A Wireless High-Sensitivity Fetal Heart Sound Monitoring System

sensors

Letter A Wireless High-Sensitivity Fetal Heart Sound Monitoring System

Jianjun Wei 1,*, Zhenyuan Wang 1 and Xinpeng Xing 2

1 School of Telecommunications Engineering, Xidian University, Xi’an 710072, China; [email protected] 2 Shenzhen International Graduate School, Tsinghua University, Shenzhen 518055, China; [email protected] * Correspondence: [email protected]

Abstract: In certain cases, the condition of the fetus can be revealed by the fetal heart sound. However, when the sound is detected, it is mixed with noise from the external environment as well as internal disturbances. Our exclusive sensor, which was constructed of copper with an enclosed cavity, was designed to prevent external noise. In the sensor, a polyvinylidene fluoride (PVDF) piezoelectric film, with a frequency range covering that of the fetal heart sound, was adopted to convert the sound into an electrical signal. The adaptive support vector regression (SVR) algorithm was proposed to reduce internal disturbance. The weighted-index average algorithm with deviation correction was proposed to calculate the fetal heart rate. The fetal heart sound data were weighted automatically in the window and the weight was modified with an exponent between windows. The experiments show that the adaptive SVR algorithm was superior to empirical mode decomposition (EMD), the self-adaptive least square method (LSM), and wavelet transform. The weighted-index average algorithm weakens fetal heart rate jumps and the results are consistent with reality.

Keywords: fetal heart sound; fetal heart rate; PVDF piezoelectric ﬁlm; automatic weight; weighted- index average

Citation: Wei, J.; Wang, Z.; Xing, X. A 1. Introduction Wireless High-Sensitivity Fetal Heart The condition of the fetus in utero can be monitored effectively with the utility of Sound Monitoring System. Sensors 2021, 21, 193. https://doi.org/ fetal heart sound. It can be used to judge whether the intrauterine fetus is hypoxic and to 10.3390/s21010193 reflect the condition of the fetal central nervous system, amongst other things. Meanwhile, the data can be sent to an obstetrician. In this way, contact between the obstetrician and Received: 16 November 2020 the gravida is significantly increased, allowing potential problems to be found early, and Accepted: 26 December 2020 reducing the incidence of fetal mortality. Fetal heart sound acquisition methods can be Published: 30 December 2020 divided into active and passive forms. On the basis of the Doppler principle and dynamic weighting matrices, Hamelmann improved the signal-to-noise ratio effectively and was Publisher’s Note: MDPI stays neu- able to obtain the fetal heart rate [1]. Khandoker obtained more information related to fetal tral with regard to jurisdictional clai- heart sound based on the same principle [2]. The active method transmits ultrasonic signals ms in published maps and institutio- to the abdomen of the pregnant woman and changes the fetal environment. Hiroyuki nal affiliations. Sato designed a fetal heart sound sensor whose size is about 19.2 cm3 [3]. An enhanced non-intrusive monitoring method was proposed by Andreotti and the accuracy of fetal heart rate monitoring was increased [4]. Through these acquisition methods, the gained signal is close to the condition in the uterus. However, noise and disturbance still exist Copyright: © 2020 by the authors. Li- censee MDPI, Basel, Switzerland. and significantly deteriorate the signal quality. A phonography-based measurement of the This article is an open access article fetus was proposed by Kovacs. Segment structures with frequency splitting were used to distributed under the terms and con- distinguish the sound signal from the disturbance signals, such as those caused by hiccups, ditions of the Creative Commons At- the body’s rotation, and limb movements. However, the sound signal can only reflect tribution (CC BY) license (https:// the fetal breathing movement and not the condition of the fetal heart [5]. The contact creativecommons.org/licenses/by/ force of the sensor affects the quality of the acquired fetal heart sounds and there exists an 4.0/). optimum contact force for the sensors [6]. During recording, the high-quality segments

Sensors 2021, 21, 193. https://doi.org/10.3390/s21010193 https://www.mdpi.com/journal/sensors Sensors 2021, 21, x FOR PEER REVIEW 2 of 11

an optimum contact force for the sensors [6]. During recording, the high-quality segments can be automatically identified and selected, but the interference from the environment remains [7]. The aim of this study was to acquire a high-quality fetal heart signal and Sensors 2021, 21, 193 2 of 12 obtain the accurate fetal heart rate.

2.can Structure be automatically of identiﬁedthe Fetal and Heart selected, Sound but the interference Monitoring from the System environment remains [7]. The aim of this study was to acquire a high-quality fetal heart signal and obtainThe the accurate wireless fetal heart fetal rate. heart sound monitoring system is composed of two parts: fetal heart2. Structure sound of the collection Fetal Heart Sound and Monitoring processing. System Fetal heart sound collection requires a fetal heart soundThe wirelesssensor fetal and heart a soundcollection monitoring module. system is Fetal composed heart of two sound parts: fetal processing includes de-noising heart sound collection and processing. Fetal heart sound collection requires a fetal heart ofsound the sensor fetal and heart a collection sound module. and Fetal extraction heart sound processingof the fetal includes heart de-noising rate. of The whole structure is shown inthe Figure fetal heart 1 sound. Figure and extraction 1a shows of the fetalfetal heart heart rate. Thesound whole collection structure is shown and Figure 1b shows fetal heart in Figure1. Figure1a shows fetal heart sound collection and Figure1b shows fetal heart soundsound processing processing at the terminal. at the terminal.

