE Trans. Biomed. Eng., vol. 32, no. 3, pp. 230-236, 1985.
[http://dx.doi.org/10.1109/TBME.1985.325532] [PMID: 3997178] , 10I.E. Sigurdardottir, "R-wave detection algorithms using adult and fetal ECG signals", M. Thesis, Chalmers University of Technology, Gothenburg, Sweden, 2013.], a technique originally designed to automatically detect R-peaks in adult tracings. Here, PTA accuracy in detecting R-peaks in FECG was evaluated and compared against IFPTA; our proposed adaptation of PTA to fetal cases that includes an adjustment of the PTA parameters and a corrector to minimize false detections.

Both PTA and IFPTA were initially tested on the “Abdominal and Direct Fetal Electrocardiogram Database” [12J. Jezewski, A. Matonia, T. Kupka, D. Roj, and R. Czabanski, "Determination of fetal heart rate from abdominal signals: evaluation of beat-to-beat accuracy in relation to the direct fetal electrocardiogram", Biomed. Tech. (Berl.), vol. 57, no. 5, pp. 383-394, 2012.
[http://dx.doi.org/10.1515/bmt-2011-0130] [PMID: 25854665] ] of PhysioNet, specifically thought for testing and evaluating automatic processing procedures on FECG [7A. Agostinelli, A. Sbrollini, L. Burattini, S. Fioretti, F. Di Nardo, and L. Burattini, "Noninvasive fetal electrocardiography part II: Segmented-Beat Modulation Method for signal denoising", Open Biomed. Eng. J., vol. 11, 2017. (Ahead of print), 12J. Jezewski, A. Matonia, T. Kupka, D. Roj, and R. Czabanski, "Determination of fetal heart rate from abdominal signals: evaluation of beat-to-beat accuracy in relation to the direct fetal electrocardiogram", Biomed. Tech. (Berl.), vol. 57, no. 5, pp. 383-394, 2012.
[http://dx.doi.org/10.1515/bmt-2011-0130] [PMID: 25854665] ]. The database included only 5 records, but offered great advantage of providing simultaneously acquired DFECG (gold standard) and 4-channel IFECG tracings. Large FECG databases are more easily available (for example in Physionet there is the “Non-Invasive Fetal Electrocardiogram Database” which includes 55 cases [13A.L. Goldberger, L.A. Amaral, L. Glass, J.M. Hausdorff, P.C. Ivanov, R.G. Mark, J.E. Mietus, G.B. Moody, C.K. Peng, and H.E. Stanley, "PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals", Circulation, vol. 101, no. 23, pp. E215-E220, 2000.
[http://dx.doi.org/10.1161/01.CIR.101.23.e215] [PMID: 10851218] ]) but typically miss IFECG, and thus are more useful for clinical applications of already tested algorithms. On the other hand, DFECG databases are more rare, since it would not necessarily be ethical to perform invasive monitoring on low-risk women in labor.

Our results indicate that both PTA and IFPTA accuracy (quantified using PPV, SE and F1) increased by increasing SNR, even though with each value of SNR, IFPTA performed better than PTA. Both the methods performed well and in a comparable way when applied to DFECG tracings (high SNR), thus indicating that PTA could be used in these recordings (as in [9J. Pan, and W.J. Tompkins, "A real-time QRS detection algorithm", IEEE Trans. Biomed. Eng., vol. 32, no. 3, pp. 230-236, 1985.
[http://dx.doi.org/10.1109/TBME.1985.325532] [PMID: 3997178] ]), whereas the difference between the PTA and IFPTA was significant in IFECG tracings, where PTA provided much less accurate R-peak detection than IFPTA. Thus, PTA should not be used in IFECG applications, where IFPTA is preferable.

R-peak detection accuracy was found to be channel-dependent, due to the fact that SNR is channel-dependent (Table 1). Thus, with the channels being simultaneously acquired, only the most accurate R-peak detection may be considered and, if needed, applied to all channels. Since accuracy increases with SNR, it would be reasonable to choose the lead characterized by the highest SNR. However, SNR may not be immediately available and, often, an accurate R-peak identification is needed for a reliable SNR quantification [7A. Agostinelli, A. Sbrollini, L. Burattini, S. Fioretti, F. Di Nardo, and L. Burattini, "Noninvasive fetal electrocardiography part II: Segmented-Beat Modulation Method for signal denoising", Open Biomed. Eng. J., vol. 11, 2017. (Ahead of print)]. Thus, SNR is not useful to identify the best channel for R-peak detection prior and an indirect SNR measure is desirable. HRV necessarily increases in case of both false-positive and false-negative detections. Consequently, HRV was used as an indirect measure of SNR. HRV was indeed found to inversely correlate with R-peak detection accuracy, with IFPTA accuracy being always better than PTA accuracy at each value of HRV. When using IFPTA in IFECG applications, by choosing the optimal channel as the one with the lowest HRV (which almost always corresponded to the channel with the highest PPV, SE and F1), IFPTA detection accuracy was high (PPV=94%, SE=95% and F1=94%) and approached DFECG (PPV=99%, SE=100%, and F1=99%). Thus, IFPTA allows a good tradeoff between using IFECG instead of DFECG, given IFECG noninvasiveness and longer period of applicability, and the need of having a good R-peak detection accuracy to allow reliable clinical evaluations on HR and further signal processing [7A. Agostinelli, A. Sbrollini, L. Burattini, S. Fioretti, F. Di Nardo, and L. Burattini, "Noninvasive fetal electrocardiography part II: Segmented-Beat Modulation Method for signal denoising", Open Biomed. Eng. J., vol. 11, 2017. (Ahead of print)]. Superiority of IFPTA over PTA was confirmed when analyzing the "Set A of the Noninvasive Fetal ECG" [13A.L. Goldberger, L.A. Amaral, L. Glass, J.M. Hausdorff, P.C. Ivanov, R.G. Mark, J.E. Mietus, G.B. Moody, C.K. Peng, and H.E. Stanley, "PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals", Circulation, vol. 101, no. 23, pp. E215-E220, 2000.
[http://dx.doi.org/10.1161/01.CIR.101.23.e215] [PMID: 10851218] ], also from Physionet, which included 100 IFECG tracings with R-peak annotations.

In conclusion, in IFECG applications, IFPTA is to be preferred over PTA, since it provides a much higher R-peak detection accuracy.