This was a retrospective analysis of fetal-maternal magnetocardiograms (MCGs) recorded from women who were enrolled in a study that was designed to measure the effect of maternal physical activity on the longitudinal development of fetal cardiac autonomic control. Methods and results of the study are previously published [12] [16] but will be presented briefly here. The study was approved by the Kansas City University of Medicine and Biosciences and University of Kansas Medical Center Institutional Review Boards, following the tenets of the Declaration of Helsinki. Recruitment was limited to low-risk, 20–35 year-old women carrying singleton pregnancies. All the women agreed to the data acquisition and for their records to be used in the study. They provided written informed consent stating as much for themselves and their unborn children prior to participation. All data were anonymized and de-identified prior to analysis.

Women were categorized into the Exercise group based on their response to the Modifiable Physical Activity Questionnaire (MPAQ) [17]. Women who engaged in moderate to vigorous aerobic exercise for a minimum of 30 minutes, 3 times per week throughout pregnancy were assigned to the Exercise group. Those who did not meet the above criterion were assigned to the Control group.

Data were acquired using an investigational 83-channel dedicated fetal biomagnetometer (CTF Systems, Inc., subsidiary of VSM MedTech Ltd., Vancouver, Canada). A continuous, 18 minute simultaneous fetal-maternal MCG was recorded for each subject. Women were tested in a magnetically shielded room to eliminate the influence of electromagnetic artifacts. The pregnant subjects were comfortably seated in a slightly reclined position in front of the biomagnetometer and monitored via video camera and microphone. The data were acquired using a 300 Hz sampling rate and recording filter of 0–75 Hz.

Data were digitally filtered between 1 and 40 Hz offline (bidirectional fourth-order Butterworth filter) then subjected to independent component analysis (ICA) using EEGLAB toolbox (version 4.311). ICA is a blind-source separation technique used to segregate the contributions from multivariate and spatially distinct electrophysiological sources into individual components (e.g., maternal heart, fetal heart, fetal diaphragm). Maternal and fetal MCG components were identified and R-peaks marked using software developed by our team [12]. We previously established through longitudinal measures obtained at 28, 32 and 36 weeks gestational age that group differences emerged at 36 weeks. Therefore, analysis for fetal-maternal synchronization characteristics was limited to data collected at 36 weeks. Datasets with excessive artifacts (ectopic beats, preventricular or preatrial contractions) were excluded from the analysis. Forty datasets were used for this retrospective analysis (Control: n = 19; Exercise: n = 21).

Each bivariate data set consisted of the fetal and maternal R peak times. From these data sets, timings of the fetal R peaks relative to maternal R peaks were determined. Synchrograms were constructed by plotting the phases of the fetal beats with respect to m maternal RR interval cycles (1≤ m ≤4) over time [13]. As in our previous work, we normalized the timing of these phases φ_mat to values between 0 and 1 (corresponding to 0 and 2π). Episodes in which the phase timing φ_mat remained constant for at least 10 s were considered to indicate fetal-maternal heart rate synchronization. In such episodes, the fetal beats within the maternal RR interval cycle were visible as n parallel horizontal groupings of data points (see Fig. 1). The identification of these groupings as synchronization epochs (SE) was performed using an algorithm developed by Toledo et al. [18], whereby the threshold values were dependent on the number of maternal cycles (threshold = 0.03*m) [13]. Based on the ranges of maternal and fetal heart rates, n was set 2≤ n ≤9 and we examined the non-redundant n:m combinations 2:1, 3:2, 4:3, 5:3, 7:3, 5:4, 7:4, 9:4.

In each 18 min. mother-fetus data set, we determined maternal and fetal HRV, the maternal respiratory rate and the number of SE found was noted. Each SE was characterized by its duration, n:m ratio, maternal and fetal RR interval duration and HRV, their bivariate variability Δq, the maternal phase φ_mat and maternal respiratory rate. The occurrence and characteristics of the SE data were examined in the maternal-fetal pairs of the two groups.

With respect to the calculation of φ_mat we took the following into consideration. We postulate that, in the case of physiological interaction, there will be a more or less fixed offset of the fetal beat timing with respect to the timing of the initial maternal beat, i.e. the φ_mat values will be similar. For instance, for any identified SE with a 2:1 combination, there will be two values of φ_mat: φ_mat1 and φ_mat2. For those SE which result from synchronization, the φ_mat1 values (and φ_mat2 values, respectively) will be similar. The φ_mat1 and φ_mat2 values of those SE which do not reflect physiological interaction but result from chance will be arbitrary and can take on any value between 0 and 1. Importantly, for n:m combinations with more than one beat (m >1), the value of φ_mat1 for synchronized SE will take on m different values depending on which of the m beats is first identified by the Toledo algorithm [18]. E.g. the offset φ_mat1 in a 3:2 SE with phase preference may take on one of two values, depending on whether the first of the 2 maternal RR interval identified (m = 2) contains 1 or 2 of the fetal beats (n = 3) (see also Fig. 1). In the analysis of possible phase preferences in the SE, we took this into account by adjusting the φ_mat values as follows: φ_mat-adj = φ_mat mod (1/m). φ_mat-adj was calculated only for the SE found in n:m combinations with at least 10% of all SE found (i.e. 3:2, 4:3, 5:3, 7:4, see Results).