fetal heart

sensor

amplifier A/D

coaxial line (a) Fetal heart collecting

fetal heart rate

audio noise pretreatment output reduction

cardiotocography

(b) Fetal heart sound processing FigureFigure 1. Structure1. Structure of the fetal of heart the sound fetal monitoring heart sound system. monitoring system. In the system, the fetal heart sound is converted into an electrical signal by the exclusive high-sensitivityIn the system, sensor. the Then, fetal the weak heart electrical sound signal is is converted amplified and convertedinto an electrical signal by the exclu- into a digital signal and transmitted to the terminal via Bluetooth. When the signal arrives, siveit is processed high-sensitivity through data conversion, sensor. normalization,Then, the meanweak removal, electrical and so on.signal The is amplified and converted original data are converted to double-precision floating-point operands for the subsequent intooperations. a digital The application signal of and zero-mean transmitted and normalization to the speed terminal up the algorithm via Bluetooth. so that When the signal arrives, itthe is time processed demands of real-world through scenarios data are conversion, met. The noise is reducednormalization, after pretreatment, mean removal, and so on. The originalthe fetal heart data sound are is extracted, converted and the fetal to heart double rate is-precision calculated. floating-point operands for the subse- quent3. Fetal Heartoperations. Sound Collection The application of zero-mean and normalization speed up the algorithm The disturbance to the fetal heart sound comes from two factors. The first is the sointerference that the from time the pregnant demands woman, of such real as- heartbeats,world scenarios breathing, rhythmic are met. stomach The noise is reduced after pretreatment,movements, and the the fetal echo of heart the fetal sound heart sound is extracted, on the internal and wall. the The fet secondal heart is rate is calculated. the external disturbance, such as random noise and power-line interference. The body interference is mixed in before the signal reaches the sensor and must be reduced using 3.software Fetal at Heart the terminal. Sound External Collection disturbance can be reduced to a minimum by using a proper sensor. The disturbance to the fetal heart sound comes from two factors. The first is the interference from the pregnant woman, such as heartbeats, breathing, rhythmic stomach movements, and the echo of the fetal heart sound on the internal wall. The second is the external disturbance, such as random noise and power-line interference. The body interference is mixed in before the signal reaches the sensor and must be reduced using software at the terminal. External disturbance can be reduced to a minimum by using a proper sensor.

3.1. Fetal Heart Sound Sensor The frequency of fetal heart sound is 10–400 Hz, and this includes components with frequencies lower than 20 Hz. The low-frequency limitation of the traditional microphone is 20 Hz, so the integrity of the fetal heart sound cannot be guaranteed in this way. Poly- vinylidene fluoride (PVDF) piezoelectric film can convert sound into electrical signals. The material has a wide frequency response, i.e., 0–400 MHz, and can gather more low-

Sensors 2021, 21, x FOR PEER REVIEW 3 of 11

frequency information. The impedance of PVDF is close to that of the human body and the signal reflection is small when they are contacted. Furthermore, the material is thin, Sensors 2021, 21, 193 light, soft, chemically stable, and it can be cut arbitrarily.3 Fabrication of 12 is simple and the cost is low.

3.1. FetalThe Heart key Sound to Sensor the amplifier connected with the sensor is the connection of the amplifier withThe PVDF frequency piezoelectric of fetal heart sound film. is 10–400 When Hz, and the this ambient includes components air vibrates, with the film vibrates. The varia- frequencies lower than 20 Hz. The low-frequency limitation of the traditional micro- phonetion isin 20 charge Hz, so the between integrity of the the fetal top heart and sound bottom cannot of be guaranteedthe film inproduces this an electrical signal. The way.charge Polyvinylidene between fluoride the (PVDF)top and piezoelectric bottom film is can the convert sam sounde with into a electrical reverse polarization when the film signals. The material has a wide frequency response, i.e., 0–400 MHz, and can gather more low-frequencyvoltage acts information. as a reference. The impedance For of this PVDF reason, is close to thatthe of signal the human was body put into the input of the differ- andential the signal amplifier. reflection The is small common when they arenoise contacted. in the Furthermore, two paths the material was cancelled is after difference ampli- thin, light, soft, chemically stable, and it can be cut arbitrarily. Fabrication is simple and thefication cost is low. due to the symmetry of the original signal. The ports of the transmission lines connectThe key the to the surface amplifier connected and the with bottom the sensor of is thethe connection film; the of the other amplifier ports connect the positive and with PVDF piezoelectric film. When the ambient air vibrates, the film vibrates. The variationnegative in charge terminals between theof topthe and differential bottom of the film amplifier. produces an Welding electrical signal. cannot be used because the film Thebecomes charge between deformed the top and at bottom very is high the same temperatures, with a reverse polarization and can when even the degrade completely. There- film voltage acts as a reference. For this reason, the signal was put into the input of the differentialfore, a perforation amplifier. The common method noise is in preferred. the two paths was The cancelled ground after differenceof the circuit is connected to the cop- amplificationper shell dueof the to the sensor symmetry and of the the original external signal. The disturbance ports of the transmission is shielded, especially to EMC in the lines connect the surface and the bottom of the film; the other ports connect the positive andenvironment. negative terminals The of the antenna differential effect amplifier. occurs Welding when cannot bethe used normal because theconductor line is used to connect filmthe becomes module deformed and the at very disturbance high temperatures, with and fixed can even frequency degrade completely. injections. The coaxial lines are used Therefore, a perforation method is preferred. The ground of the circuit is connected to the copperbetween shell of the the sensor film and and the externalthe differential disturbance is shielded, amplifier. especially The to EMC collected in the fetal heart sound signal is environment.transmitted The antennathrough effect the occurs wire when thein normalthe coaxial conductor lines, line is used while to connect a woven wire cloth is connected the module and the disturbance with fixed frequency injections. The coaxial lines are used betweenwith the the filmground and the to differential shield amplifier. it from The disturbance. collected fetal heart sound signal is transmitted through the wire in the coaxial lines, while a woven wire cloth is connected with theThe ground fetal to shield heart it from sound disturbance. sensor was designed according to the principle of a stethoscope andThe aerodynamics. fetal heart sound sensor It was designed used accordingin the closed to the principle environment of a stethoscope and is showed in Figure 2. and aerodynamics. It was used in the closed environment and is showed in Figure2.

