Flexible piezoelectric sensor for pregnant recognition based on the pulse‐taking procedure in traditional Chinese medicine

Traditional Chinese medicine (TCM) is an experienced‐based discipline and plays an important role in the current medical system. Digitalizing key features is the best way to con‐duct an in‐depth analysis of TCM. However, due to its complex generation mechanism, the crucial pulse‐taking procedure is hard to be digitalized and analyzed. This article fabricated a flexible piezoelectric sensor to gather data for the pulse‐taking procedure. The sensitivity of the proposed flexible sensor reached 0.34 V/kPa, so we adopted peripheral circuits to make its output signals stay in a reasonable range (± 0.5 V) for convenient data collection. We used the sensor to recognize whether a woman is pregnant since it has solid golden standards. The pulse signal was fused via mapping the signal spectrogram into a 3‐dimension tensor. The fused matrix was then processed by GoogleNet and reached a 90.48% correction rate. This result shows that the pulse condition has statistical differences, which solidified the objectivity of pulse diagnosis and escalated the TCM's digitalization level.


INTRODUCTION
The method and structure of traditional Chinese medicine is built upon tremendous medical practices in the past 2000 years. 1 This "practices to rules" path is similar to the research paradigm of artificial intelligence based on big data (practice data to modeled rules).In this term, with the development of information technology and advanced algorithms, the characteristics of traditional Chinese medicine are approaching objectification, digitalization, and intellectualization.
A typical diagnosis procedure in TCM involves four sections, namely, observing the patient's status, hearing the patient's sounds, asking the patient's feelings, and taking the patient's pulse.Every TCM doctor is required to master those skills.The former three sections can be smoothly converted to digital data via writing medical records or establishing imaging systems. 2,3The corresponding intellectualization can be easily obtained and has already been deeply investigated by the AI community. 4,5However, due to the special "three regions and their nine subdivisions" rules, 6 pulse taking is hard to teach and learn for young TCM doctors, not to mention for computers.The exact mechanism used in pulse taking is highly complex, and there have been many types of descriptions and theories in Chinese medical literature, but many such theories are often more subjective than objective.Thus, collecting reasonable pulse conditions requires experienced TCM physicians to find exact pulse positions (Cun, Guan, and Chi) and apply correct pressure on them.Thus, wearable sensors are highly desired without interfering with the doctor's operation.Only in this way, signals in the pulse-taking procedure can be maximumly reserved and transmitted to the computer.
Pulse taking procedure is transmitting the pulse-caused pressure into finger tactile feelings.Widely used medical sensors usually have rigid bases, so their large volumes and tough surfaces will greatly hinder the doctor's judgment of both finger pressures applied and pulse positions.4][15][16] Unlike traditional sensors, these devices possess identical interfaces to human skin, which can be integrated into the human body without causing uncomfortable feelings.According to the above excellent features, flexible pressure sensors are the proper choice for digitalizing the TCM pulse-taking procedure.
During pulse taking procedure, professional TCM physicians can feel the skin deformation caused by pulse waves.Fundamentally, the "feeling" is caused by deformation applied pressures on tactile receptors, so a flexible pressure sensor is manufactured to collect the pulse data.At present, sensors are manufactured based on two mechanisms: the piezoresistive effect and the piezoelectric effect.Piezoresistive sensors 17,18 have a simple manufacturing procedure but suffer from temperature drift and crosstalk.][21][22] On the other hand, according to the similarity of TCM and AI research, machine learning models naturally become the best choice for pulse signal recognition.Most existing studies used one-dimensional pulse signals or reduced the dimension of multichannel data to one, to identify pulse conditions. 23,24These works have certain significance and reached a moderate correction rate, but fail to conduct data fusion as our brain did, where the latter can process the three-channel data synchronously.Moreover, pulse conditions are subjective, which are not proper golden standards.
Currently, the lack of feasible flexible pulse sensors, the data fusion problem, and the missing golden standards greatly delay the development of practical flexible pulse-taking and analysis systems in TCM.Thus, in this article, we manufactured a low-noise pulse sensor with PVDF piezoelectric film and collected pulse signals from 107 women (71 pregnant and 36 non-pregnant).In this way, the golden standard of the pregnant dataset can be established by objective measurements.Then, we plotted the spectrogram for the 3-channel pulse data and mapped them into a 224 × 224 × 3 tensor.The data was processed by GoogleNet.The high result accuracy (90.48%) on the validation set proves the feasibility of the flexible pulse-taking and analysis system, which solidified the objectivity of traditional Chinese medicine and promote its digitalization level.