With respect to the maternal respiratory rate, no direct measurement of the mothers’ breathing was performed during the data acquisition. Therefore the instantaneous maternal respiratory rate was estimated on the basis of ECG derived respiration (EDR) [19]. EDR evaluates and quantifies the respiration induced variations of the R-wave amplitude in the ECG. This approach was implemented here on the MCG signals of the mothers and permitted an assignment of respiratory rates on a beat-to-beat basis. The respiratory rate of a complete 18 min. data set or within the single SE was characterized by the median of the beat-to-beat rates.

Due to natural variations in both maternal and fetal heart rates, spurious transient episodes of constant phase timing of fetal beats may occur without any basis in physiological processes. In order to distinguish between such chance coordination and real synchronization, we constructed surrogate data sets in which the timing relationships between the fetal and maternal heart beats were destroyed. This was done by creating so-called twin surrogate data sets of the maternal data based on the recurrence matrix [20]. Details of the surrogate data approach can be found in [21] and the details of the calculation of the twin surrogates can be found in [14]. For the calculation of the twin surrogates we set the embedding dimension to 3 and used a delay of 1. In order to achieve stable statistical results we produced 10 surrogate data sets for each original set. The surrogate data sets consisting of the shuffled maternal R times and the original fetal R times were processed and analyzed in the same fashion as the original data and the results were compared to those of the original data. If no differences between the surrogate and original data become apparent, then the results in the original data may be due simply to chance and have no physiological significance.

Values are given as mean ± standard deviation (SD) or as median with the 90% confidence intervals (CI), as appropriate. In the presentation of the results in the text, Tables and Figures, it was at times expedient to divide the number of surrogate SE by ten in order to present the results of the original and surrogate data in the same order of magnitude. The analyses were nonetheless performed on all data.

HRV was quantified in the time domain for both the maternal and fetal RR time series using the SD of all normal-to-normal beats (SDNN) as a measure for overall variance and the root mean square of successive differences (RMSSD) of the normal-to-normal beats as a measure for short term variability. The bivariate variability Δq was calculated on the basis of the means and SDs of the fetal and maternal RR intervals as Δq = sqrt[(RR_mat/RR_fet² × SD_fet)²+(1/RR_fet × SD_mat)²], where RR = mean RR interval, SD = SDNN and the subscripts mat and fet stand for the corresponding maternal and fetal data [22]. Δq has no dimension.

The order of the n:m combinations used to display the data is: 5:4, 4:3, 3:2, 5:3, 7:4, 2:1, 9:4, 7:3. This corresponds to a decrease in the ratio of fetal to maternal beats (0.80, 0.75, 0.67, 0.60, 0.57, 0.50, 0.44, 0.43, respectively) and indicates an order starting with similar fetal and maternal heart rates to fetal heart rates much faster than the maternal. The differences in distributions of the number of SE over the n:m combinations were tested using Pearson’s Chi² test. The Mann-Whitney U test was used to examine differences between the Control and Exercise subjects with respect in the number of SE, SE duration, the maternal and fetal HRV measures and maternal respiratory rates. This test was also used to examine the corresponding group differences between the original and surrogate data. The differences in the duration of the SE in the different n:m combinations was tested using the Kruskal-Wallis test. The dependency of the number of SE per data set on HRV measures as well as on respiratory rate was quantified on the basis of linear regression. The strength of the relationships was given using the correlation coefficient r and the amount of variance explained by the coefficient of determination r². The frequency distribution of the φ_mat-adj values was examined for each n:m combination, both for the original and surrogate data, in histograms with a 5*n bin size. The distribution of the phases in the histograms was tested by fitting partial Fourier series to the data using the general linear least-squares model and the significance of the fit was estimated by the zero-amplitude test (F-statistic) [23]. Statistical significance was considered achieved at the level p<0.05 and all values were interpreted in an explorative manner.

RR interval duration in the data sets of the mothers varied considerably, ranging from 573 ms to 1028 ms (heart rate range: 58–105 bpm). The values did not differ significantly between those who had not performed exercise and those who had. Overall variance (SDNN) was also similar but short term variability (RMSSD) tended to be higher in the Exercise group (Table 1). In the fetuses, RR interval duration ranged from 373–518 ms (heart rate range: 116–161 bpm) and tended to be shorter in the fetuses whose mothers had not performed exercise. Both overall variance and short term variability were significantly greater in the Exercise group (Table 1). The combined fetal-maternal variance, quantified on the basis of Δq, showed that the data sets of Control group had significantly lower variance than those of the Exercise group: 0.134±0.024 vs. 0.172±0.041, p = 0.004. We did not find significant differences between the maternal HRV measures in the original 18 min. data sets and those in the surrogate data sets (Table 1).

