CRCM Journal

Introduction

Elevated blood pressure variability (BPV) may indicate impaired cardiovascular regulation. As a risk factor for various health outcomes, blood pressure exhibits dynamic fluctuations in blood vessels, with systolic and diastolic blood pressures in its waveform [1]. The current evidence regarding its clinical relevance has several gaps and is based on heterogeneous studies with limited standardization of methods for BPV assessment [2]. In traditional Chinese medicine (TCM), practitioners sense the radial artery pulse in the blood pressure waveforms through the fingertips, especially using the “slippery pulse” pattern for fetal sex diagnosis [3]. However, variations in this pattern within the pulse waveform are mainly documented by explanatory words in classical literature. Pulse diagnosis requires apprenticeships and precise tactile sensations of the finger waveform, which are difficult to transmit and lack scientific data. This leads to differences in the interpretation of pulse waveforms. In this study, we employed scientific measurement methods to record pulses in pregnant women and applied machine learning techniques to classify fetal sex with the objective of exploring the impact of pregnancy on wrist pulse waveforms and potentially determining fetal sex.

The use of the “slippery pulse” pattern for diagnosing pregnancy has been recorded in various TCM literature sources. However, these studies mostly focused on explanations that lacked substantial quantitative data. For example, the foundational TCM text HuangTi-Nei-Ching indicates pregnancy through a “normal pulse but appearing weak” and “significant movement in the Shaoyin meridian” [4, 5]. Another instance is found in the “differential diagnosis of the Qi,” where the pulse can reveal the fetal sex: a larger left pulse indicates a male fetus, whereas a larger right pulse suggests a female fetus. The practice of using TCM pulse diagnosis for pregnancy and sex identification has been employed over time. The current research has delved into digital sensing measurement, revealing that measuring pressure amplitude with position differences in the radial pulse could assist in identifying fetal sex [6]. However, its accuracy remains questionable, owing to its reliance on the tactile sensation of the finger. This reliance on oral tradition, which is susceptible to distortion, adds complexity to discerning the exact meaning of foundational historical texts. Although historical insights are not always conclusive, they offer promising pathways for analyzing pregnant pulse waveforms.

Scientifically measuring and analyzing arterial pulse waves using sensors and computers are essential in clinical practice, encompassing aspects such as understanding how physiological factors impact pulse dynamics, diagnosing diseases, evaluating drug responses, and conducting epidemiological studies [7]. Various devices have been developed to facilitate these measurements [8]. For data preprocessing, the transform-based method, which is one of the four typical methods, delves into the frequency spectrum of the radial arterial pulse using a Fourier transformation [9]. Currently, the application of harmonic modes from transformed pulse data has proven useful in diagnosing conditions, such as abnormal liver function [10, 11], type 2 diabetes [12], coronary artery disease [13], pregnancy[14], and death [15, 16]. Additionally, the authors previously introduced the concept of organs and TCM meridians resonating with specific Fourier components of pressure waves [17, 18], highlighting the detection of disrupted hepatic, renal, or splenic qi in the pulse [16]. Moreover, AI algorithms can go beyond mere classification tasks; they can analyze vast datasets to detect early signs of diseases and conditions, thereby facilitating computational TCM diagnosis [19]. The authors also utilized AI classification techniques in a study of kidney insufficiency [20].

In this study, we addressed the gap in scientifically quantified pulse diagnosis by employing pressure sensors to measure pulses, utilizing a Fourier transform-based method to analyze the signals into harmonic components, and applying machine-learning techniques to support fetal gender classification. Furthermore, by establishing a correlation between harmonic data, monthly pregnancy changes, and fetal sex, our study contributes to a more comprehensive understanding of the intricate dynamics of wrist pulse waveforms during pregnancy. The significance of this study is that it bridges the gap between TCM practices and modern medical techniques but also opens avenues for further research into utilizing pulse waveform analysis for accurate diagnosis and monitoring of pregnancy-related changes.

Materials and Methods

Data were collected from pregnant women, and measurements were taken simultaneously from both hands of these pregnant women. Simultaneous measurements in both hands increased the complexity of the data collection, particularly considering that pregnant women are a high-risk group. Pregnant women who declined to undergo simultaneous measurements in both hands were excluded from the study. Therefore, a cohort of 86 participants, including 36 pregnant females, were examined using individual monthly measurements. Among them, there were 16 female fetuses and 20 were male fetuses. Measurements were obtained at least once a month, resulting in ten values for both hands. The remaining 50 participants were healthy females who did not menstruate and did not exhibit any apparent symptoms. Measurements were obtained from this group only once and served as the control group.