Figure 2. Fetal heart sound sensor. Figure 2. fetal heart sound sensor. Figure2a presents the principle of the sensor and the position of the film. Figure2b is a 3D profile. The film was fixed by two clips in the center of the sensor and the conductor lines wereFigure drawn. 2a presents the principle of the sensor and the position of the film. Figure 2b is a 3D profile. The film was fixed by two clips in the center of the sensor and the conductor lines were drawn.

3.2. Fetal Heart Sound Collection Module The original signal collected by the sensor is processed by the collection module in the fetal heart sound monitoring system. The module includes an amplifier, Bluetooth, a charging circuit, and a feed circuit. They are shown in Figure 3.

Sensors 2021, 21, 193 4 of 12

3.2. Fetal Heart Sound Collection Module Sensors 2021, 21, x FOR PEER REVIEW The original signal collected by the sensor is processed by the collection module in 4 of 11 the fetal heart sound monitoring system. The module includes an ampliﬁer, Bluetooth, a charging circuit, and a feed circuit. They are shown in Figure3.

Charging Power Module

S+ +

RG OPAA0 Bluetooth S- -

Figure 3. The structure of fetal heart sound collection module. Figure 3. The structure of fetal heart sound collection module. The amplification multiple is controlled by the feedback resistor, RG, in the amplifier. The relationship between resistance and amplification multiple is: The amplification multiple is controlled by the feedback resistor, RG, in the amplifier. 49.4 kΩ G = + 1 (1) The relationship between resistanceRG and amplification multiple is: The amplification multiple was 12 after the experiment; it corresponds49.4 kΩ to a resistance of 4.7 k. 퐺 = + 1 (1) 푅 The sensor in contact with the skin has zero potential just like퐺 the reference with the differential amplifier, because the positive and negative terminals of the amplifier are connectedThe with amplification the top and bottom multiple surfaces of thewas PVDF. 12 Bioelectricityafter the experiment; exists in the skin it corresponds to a resistance andof 4.7 disturbs k. the collected sound. Through the adoption of this structure, the disturbance induced by human bioelectricity was reduced. The fetal heart sound is sent to a terminal, such asThe a smart sensor phone, viain Bluetooth,contact afterwith amplification. the skin Thehas smart zero phone potential is the main just like the reference with the communicationdifferential technology amplifier, in daily because life and almost the allpositive have Bluetooth and capabilities,negative making terminals of the amplifier are con- this relevant for real-world applications. The Bluetooth connection-oriented approach, Santanected Cruz with Operation the (SCO), top wasand adopted bottom and the surfaces data were transmittedof the PVDF. in real time Bioelectricity from exists in the skin and thedisturbs collection the module collected to the smart so phone.und. The Through rechargeable the lithium adoption battery was of charged this structure, the disturbance in- using a charging circuit when it is running low. The positive and negative voltage sources, whichduced were by produced human by the bioelectricity feed circuit, supplied was Bluetooth reduced. and the The amplifier. fetal heart sound is sent to a terminal, suchThe as fetal a heartsmart sound phone, was amplified via threeBluetooth, times in wireless after fetal amplification. sound monitoring The smart phone is the main system: in the sensor, in the analog amplifier, and at the terminal. The amplification in thecommunication sensor ensures the weak tech fetalnology heart sound in daily can be easily life detectedand almost by the convector. all have The Bluetooth capabilities, making amplification at the terminal ensures the signal in the screen is easy to observe. Only the amplificationthis relevant in the amplifier for real amplifies-world the actual applications. signal and represses The the Bluetooth noise. connection-oriented approach, Santa Cruz Operation (SCO), was adopted and the data were transmitted in real time from 4. Adaptive Noise Reduction of the Fetal Heart Sound the Whencolle thection fetal heartmodule sound reachedto the thesmart sensor, phone. it was mixed The with rechargeable the noise of the lithium battery was charged pregnantusing a woman’s charging body circuit so that the when collected it signal is running contained low. a lot of The disturbance. positive The and negative voltage sources, random noise was superposed to the low-frequency signal and burrs appeared in the signal curve.which This were is equal produced to a band whose by center the isfeed the fetal circuit, heart sound supplied signal and Bluetooth the width is and the amplifier. the magnitudeThe fet ofal the heart noise. The sound aim of noisewas reduction amplified is to find three the center times line fromin wireless the fetal sound monitoring system: in the sensor, in the analog amplifier, and at the terminal. The amplification in the sensor ensures the weak fetal heart sound can be easily detected by the convector. The amplification at the terminal ensures the signal in the screen is easy to observe. Only the amplification in the amplifier amplifies the actual signal and represses the noise.

4. Adaptive Noise Reduction of the Fetal Heart Sound When the fetal heart sound reached the sensor, it was mixed with the noise of the pregnant woman’s body so that the collected signal contained a lot of disturbance. The random noise was superposed to the low-frequency signal and burrs appeared in the signal curve. This is equal to a band whose center is the fetal heart sound signal and the width is the magnitude of the noise. The aim of noise reduction is to find the center line from the disturbed signal [8,9]. The adaptive noise reduction of the fetal heart sound was adopted to achieve this goal. The training operation in the support vector regression (SVR) (Supplementary material) algorithm is relatively long, and the continual readjustment of C, the regular parameter, and 훾, the bandwidth of the Gauss kernel, cannot satisfy real-time requirements. For this reason, the length of the trained data sample must be controlled. If it is too long, the time requirements cannot be met; if it is too short, the accuracy of C and 훾 cannot be satisfied. The fetal heart sound is periodic and the parameters can, in theory, only be determined in one period. The fetal heart rate varies from 120 to 160 beats per minute and two periods are contained in one second, so the data in one second were selected to train and to reduce noise. The fetal heart rate may change slowly and C and 훾 change with it.