Material and sensor preparation
In this article, we bought the pristine PVDF powder from Arkema, Co, Ltd.The Fourier-transform infrared spectroscopy (FTIR) is conducted by Spectrum 100, PerkinElmer, Co, Ltd. to characterize the material category.As shown in Figure 1A, the PVDF powder was first solved in the solution with dimethylformamide (DMF): Acetone = 3: 1. PVDF film.We used the DMF-acetone-PVDF solution to cast PVDF film under room temperature (24 • C) and 50% moisture.The PVDF film is poled under external electric field to increase the proportion of β phase.The DC voltage we applied was 10 kV/cm, and the PVDF was stretched for 50% to facilitate the poling procedure.The temperature during preparation procedure was 60 • C. The sensor is a sandwich structure (electrode/PVDF/electrode) generated by screen printing and it was encapsulated by thermoplastic polyurethanes (TPU) for further protection, as shown in Figure 1B.Moreover, in Figure 1C, the PVDF sensing film can be bent by a tweezer, proving its flexibility.After the calibration, the sensor (12 mm diameter, 28 μm thickness) was mounted on a rubber finger cot (Arukuhito Co, Ltd) to facilitate the application, as shown in Figure 1C.The electro-mechanical property of the sensor is tested in three aspects to show its sensitivity, stability, and resolution.During TCM pulse taking procedure, experienced physicians often use their fingertip centers to feel the pulse.Thus, to sense identical signals as the doctor does, the sensor center should overlap with the fingertip center.Since the center point of a circle is the easiest to find, we used this kind of symmetric layout to form our sensor shape.

Peripheral circuit
The sensor size in this article is small (12 mm diameter), so the voltage generated by pulse pressure may be wrongly sampled by the 12-digit internal ADC in STM32 (MUC resolution is approximately 3.3/4096 ≈ 8 mV).To deal with this problem, we adopted a charge amplification circuit as the front module to convert charge variation to voltage changes.
Additionally, an active low-pass filter is applied for eliminating the interference from the 50 Hz working frequency without attenuating the <10 Hz pulse frequency component.To achieve a convenient application procedure, we used a ICL7660M chip to generate -5 V by normal DC power source in upper computer.The modified signal was sampled by the internal 12-bit internal ADC in STMF103C8T6 and then sent to the upper computer for advanced analysis.The oscillator was set as 3.2768 kHz.Moreover, the maximum frequency component of human pulse signals was around 30 to 40 Hz.According to Nyquist Sampling Theorem, the sampling rate can be no less than 60 Hz.Thus, we used a 200 Hz sampling rate and cut 4096 points (about 20 s) to reserve sufficient information.The system architecture is shown in Figure 2A.