This resulted in 420 measurements, comprising 208 among female fetuses and 212 among male fetuses, as depicted in (Figure 1a) for the monthly measurement count and (Figure 1b) for the weekly measurement count. As pregnant participants visited the outpatient department (OPD) between 10 and 30 times during the study, repeated measurements of harmonic magnitudes were linked to the records of individual pregnant women. The entire study was conducted at the OPD over a two-year period. The exclusion criteria were irregular heart rhythms, arterial catheters on the wrists, or sensitive skin conditions. Participants were also required to abstain from consuming alcoholic and caffeinated beverages. One hour before the experiment, participants were instructed to fast and rest. Ten minutes before the test, blood pressure and heart rate of the pregnant participants were recorded. The results of these recordings are shown in (Figures 1c and 1d), respectively. All experiments were conducted with the permission of the participants, and informed consent was obtained after explaining the research purpose. The study was approved by the ethical committee of China Medical University Hospital, Taiwan (No. CMUH108-REC1-135). All the procedures performed in this study were in accordance with the relevant guidelines and regulations of the ethical standards of the institutional research committee and the Declaration of Helsinki.

Figure 1: The number of participants and their basic physiological data statistics. Statistical data on the number of (a) monthly and (b) weekly measurements. Monthly variations in (c) blood pressure and (d) heart rate were measured for all the participants.

Non-invasively measured pulse pressures of the right and left radical arteries were synchronously obtained for each participant using a pressure transducer (PSL-200GL, Kyowa Electronic Instrument Co. Ltd., Japan). The device was fixed on the participant’s skin using Scotch™ tape along with an adjustable belt with a small button to provide a suitable pressure on the transducer. Pulse pressure readings were recorded for both the left and right hands. Precise measurements were obtained by identifying the position at which the wrist bone and tendon on the thumb side (known as the “Guan” point in TCM) exhibited the highest pulse magnitude, as shown in (Figure 2a). Subsequently, the sensor was placed at this specific location. Our successful measurement criterion was to capture the pulse with the highest magnitude [11, 18].

Figure 2: Experimental setup and pulse waveform analysis. (a) Measurement setup for detecting pulse waveforms, achieved by placing a pressure sensor on the wrist to obtain signals. (b) The inset figure illustrates the pulse waveform. Each independent waveform can be transformed into the frequency domain through Fourier transformation to obtain various peaks. Each peak is sequentially numbered, representing individual harmonics.

The transducer output was connected to an Apple MacBook for analysis, using an A/D converter with a sampling rate of 500 data points/s. The pulse spectrum was subjected to Fourier transformation using a one-pulse period within a 6-second interval [17], as depicted in the inset of (Figure 2b) (inset). Following the Fourier transformation, we examined the magnitudes (Figure 2b) and corresponding phases of the harmonics along with their mean values. We developed analysis codes using MATLAB (MathWorks, 2022b, USA) and subsequently compared the magnitudes and phases of the first ten harmonics among each group, as shown in (Figure 3). Statistical analyses of these outcomes were conducted using SigmaPlot 14.0. We constructed a two-level hierarchical linear model (HLM) [21, 22] or a mixed model [23] to analyze the harmonic magnitude and group differences. The level-one is used to estimate the means of longitudinal intensive data (repeated harmonic magnitudes) and the level-two is used to estimate the gender differences in the average harmonic magnitudes. Descriptive analyses of harmonic magnitude data were performed using the SPSS 22.0 software package (SPSS Inc., Chicago, IL, USA), and the SPSS mixed module was used to conduct the HLM. For the machine learning components, MATLAB functions were applied, including predictor importance estimates by permutation using a random forest of regression trees, trained regression containing boosting 100 regression trees, and k-nearest neighbor classifier.

Figure 3: Top 10 Harmonics for normal women and pregnancy and their variation. (a) Right hand measurements and (b) Left hand measurements. (c) Monthly variation of the second harmonic based on fetal gender and hands. (d) Average and standard deviation of modified gender differences with harmonics. Here, ‘f’ denotes P < 0.01, and ‘*’ denotes P < 0.05.

In this study, we hypothesized that utilizing a highly accurate sensor at the ‘Guan point’ could effectively capture the necessary information to replace the three-position detection method mentioned in TCM [3]. This assumption is based on the idea that organs and their associated meridians resonate with specific Fourier components of the pressure waveform [17, 18]. Although other studies have often utilized different positions and pressures to obtain pulse signals, we argue that acquiring signals from various positions with varying pressures merely changes local boundary conditions. Consequently, these changes in the boundary conditions result in different local resonance behaviors, enabling the filtration or enhancement of signals in different frequency bands. Therefore, we assumed that employing a high-resolution sensor would be sufficient to compensate for these changes in boundary conditions and strengthen the signal.

Additionally, we considered the potential influence of various external factors on the participants, such as stress levels, physical activity, eating habits, sleeping patterns, and environmental conditions, which were not controlled for in this study. From a TCM perspective, these external factors may influence changes in harmonic values, potentially affecting the accuracy of our measurements. However, owing to the high-risk nature of pregnant women, we were unable to impose strict control over their lifestyle.