Sensors 2021, 21, 193 5 of 12

disturbed signal [8,9]. The adaptive noise reduction of the fetal heart sound was adopted to achieve this goal. The training operation in the support vector regression (SVR) (Supplementary Materi- als) algorithm is relatively long, and the continual readjustment of C, the regular parameter, and γ, the bandwidth of the Gauss kernel, cannot satisfy real-time requirements. For this reason, the length of the trained data sample must be controlled. If it is too long, the time requirements cannot be met; if it is too short, the accuracy of C and γ cannot be satisfied. The fetal heart sound is periodic and the parameters can, in theory, only be determined in one period. The fetal heart rate varies from 120 to 160 beats per minute and two periods are contained in one second, so the data in one second were selected to train and to reduce noise. The fetal heart rate may change slowly and C and γ change with it. The threshold, θ, was set, and when the mean square error (MSE) was larger than θ, the current data were trained again to obtain an optimum C and γ. Cross-validation in deep-learning algorithms was used to monitor the status and prevent overfitting in training [10–14]. If MSE was smaller than θ, the current C and γ were regarded as perfect, the result was output, and the new data continued to reduce noise. The adaptive SVR noise reduction algorithm does not contain manually defined C and γ and obtains optimum parameters through self-learning according to the character of the real-time signal. Thus, the system adaptability was strengthened.

5. Extraction of Fetal Heart Rate The direct calculation method or time-domain windowing can be used to extract the fetal heart rate [15,16]. The signal peak’s positions may drift and any tiny variation in denominator can be amplified. The direct calculation method outputs unsteadily and is not fit for real-time applications. Time-domain windowing was proposed to weaken the defect in the direct calculation method to some degree, as tiny changes in fetal heart rate are not uncommon [17]. If a spike is low that it is missed, the result will decrease by the smallest unit; meanwhile, if the disturbance is detected as a fetal heart sound peak, the result will increase by the smallest unit. The weight associated with each peak interval causes the calculation of fetal heart rate to be incorrect in the condition of a missed or incorrect detection. Although the mean can weaken this disturbance, the calculation of fetal heart rate will be significantly influenced if the missed or incorrect detection is severe. The fetal heart sound data were weighted automatically in the window so that the noise was removed and the calculation accuracy of fetal heart rate was increased. The ability of the automatic weight algorithm to eliminate disturbance decreases when considerable noise is apparent in the window. In extreme cases, the noise may overwrite the data and cause the output to fluctuate greatly. It must be restrained so that the fetal heart rate value does not change greatly as compared to the previous interval. The weighted-index average method was adopted to modify the weight between windows. The weighted coefficients of each data point decreased exponentially with time, and the closer the data point, the larger the weighted coefficient. This can be shown as:  v =  0 0,  = · + − ·  v1 β v0 (1 β) η1 v2 = β·v1 + (1 − β)·η2 (2)   ......  vt = β·vt−1 + (1 − β)·ηt

where vt is the optimal value after weighted-index average at t, ηt denotes the calculated fetal heart rate in the window at t, β is length factor between 0 and 1. vt is approximately equal to the mean of 1/(1 − β) data:

2 t vt = (1 − β)·ηt + (1 − β)·β ·ηt−1 + ... + (1 − β)·β ·η0 (3) Sensors 2021, 21, 193 6 of 12

where: 1 1 (1 − ρ) ρ ≈ ≈ 0.36 (4) e Let ρ = 1 − β, so: 1 1 (β) 1−β ≈ ≈ 0.36 (5) e The influence factor fell to about a third after 1/(1 − β) data. This had great influence on the result in the previous calculation because v0 initially equals zero. For this reason, the deviation should be modified and the iterative formula is:

β·v − + (1 − β)·η v = t 1 t (6) t 1 − βt

The numerator has important value when ‘t’ is small. It is demonstrated that 1 − βt is closing to 1 with the increment of t and the difference can be ignored after 1/(1 − β) data. The output of the fetal heart rate becomes more stable after a precise calculation in the windows and weighted-index modiﬁcation outside the windows. The proposed extraction algorithm of fetal heart rate is shown in Algorithm 1.

Algorithm 1: Proposed Fetal Heart Rate Extraction Algorithm Input: D, fetal heart sound data with the time of T seconds, H, peak margin matrix, F0 = 0, t = 0, w = 4. 1. the fetal heart sound data were obtained from t to t + w in D; 2. sampling down, from 8 k to 1 k; 3. ﬁnd out the position of m peaks; 4. calculate the interval of m−1 peak and sort them by size (largest to smallest) to obtain matrix P; 5. calculate the time difference of m−2 peak interval and obtain matrix E; 6. ﬁnd out the index i, correspond to the minimum in E; 7. set E(i) = +∞, put P(i) and P(i + 1) into matrix H; 8. repeat step 5 and 6 until min(E(i)) = ∞; 9. calculate the mean of element in matrix H, At = mean(H); 10. calculate the fetal heart rate in window: ft = 60/At; = β·Ft−1+(1−β)· ft 11. modify the fetal heart rate Ft 1−βt ; 12. move window backward: t = t + 1; 13. repeat 1–11 until t + w > T; Output: F0, F1, L, Ft−1.

The amount of data was reduced through sampling down at step 2, so the time cost of the algorithm declined and real-time requirements were easy to achieve. The interval between adjacent peaks was calculated at step 4. The data were automatically weighted inside of the windows from step 3 to step 9. The fetal heart rate was modiﬁed by the weighted-index average method between windows from step 10 to step 13.