Pulse-taking experiments
With the developed system in this article, we collected the pulse waves of 36 healthy non-pregnant volunteers from staff and interns in Tongde Hospital of Zhejiang Province.On the other hand, 71 healthy pregnant women were recruited for pulse signal acquisition at the obstetric clinic of Tongde Hospital of Zhejiang Province.Ages of all participants are controlled within the range of 24-34.Before the experiment, volunteers were abstained from alcohol, caffeine, and all kinds of drugs.Pregnant women with multiple pregnancies, abnormal menstrual cycles, chronic hypertension, diabetes, anemia, and any other known diseases during pregnancy were excluded from this study.
In the beginning, participants were asked to sit down and keep silent to adjust their breathes for 5 min first.Then, the TCM doctor assessed and collected the pulse waves in Cun, Guan, and Chi places of the radial artery for 60 s by the flexible piezoelectric sensor.To eliminate the effect of transient status, we choose the data from the 20-40 s range to form the data sample for the latter data processing procedure.Data collection took place at a TCM laboratory in the Tongde Hospital of Zhejiang Province from June to October 2021.The laboratory temperature was kept at 22 • C throughout the The pulse signal sensing system and the collected signal: (A) the peripherals required during the pulse taking procedure; (B) one sample of the collected 3 channel pulse signals corresponding to Cun, Guan and Chi study period, as fluctuations in ambient temperature are known to affect the pulse. 25One sample of the collected signal dataset is shown in Figure 2B.
This study was approved by the Ethics Committee of Tongde Hospital of Zhejiang Province.All experiments were performed in accordance with the Declaration of Helsinki.Each volunteer understood the purpose of this study, and they agreed to take part after reading the Participant Consent Form.During the experiment, the device is not directly attached on human skin but a finger cot, where the doctor can wear it during diagnosis and take it off after the pulse taking.After wearing the functionalized finger cot, positions of Cun, Guan and Chi will be found following the doctor's experiences.To be mentioned, all participants have signed their names on the informed consent before inclusion in the study.

Data fusion and analysis
The diagnosis procedure of traditional Chinese medicine is similar to data-driven artificial intelligence.Thus, we chose to use the highly nonlinear deep neural network (DNN) for the pulse analysis and recognition task.Pulse signals are parodic signals, so their pattern can be found by mapping into an image with a proper size. 26Meanwhile, for parodic signals, the frequency domain can provide more intuitive information.Thus, we adopted the continuous wavelet transformation (CWT) to form a spectrogram so that information from both the time domain and frequency domain can be displayed.The CWT can be conducted by: where a and b are the scale and shift factor, respectively, and the star represents the complex conjugate.In 1-dimensional transformation, x is the time, f (x) indicates the signal value, and (x) represents the wavelet basis.To reserve feature as much as possible, we adopted a Morse wavelet during CWT 27 : in which the a , act as normalization value, and U() is the unit step function.The CWT performance can be optimized by adjusting b and r, where we set these values as 3 and 60, respectively.Moreover, since the pulse data contains three channels, we can naturally map spectrograms from Cun, Guan, and Chi toward red, green, and blue (RGB) channels.The merit of this method is that it provides a high-dimensional way to preserve original signal information, which can reduce the involvement of expert experience to its minimum.Three spectrograms from Cun, Guan, and Chi are shown in Figure 3A.
In terms of data analysis, GoogleNet is a novel convolutional neural network (CNN) structure with highly accurate output.It is a multilayer structure consisting of convolution layers, pooling layers, inspection layers, full connection layers, and output layers.Compared with images in real life, pulse signals contain far less information, and the classification task can be sufficiently conducted by GoogleNet.Moreover, to fit the input requirement of GoogleNet, we reformed spectrograms as three 224 × 224 gray figures.Then, we stack three spectrogram in on tensor and get a 224 × 224 × 3 RGB image, which is the input form of our neural network.The mapping procedure is show in Figure 3B.

Characterization results
The FTIR results of the sample are shown in Figure 4A.The absorption peaks of the infrared spectra at 762 and 974 cm −1 are clear features of α-crystal shape.Features of β-crystal shape can be observed at 842 cm −1 .These two shapes indicate that we used the correct type of PVDF powders.The SEM image of the screen-printed electrode is shown in Figure 4B.The flat and dense surface indicates that the silver paste is successfully cured without leaking points.In Figure 4C, the SEM image of the film cross-section confirmed the electrodes are highly combined so the transmission quality of the piezoelectric signal will not be deteriorated in this part.Moreover, the sensitivity test in Figure 5B shows that the gauge factor of the PVDF-based flexible pulse sensor reached 0.34 V/kPa.The repeatability depicted in Figure 5A guaranteed the pulse data stability during multiple diagnosis procedures.Moreover, the stable outputs in Figure 5C under different external pressure indicate an excellent resolution.The stably raised step height indicates a high linearity in the pressure-voltage relation and will cause a small error during the pulse-taking test.