In the 40 data sets examined, we found a total of 676 SE in the original data, considerably more in the Control than in the Exercise group (403 vs. 273). Most of the SE (88%) were in the n:m combinations 4:3, 3:2, 5:3 and 7:4 (Table 2). The distributions over the n:m combinations differed significantly overall between the Control and Exercise groups (p<0.001) and in particular for n:m = 4:3, 3:2, 2:1 and 9:4.

In the 10 sets of surrogate data, on average 687.8 SE/set were found, a few more than in the original data. The ratio of SE found in the Control and Exercise groups was similar to that in the original data (402.2 vs. 285.6). The distribution over the n:m combinations was not different between the original and surrogate data (p = 0.50, Table 2). Within the surrogate data, the distributions over the n:m combinations differed significantly overall between the Control and Exercise groups (p<0.001) and were similar to those in the original data.

In the original data the number of SE per subject was 16.9±7.4. There were significantly more SE in the Control group than in the Exercise group (21.2±5.9 vs. 13.0±6.5, p<0.001). Compared to the original data, the number of SE per subject in the surrogate data was not significantly different (17.2±7.2, p = 0.84). As in the original data, the number was higher in the surrogate Control group than in the surrogate Exercise group (21.2±5.4 vs. 13.6±6.9, p = 0.001).

The number of SE found in a single 18 min. data set was significantly dependent on both fetal and maternal HRV as well as their combined variance (Table 3). By way of example, the relationship between Δq and n(SE) is shown in Fig. 2.

Valid SE were defined as having a duration ≥10 s. The mean duration of the SE in the original data was 16.4±8.8 s (90% CI: 10–31 s). No differences were found between the Control and Exercise groups nor in their respective distributions over n:m combinations. Due to the higher number of SE per subject in the Control group, the total SE time per subject was significantly longer in this group compared to the Exercise group (366±141 s vs. 215±121 s, p = 0.003).

SE duration was shorter in the surrogate data (15.8±7.3, 90% CI: 10–30 s; p = 0.035). In contrast to the original data, the durations of the surrogate Exercise SE were significantly shorter than those of the Control (13.7±1.1 vs. 12.7±1.1 s, p<0.015). Also the distributions over the n:m combinations was not uniform (p<0.001). But, as in the original data, the total duration of SE episodes per subject was longer in the Control group (348±128 s vs. 202±111 s, p = 0.001).

For each SE that was identified, we calculated the mean RR interval, SDNN and RMSSD for both the mother’s and the fetus’ data. Comparing the subjects’ values between the Control and Exercise groups in the original data, we found no differences with respect to RR interval (Table 1). There was a tendency for slightly higher fetal and maternal SDNN and maternal RMSSD in the exercise group but no statistically significant differences could be found. Examination of the bivariate fetal-maternal variability also showed only a statistical trend in the difference between the Δq values of the Control and Exercise subjects: 0.043±0.011 vs. 0.054±0.030, p = 0.078. The results in the surrogate data were similar, whereby a significant difference was found for fetal RMSSD (Table 1) as well as for Δq (0.050±0.012 vs. 0.067±0.035, p = 0.034).

The distributions of the SE phases φ_mat-adj was examined for the n:m combinations 4:3, 3:2, 5:3 and 7:4. Histograms were constructed for each n:m combination with 5*n bins. In the original data there was a tendency to non-uniform distributions with roughly n peaks and troughs (Fig. 3). The fit to a sinus function was statistically significant for 4:3 (p = 0.025) and 7:4 (p = 0.015) and with a trend for 5:3 (p = 0.083). In the surrogate data, the distribution patterns were not clear and no statistical significance was found, with only a trend for 7:4 (p = 0.093).

The median maternal respiratory rates found in the subjects’ 18 min. data sets ranged from 12.8 to 20.7 cpm (mean ± SD: 17.0±1.9 cpm). Although the rates were somewhat higher in the Control group, this difference was not statistically significant (Control vs. Exercise: 17.4±1.8 cpm vs. 16.6±1.9 cpm, p = 0.196). Comparing the respiratory rates of the subjects’ SE gave similar results at somewhat slower rates: 16.3±2.1 vs. 15.6±2.1, p = 0.376.

There was a significant relationship between the number of SE found per data set and the respiratory rate: n(SE) = 1.3 * respiratory rate –5.4, p_slope = 0.034 (see also Fig. 4). Separating the subjects into groups with respiratory rates above and below 18 cpm [14] showed that those with slower rates had significantly fewer SE (15.1±6.9 vs. 20.5±7.5, p<0.001). Examining the Control and Exercise groups separately gave similar results (Control: 19.5±5.4 vs. 24.1±5.9, p<0.001 and Exercise: 11.7±6.0 vs. 16.3±7.2, p<0.001), whereby the Exercise subgroup values also tended to be lower than the corresponding Control values.

We also found a dependency of the total SE duration per data set on respiratory rate (total duration(SE) = 26 * respiratory rate –153, p_slope = 0.040) and examination of the subgroups showed results similar to the number of SE but these results were not statistically significant.