6. Experiment Result 6.1. Realization of the Collection Module The fetal heart sound sensor was made of copper according to Figure2, and the external disturbance was prevented by means of the airtight collection environment. The circular surface under the sensor was placed on the abdomen of the pregnant woman and all the energy, resulting from the enclosed air vibrating due to the fetal heart sound, was collected and sent to the PVDF film. Physical amplification occurs when the signal collected on the larger inductive surface is applied to the relatively small film. The energy loss is negligible inside the enclosed cavity and the weak fetal heart sound can be detected and collected easily. The external electromagnetic environment is complex and copper cover can block the interference in a great degree. The portable wireless fetal sound monitoring system’s collection circuit was made up of a two-layer printed circuit board (PCB) and is shown in Figure4. Sensors 2021, 21, x FOR PEER REVIEW 6 of 11

The numerator has important value when ‘t’ is small. It is demonstrated that 1 − 훽푡 is closing to 1 with the increment of t and the difference can be ignored after 1/(1 − 훽) data. The output of the fetal heart rate becomes more stable after a precise calculation in the windows and weighted-index modification outside the windows. The proposed extraction algorithm of fetal heart rate is shown in Algorithm 1.

Algorithm 1: Proposed Fetal Heart Rate Extraction Algorithm Input: D, fetal heart sound data with the time of T seconds, H, peak margin matrix, F0 = 0,t = 0,w = 4. 1. the fetal heart sound data were obtained from t to t+w in D; 2. sampling down, from 8 k to 1 k; 3. find out the position of m peaks; 4. calculate the interval of m-1 peak and sort them by size (largest to smallest) to obtain matrix P; 5. calculate the time difference of m-2 peak interval and obtain matrix E; 6. find out the index i, correspond to the minimum in E; 7. set 퐸(푖) = +∞, put P(i) and P(i + 1) into matrix H; 8. repeat step 5 and 6 until 푚푖푛(퐸(푖)) = ∞; 9. calculate the mean of element in matrix H, At = mean(H); 10. calculate the fetal heart rate in window: ft = 60/At; 훽∙퐹 +(1−훽)∙푓 11. modify the fetal heart rate 퐹 = 푡−1 푡; 푡 1−훽푡 12. move window backward: t = t + 1; 13. repeat 1–11 until t + w > T; Output: F0, F1, L, Ft-1.

The amount of data was reduced through sampling down at step 2, so the time cost of the algorithm declined and real-time requirements were easy to achieve. The interval between adjacent peaks was calculated at step 4. The data were automatically weighted inside of the windows from step 3 to step 9. The fetal heart rate was modified by the weighted-index average method between windows from step 10 to step 13.

6. Experiment Result 6.1. Realization of the Collection Module The fetal heart sound sensor was made of copper according to Figure 2, and the external disturbance was prevented by means of the airtight collection environment. The circular surface under the sensor was placed on the abdomen of the pregnant woman and all the energy, resulting from the enclosed air vibrating due to the fetal heart sound, was collected and sent to the PVDF film. Physical amplification occurs when the signal collected on the larger inductive surface is applied to the relatively small film. The energy loss is negligible inside the enclosed cavity and the weak fetal heart sound can be detected and collected easily. The external electromagnetic environment is complex and copper cover can block the interference in a great degree.

Sensors 2021, 21, 193 The portable wireless fetal sound monitoring system’s7 of 12 collection circuit was made Sensors 2021, 21, x FOR PEER REVIEW 7 of 11 up of a two-layer printed circuit board (PCB) and is shown in Figure 4.

The circuit was minimized, its power was reduced, saving space and improving func- tionality. The PCB was designed as a circular plate in order to fit the sensor and reduce the general size. The left image in the Figure 4 is the front, which is responsible for the transmission of the wireless signal. The right image is the back, which is for power man- agement, signal input, and amplification.

6.2. Experiment of Adaptive Noise Reduction FigureFigure 4. Collection 4. Collection circuit of the circuit portable of wireless the portable fetal sound monitoring wireless system. fetal sound monitoring system. The fetal heart sound data were cut into 0.5 s, 휀 = 0.1, 휃 = 0.1. The magnitude was The circuit was minimized, its power was reduced, saving space and improving normalizedfunctionality. The and PCB the was signal designed after as a circular noise plate reduction in order to is ﬁt shown the sensor in and Figure 5. reduceThe the general noise size. in Thethe left original image in the fetal Figure heart4 is the front,sound which data is responsible is larger for and the second heart sound the transmission of the wireless signal. The right image is the back, which is for power ismanagement, almost completely signal input, and covered. ampliﬁcation. The adaptive SVR algorithm followed well in the first and

second6.2. Experiment heart of Adaptivesounds Noise and Reduction inhibited the noise effectively in the other part. The tendency of the fittingThe fetal curve heart sound was data severely were cut into restricted 0.5 s, ε = 0.1 by, θ = the0.1 .points The magnitude of the was SVR. normalized and the signal after noise reduction is shown in Figure5.

Figure 5. Adaptive Support Vector Regress (SVR) noise reduction of fetal heart sound. Figure 5. Adaptive Support Vector Regress (SVR) noise reduction of fetal heart sound. The noise in the original fetal heart sound data is larger and the second heart sound is almost completely covered. The adaptive SVR algorithm followed well in the ﬁrst and secondThe heart factorization sounds and inhibited algorithm the noise effectively with EMD, in the other the part. self The-adaptive tendency of least squares method (LSM), the ﬁtting curve was severely restricted by the points of the SVR. and the wavelet transform algorithm were employed to reduce noise in the original fetal heart sound and were compared with the de-noised results of the adaptive SVR algorithm [18,19]. In the factorization algorithm with EMD, the factorization layer was 10 and the correlation calculation was operated between the implicit matrix factorization (IMF) component after factorization and the original signal. The signal was reconstructed using half the IMF component with a large correlation coefficient. The length of the filter was 51 in the self-adaptive LSM and the learning rate was 0.0014. The wavelet basis functions in the wavelet transform algorithm were sym5 and the factorization layer was 6. The threshold was obtained using minimax methods. The coefficient was processed using the hard- threshold function and the signal was reconstructed to reduce the noise. The noise reduction results of these algorithms are shown in Figure 6.