Results

Over a span of two years, we gathered pulse pressure data from both the left and right hands of 36 pregnant women and 50 normal women. After applying the Fourier transformation to these pulse pressure waveforms, the magnitudes of the top ten harmonics for these measurements are depicted in (Figure 3a (right) and 3b (left). Analyzing the results on the right-hand side, we observed that the magnitudes of the 1^st and 2^nd harmonics were elevated in pregnant women, whereas the magnitudes of the remaining seven harmonics were lower. After conducting t-tests, all these differences were found to be significant (t-test, P<0.01, denoted by f), except for the 4^th harmonic, which showed nearly equal values between the normal and pregnant groups. On the other hand, in the case of the left hand, it was observed that the 2^nd harmonic exhibited higher values among pregnant women, and a minor elevation was seen in the 4^th harmonic. However, the remaining eight harmonics exhibit lower values. Subsequent t-tests yielded significant results for all these differences (t-test, P< 0.01, denoted by f); the sole exception was the fourth harmonic (P<0.05, denoted by *). This result highlighted the distinct nature of the 2^nd harmonic. Furthermore, this harmonic emerged as a pivotal indicator after the third month, vividly mirroring the monthly trajectory of fetal growth and gender distinctions (Figure 3c). Notably, all second-harmonic magnitudes experienced significant increases until night. Additionally, pregnancy can be identified in conjunction with the 1^st, 3^rd, and 4^th harmonics, and there is the potential to identify pregnancy [24].

To explore fetal sex distinctions, we collected data synchronously from both the left and right hands of the pregnant women. Mean values and standard errors of the mean (SE) were computed for the male and female fetal groups and are presented in Supplementary Tables S1 and S2, respectively. We assessed the sex differences by subtracting the harmonic magnitudes of females from those of males (sex difference= male–female). These gender differences in harmonic magnitudes are delineated in the corresponding columns of these tables for both the individual left- and right-hand groups. It is important to note that no statistically significant gender differences emerged in either the individual left- or right-hand group. Additionally, we initially obtained the differences between the hands and then computed the modified sex differences. Hand differences were derived by subtracting the harmonic magnitudes of the right hand from those of the left hand (hand difference= leftright). Subsequently, we utilized the individual participants’ hand differences to calculate the modified gender difference obtained by subtracting the magnitude of the female hand difference from that of the male hand difference (modified gender difference=Male(LR)-Female(L-R)). Notable sex differences were observed (Table S3 in the Supplementary). In particular, the differences in the values of the 1^st (P < 0.01), 2^nd (P< 0.01), and 3^rd (P< 0.05) harmonic components were significantly higher in male fetuses, as depicted in Figure 3d. To assess monthly variations, we also considered modified sex differences. The outcomes for identifying modified sex differences based on each harmonic and comparison within each month are presented in (Figure 4), accompanied by the detailed raw data in Supplementary (Table S4).

Figure 4: Illustrating the monthly differences of modified gender differences with harmonics. The red arrows indicate the seventh month, which served as a turning point, with monthly changes particularly evident in the first four harmonics. Blue arrows indicate the earliest signs of male fetal identification. Here, ‘f’ denotes P < 0.01, and ‘*’ denotes P < 0.05.

We input the ten harmonic values contained in the left hand, right hand, and the difference between the two hands into the decision tree analysis to obtain a streamlined method for distinguishing between male and female fetuses. We also focused on the differences between the first three months (Figure 5a) and months four–ten (Figure 5b), and these figures show the original results of the classification tree view. In the analysis of the first three months, very high prediction probabilities can be achieved, especially for female fetuses. Male fetuses also exhibited discrimination rates of approximately 0.945, except for a small portion of branches with discrimination rates of approximately 0.667, but this accounted for only three out of 68 cases. The estimated probability for female fetuses was 1.0 (28/28), whereas that for male fetuses was 0.925 (37/40). Moreover, in the analysis from months 13 to 40, the discrimination rates were not as ideal as those in the first three months. The discrimination rates for female fetuses ranged from 0.682–0.939, whereas those for male fetuses ranged from 0.753–0.886. In total, the estimated probability of female fetuses was 0.84 (123/146), while that of male fetuses was 0.77 (127/165).

Figure 5: The decision tree analysis for classifying fetal genders. (a) Results of the first 12 weekly decision tree. (b) Displays the 13-40 weekly decision tree results. The importance estimates for the 30 harmonics are presented in (c) for the first 12 weeks, and in (d) for weeks 13-40. Here, ‘F’ and ‘0’ denote female fetus, ‘M’ and ‘1’ denote male fetus, and (n/N) denotes ‘n’ for the correct number and ‘N’ for the total number. For example, the values, “F: 0.832” and “M: 0.682” represent probability. The coding is as follows: the left hand (x1-x10), the right hand (x11-x20), and the difference between both hands (x21-x30) represent their ten harmonics.

We also used the out-of-bag predictor importance estimation method, specifically, a random forest of regression trees, to validate the important parameters (Figures 5c and 5d). The results indicate that the difference between the left and right hands plays a crucial role in distinguishing between fetal sexes during the first three months. However, in the analysis of months four–ten, the parameters of the left and right hands individually also showed a significant importance. The most crucial parameter in the first three months is the 8^th harmonic of the difference between both hands, while in the later months, the 2^nd, 3^rd, and 8^th harmonics also play important roles. In addition, we analyzed the cumulative weekly data, starting from the seventh week, using both the decision tree and K-nearest neighbors (KNN) methods to statistically analyze the recognition of fetal sex (see Supplementary Material).