Data analysis result
In this article, we adopted GoogleNet 28 as the backbone network, which consists of two groups of convolutional and max-pooling layers that have 7 × 7 and 3 × 3 kernel sizes, respectively, for extracting initial features, nine inception modules for updating intermediate features, and a fully connected layer with a softmax for predicting the pregnant and non-pregnant status.In specific, each inception module contains four branches: ( The structure of the adopted GoogleNet and its whole training procedure in this article.As shown in Figure 7, after updating the network parameters of 100 epochs, the training loss was optimized toward the minimum in terms of 0.1283, and the validation loss was converged in the same region as well.Such results indicate that no over-fitting occurred in the training phase, resulting in feasible recognition results.The correction rate of our network reached 90.48% and 98.8% in the validation and training sets, respectively.The above-mentioned results proved that the pulse conditions have statistical differences, and with a large enough dataset, the intellectualized TCM diagnosis has its future.

CONCLUSION
This article provided a way to fabricate a flexible piezoelectric sensor system and functionalized it in the pulse-taking monitoring during TCM diagnosis.The sensor was integrated into a finger cot to make the doctor's operation convenient.Meanwhile, the sensitivity of the prepared sensor was 0.34 V/kPa, and a peripheral circuit is further developed to enhance its ability to monitor small deformation-caused pressures.To fabricate reasonable inputs, spectrograms for the collected pulse data were mapped into a 3-dimensional image tensor to maximumly reserve both the time and frequency information.The recognition procedure was conducted by GoogleNet, and the correction rate reached 90.48%, which demonstrated the feasibility of the flexible sensing system in pulse signal collection and analysis.Moreover, the machine learning result is sufficient to prove the significant differences of pulse conditions between healthy pregnant women and healthy normal women, which provided another solid proof for TCM pulse diagnosis.However, since unhealthy women are prone to avoid pregnant, pregnant participants with diseases are rare and hard to recruit.

F I G U R E 1
The preparation and image of the proposed sensor: (A) The preparation procedure of PVDF solution; (B) the fabrication of the sensor by screen printing; (C) the image for the sensor integrated with a finger cot.The PVDF sensing film can be bent by a tweezer, so it is a flexible sensor.Moreover, the finger cot is mounted on the doctor's fingers so he can find the location of Cun, Guan and Chi for different individuals.

F I G U R E 3
The data fusion method for the 3-channel pulse signal: (A) the spectrogram for each signal channel was mapped in R, G or B layer; (B) the layers containing time-frequency information was combined as a 224 × 224 × 3 RGB image.

F I G U R E 4
Sensor characterization: (A) the FTIR result of the PVDF powder; (B) the SEM image for the silver electrode surface; (C) the SEM image for the interface between the electrode and the PVDF layer I G U R E 5 Electro-mechanical performance of the sensor: (A) The stability test for the PVDF sensing material during 200 cycles; (B) the sensitivity test for the sensor, where GF is the gauge factor of the sensor; (C) the pressure-voltage relationship der different loads (1) 1 × 1 convolution; (2) 1 × 1 and 3 × 3 convolutions; (3) 1 × 1 and 5 × 5 convolutions; and (4) 3 × 3 max-pool and 1 × 1 convolution.This network is implemented by MATLAB 2022 (CUDA) acceleration with NVIDIA GeForce RTX 3060.The training epochs were set as 100 with the Adam optimizer, and the cost function was defined as the binary cross-entropy loss for all samples.The detailed architecture of GoogleNet is shown in Figure6.As for the activation function, a rectified linear unit (ReLu) is chosen to solve the gradient vanish and long training period problem: ReLu(x) = max(0, x).

F I G U R E 7
The performance of GoogleNet on the fused 3-channel pulse dataset: (A) the loss reduction during the 100-epoch in both validation and training set; (B) the correction rate during the 100-epoch iteration in both validation and training set During the training process, traditional stochastic gradient descent (SGD) can easily be trapped in the local minimum if the hyperparameter is badly chosen.Thus, this article adopted an adaptive moment estimation (Adam) to generate different learning rates dynamically for different parameters.In this way, the cost function can converge faster without being trapped in local minimums.