Sensors 2021, 21, 193 8 of 12

The factorization algorithm with EMD, the self-adaptive least squares method (LSM), and the wavelet transform algorithm were employed to reduce noise in the original fetal heart sound and were compared with the de-noised results of the adaptive SVR algorithm [18,19]. In the factorization algorithm with EMD, the factorization layer was 10 and the correlation calculation was operated between the implicit matrix factorization (IMF) component after factorization and the original signal. The signal was reconstructed using half the IMF component with a large correlation coefficient. The length of the filter was 51 in the self-adaptive LSM and the learning rate was 0.0014. The wavelet basis functions in the wavelet transform algorithm were sym5 and the factorization layer was 6. The Sensors 2021, 21, x FOR PEER REVIEW threshold was obtained using minimax methods. The coefficient was processed using the 8 of 11 hard-threshold function and the signal was reconstructed to reduce the noise. The noise reduction results of these algorithms are shown in Figure6.

FigureFigure 6. Comparison Comparison of noise of reduction noise of reduction the four algorithms. of the four algorithms. These four algorithms all reduced noise to some degree. The signal-to-noise ratios (SNR) of EMD decomposition, adaptive LMS, wavelet transform, and adaptive SVR were 10.36,These 14.42, 20.64,four and algorithms 22.17, respectively. all Thereduced ﬁrst fetal heartnoise sound to couldsome be separateddegree. The signal-to-noise ratios (SNR)from the of original EMD by decomposition, the factorization algorithm adaptive with EMD. LMS, However, wavelet the details transform, of the and adaptive SVR were 10.36, 14.42, 20.64, and 22.17, respectively. The first fetal heart sound could be separated from the original by the factorization algorithm with EMD. However, the details of the second fetal heart sound were lost because of noise floor, and the noise requirement cannot be satisfied [20]. Compared with factorization algorithm with EMD, self-adaptive LSM retained part of the second fetal heart sound and reduced the noise. Both the wavelet transform algorithm and the adaptive SVR algorithm performed well and were superior to the others. These two methods followed the first the second fetal heart sounds well. However, for the low noise, the adaptive SVR algorithm performed much better than the wavelet transform algorithm. The wavelet transform algorithm fluctuated without noise, while a straight line was obtained by the adaptive SVR algorithm. Furthermore, the wavelet basis function had to be selected and the hierarchy determined in the wavelet transform, and this has a great influence on noise reduction. Although similar parameters, C and 훾, should be set in the SVR algorithm, their optimal value can be obtained through self-learning by adaptive means and can be adjusted according to the characteristics of the signal.

6.3. Fetal Heart Rate Extraction Experiment Firstly, three fetal heart rate extraction algorithms were compared in one window. Then, the actual results, which were modified by the weighted-index average algorithm between windows, were inspected. Twenty fetal heart rate data points of the same group were calculated continuously using three independent algorithms. The time of the window was 4 s and it moved forward 1 s at a time. The results are shown in Figure 7.

Sensors 2021, 21, 193 9 of 12

second fetal heart sound were lost because of noise floor, and the noise requirement cannot be satisfied [20]. Compared with factorization algorithm with EMD, self-adaptive LSM retained part of the second fetal heart sound and reduced the noise. Both the wavelet transform algorithm and the adaptive SVR algorithm performed well and were superior to the others. These two methods followed the first the second fetal heart sounds well. However, for the low noise, the adaptive SVR algorithm performed much better than the wavelet transform algorithm. The wavelet transform algorithm fluctuated without noise, while a straight line was obtained by the adaptive SVR algorithm. Furthermore, the wavelet basis function had to be selected and the hierarchy determined in the wavelet transform, and this has a great influence on noise reduction. Although similar parameters, C and γ, should be set in the SVR algorithm, their optimal value can be obtained through self-learning by adaptive means and can be adjusted according to the characteristics of the signal.

6.3. Fetal Heart Rate Extraction Experiment Sensors 2021, 21, x FOR PEER REVIEW Firstly, three fetal heart rate extraction algorithms were compared in one window. 9 of 11 Then, the actual results, which were modiﬁed by the weighted-index average algorithm between windows, were inspected. Twenty fetal heart rate data points of the same group were calculated continuously using three independent algorithms. The time of the window was 4 s and it moved forward 1 s at a time. The results are shown in Figure7.

automatic weight time windowing

direct calculation fetalheart rate(beat/minute)

time Figure 7. Comparison of three fetal heart rate extraction algorithms. Figure 7. Comparison of three fetal heart rate extraction algorithms. In order to homogenize the result dimensions, the average of the peak intervals in the windows was adopted when the fetal heart rate was directly calculated. It can be seen that the minimumIn order variation to homogenize of fetal heart rate is 15the with result the time-domain dimensions, windowing becausethe average of the peak intervals in there are 15 windows in 1 min. The effect of time-domain windowing is poor for the short thetime windows window. This was may be adopted due to that fact when that three the disturbances fetal heart occurred rate in twenty was data directly calculated. It can be seen thatpoints. the Both minimum the direct calculation variation method of and fetal the automatic heart weight rate methodis 15 performedwith the time-domain windowing be- well when the noise appeared for the ﬁrst time. The inﬂuence of the noise was weakened cause there are 15 windows in 1 min. The effect of time-domain windowing is poor for the short time window. This may be due to that fact that three disturbances occurred in twenty data points. Both the direct calculation method and the automatic weight method performed well when the noise appeared for the first time. The influence of the noise was weakened by the average effect used in the direct calculation algorithm. However, when the noise appeared for the second time, it was not completely eliminated, although it was weakened on average. The average effects will decrease with excessive noise and over longer times. The disturbance was eliminated automatically so that the automatic weight algorithm results were more stable. The automatic weight algorithm has defects and it cannot satisfy the demands of a real-time system, although it improved the performance to some degree as compared with the other two algorithms. For example, the fetal heart rate rose too quickly, which lead to a jump from the sixteenth to the eighteenth calculation. In general, the fetal heart rate varies slowly and does not jump in a short time. The jumping was caused by disturbance. The fetal heart rate data obtained using the weighted-index average method with deviation correction were further processed on the basis of this fetal heart rate characteristic. The length factor, 훽 = 0.9, and the result are shown in Figure 8.

weighted-index average

automatic weight average fetalheart rate(beat/minute)

time Figure 8. Weighted-index average algorithm.

Sensors 2021, 21, x FOR PEER REVIEW 9 of 11

automatic weight time windowing

direct calculation fetalheart rate(beat/minute)

time Figure 7. Comparison of three fetal heart rate extraction algorithms.

In order to homogenize the result dimensions, the average of the peak intervals in the windows was adopted when the fetal heart rate was directly calculated. It can be seen that the minimum variation of fetal heart rate is 15 with the time-domain windowing because there are 15 windows in 1 min. The effect of time-domain windowing is poor for the short time window. This may be due to that fact that three disturbances occurred in twenty data points. Both the direct calculation method and the automatic weight method performed well when the noise appeared for the first time. The influence of the noise was weakened by the average effect used in the direct calculation algorithm. However, when the noise appeared for the second time, it was not completely eliminated, although it was weakened on average. The average effects will decrease with excessive noise and over longer times. The disturbance was eliminated automatically so that the automatic weight Sensors 2021, 21, 193 algorithm results were more stable. 10 of 12 The automatic weight algorithm has defects and it cannot satisfy the demands of a realby- thetime average system, effect used although in the direct calculation it improved algorithm. However,the performance when the noise to some degree as compared with appeared for the second time, it was not completely eliminated, although it was weakened theon other average. Thetwo average algorithms. effects will decrease For with example, excessive noise the and fetal over longer heart times. rate rose too quickly, which lead to The disturbance was eliminated automatically so that the automatic weight algorithm a resultsjump were from more stable. the sixteenth to the eighteenth calculation. In general, the fetal heart rate variesThe slowly automatic and weight does algorithm not has defectsjump and in it cannota short satisfy time. the demands The ofjumping a was caused by disturbance. real-time system, although it improved the performance to some degree as compared with Thethe otherfetal two hea algorithms.rt rate For example,data obtained the fetal heart rate using rose too the quickly, weighted which lead to a-index average method with devia- jump from the sixteenth to the eighteenth calculation. In general, the fetal heart rate varies tionslowly correction and does not jump were in a short further time. The processed jumping was caused on bythe disturbance. basis Theof this fetal heart rate characteristic. Thefetal length heart rate datafactor, obtained 훽 using= 0 the.9 weighted-index, and the averageresult method are shown with deviation in Figure 8. correction were further processed on the basis of this fetal heart rate characteristic. The length factor, β = 0.9, and the result are shown in Figure8.

weighted-index average

automatic weight average fetalheart rate(beat/minute)

time Figure 8. Weighted-index average algorithm. Figure 8. Weighted-index average algorithm. The fetal heart rate was corrected by the weighted-index average algorithm on the basis of extraction using the automatic weight algorithm. The subtle fluctuation was eliminated from the second calculation to the twelfth and the fetal heart rate tended to rise from the fourteenth to the twentieth. The jumping fetal heart rate was weakened by the weighted-index average algorithm and the variation in the fetal heart rate became more stable. The expected effect was realized and the results were consistent with reality. 7. Conclusions A wireless high-sensitivity fetal heart sound monitoring system is composed of two parts: fetal heart sound collection and fetal heart sound processing. The reported exclusive sensor, which was constructed of copper with an enclosed cavity, was designed to prevent EMI and external noise. PVDF piezoelectric film, with a frequency from DC to 400 MHz, was adopted to convert the fetal heart sound into an electrical signal without information loss. The collected fetal heart sound was amplified and transmitted to the terminal wirelessly. The adaptive SVR algorithm, which optimizes parameters according the characteristics of the real-time system and ensures output stability, was proposed to reduce noise. Fetal heart rate extraction includes operation inside and between windows. Sensors 2021, 21, 193 11 of 12

The automatic weight algorithm was proposed to eliminate missed or incorrect detections in the windows and was more stable than the two known algorithms. The weighted-index average algorithm with deviation correction was processed to restrain the calculated fetal heart rate according to the continuity in time. The wireless fetal heart sound monitoring system can process fetal heart sound correctly and in real time.

Supplementary Materials: The following are available online at https://www.mdpi.com/1424-822 0/21/1/193/s1, Figure S1: Adaptive Support Vector Regress (SVR) noise reduction algorithm. Author Contributions: J.W. and Z.W.; methodology, J.W.; software, Z.W.; validation, J.W., Z.W. and X.X.; formal analysis, J.W.; investigation, Z.W.; resources, X.X.; data curation, Z.W.; writing— original draft preparation, J.W.; writing—review and editing, X.X.; visualization, Z.W.; supervision, J.W.; project administration, J.W.; funding acquisition, J.W. All authors have read and agreed to the published version of the manuscript. Funding: This study was funded by the National Natural Science Foundation of China through the research projects (Grant No. 61972438, 61876143). Institutional Review Board Statement: Not applicable. Informed Consent Statement: Not applicable. Data Availability Statement: Data sharing not applicable. Conﬂicts of Interest: The authors declare that they have no conﬂict of interest.

References 1. Hamelmann, P.; Mischi, M.; Vullings, R.; Kolen, A.F.; Schmitt, L.; Joshi, S.; van Laar, J.O.E.H.; Bergmans, J.W.M. Flexible sensor matrix with dynamic channel weighting for improved estimation of the fetal heart rate by doppler ultrasound. In Proceedings of the 2017 IEEE International Ultrasonics Symposium (IUS), Washington, DC, USA, 6–9 September 2017; pp. 1–4. 2. Khandoker, A.H.; Alangari, H.M.; Marzbanrad, F.; Kimura, Y. Investigating fetal myocardial function in heart anomalies by Doppler myocardial performance indices. In Proceedings of the 39th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), Seogwipo, Korea, 11–15 July 2017; pp. 2197–2200. 3. Sato, H.; Yoshimura, K.; Nakamoto, H.; Ishibashi, D.; Nakata, Y.; Yaginuma, Y.; Masui, S. 19.2 cm3 flexible fetal heart rate sensor for improved quality of pregnancy life. In Proceedings of the 2016 IEEE Biomedical Circuits and Systems Conference (BioCAS), Shanghai, China, 17–19 October 2016; pp. 140–143. 4. Andreotti, F.; Grasser, F.; Malberg, H.; Zaunseder, S. Non-Invasive Fetal ECG Signal Quality Assessment for Multichannel Heart Rate Estimation. IEEE Trans. Biomed. Eng. 2017, 64, 2793–2802. [CrossRef][PubMed] 5. Kovacs, F.; Goda, M.A.; Hosszu, G.; Telek, T. A Proposed Phonography-Based Measurement of Fetal Breathing Movement Using Segmented Structures with Frequency Splitting. In Proceedings of the 2020 42nd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), Montreal, QC, Canada, 20–24 July 2020. 6. Lakshmi, G.S.; Bhaskaran, A.; Bharadwaj, S.U.; Arora, M. Effect of Contact Force on Foetal Heart Sound Recordings. In Proceedings of the 2020 IEEE International Conference on Electronics, Computing and Communication Technologies (CONECCT), Bangalore, India, 2–4 July 2020. 7. Eiríksdóttir, D.H.; Sæderup, R.G.; Riknagel, D.; Zimmermann, H.; Plocharski, M.; Hansen, J.; Struijk, J.J.; Schmidt, S.E. Quality Assessment of Maternal and Fetal Cardiovascular Sounds Recorded from the Skin Near the Uterine Arteries during Pregnancy. In Proceedings of the 2019 Computing in Cardiology (CinC), Singapore, 8–11 September 2019. 8. Shengwei, W.; Yanni, L.; Jiayu, Z.; Jiajia, L. Agricultural price fluctuation model based on SVR. In Proceedings of the 9th International Conference on Modelling, Identification and Control (ICMIC), Kunming, China, 10–12 July 2017; pp. 545–550. 9. Mesquita, D.P.; Freitas, L.A.; Gomes, J.P.; Mattos, C.L. LS-SVR as a Bayesian RBF Network. IEEE Trans. Neural Netw. Learn. Syst. 2020, 31, 4389–4393. [CrossRef][PubMed] 10. Ni, L.; Li, J.; Lin, S.; Xin, D. A method of noise optimization for Hardware Trojans detection based on BP neural network. In Proceedings of the 2nd IEEE International Conference on Computer and Communications (ICCC), Chengdu, China, 14–17 October 2016; pp. 2800–2804. 11. Isogawa, K.; Ida, T.; Shiodera, T.; Takeguchi, T. Deep Shrinkage Convolutional Neural Network for Adaptive Noise Reduction. IEEE Signal Process. Lett. 2018, 25, 224–228. [CrossRef] 12. Wu, C.; Ma, J.; Zhang, X. User space transformation in deep learning based recommendation. J. Syst. Eng. Electron. 2020, 31, 674–684. 13. Daryanavard, S.; Porr, B. Closed-Loop Deep Learning: Generating Forward Models with Backpropagation. Neural Comput. 2020, 32, 2122–2144. [CrossRef][PubMed] 14. Zhu, H.; Wang, W.; Leung, R. SAR Target Classification Based on Radar Image Luminance Analysis by Deep Learning. IEEE Sens. Lett. 2020, 4, 1–4. [CrossRef] Sensors 2021, 21, 193 12 of 12

15. Romano, M.; Iuppariello, L.; D’Addio, G.; Clemente, F.; Amato, F.; Cesarelli, M. Computerised simulation of fetal heart rate signals. In Proceedings of the 2017 E-Health and Bioengineering Conference (EHB), Sinaia, Romania, 22–24 June 2017; pp. 185–188. 16. Fuentealba, P.; Illanes, A.; Ortmeier, F. Cardiotocographic Signal Feature Extraction through CEEMDAN and Time-Varying Autoregressive Spectral-Based Analysis for Fetal Welfare Assessment. IEEE Access 2019, 7, 159754–159772. [CrossRef] 17. Gurun, G.; Zahorian, J.S.; Sisman, A.; Karaman, M.; Hasler, P.E.; Degertekin, F.L. An Analog Integrated Circuit Beamformer for High-Frequency Medical Ultrasound Imaging. IEEE Trans. Biomed. Circuits Syst. 2012, 6, 454–467. [CrossRef][PubMed] 18. Sbrollini, A.; Strazza, A.; Caragiuli, M.; Mozzoni, C.; Tomassini, S.; Agostinelli, A.; Morettini, M.; Fioretti, S.; Di Nardo, F.; Burattini, L. Fetal Phonocardiogram Denoising by Wavelet Transformation: Robustness to Noise. In Proceedings of the 2017 Computing in Cardiology (CinC), Rennes, France, 24–27 September 2017; Volume 44, pp. 1–4. 19. Kahankova, R.; Martinek, R. Comparison of Fetal Phonocardiogram Wavelet Denoising Methods. In Proceedings of the 2018 IEEE Signal Processing in Medicine and Biology Symposium (SPMB), Philadelphia, PA, USA, 1 December 2018; pp. 1–4. 20. Wang, Z.; Wei, J.; Li, X. Adaptive SVR Denoising Algorithm for Fetal Monitoring System. In Proceedings of the 2018 10th International Conference on Wireless Communications and Signal Processing (WCSP), Hangzhou, China, 18–20 October 2018; pp. 1–5.