Portable Face-Shielding Device Based on sEMG Considering the COVID-19 Scenario

Wearing a mask greatly reduced the possibility of infection during the COVID-19 pandemic. However, major inconveniences occur regarding patients with upper limb amputations, as they cannot independently wear masks. As a result, bacterial contamination is caused by medical staff touching the quilt when helping. Furthermore, this effect can occur with ordinary people due to accidental touch. This research aims to design an automatic and portable face shield assistive device based on surface electromyography (sEMG) signals. A concise face shield-wearing mechanism was built through 3D printing. A novel decision-making control method regarding a feature extraction model of 16 signal features and a Softmax classification neural network model were developed and tested on an STM32 microcontroller unit (MCU). The optimized electrode was fabricated using a carbon nanotube (CNT)/polydimethylsiloxane (PDMS). The design was further integrated and tested, showing a promising future for further implementation.


Introduction
In recent years, the COVID-19 epidemic has been rampant. As a result, face shielding has been efectively used to avoid the spread of the corresponding virus [1]. In the absence of a shelter-in-place strategy, infections, hospitalizations, and deaths were reduced by 37.7%, where the interquartile range (IQR) was 36.1-39.4%, 44.2% (IQR: 42.9-45.8%), and 47.2% (IQR: 45.5-48.7%), respectively, as nonmedical masks were worn by 75% of the population reduced [2]. Te use of general facemasks has signifcant benefts, especially if they are adopted earlier, and at least some beneft is realized across a range of epidemic intensities. Furthermore, masks can be used in addition to other interventions, such as social distancing and hygienic measures, ultimately resulting in a nonlinear decrease in epidemic mortality and healthcare system burden [3]. During the COVID-19 outbreak, many new developments related to healthcare utilized electromyography (EMG) or electroencephalogram (EEG). Setiawan et al. proposed a stroke rehabilitation monitoring method using EEG [4], Bano and Hussain could identify patients that were infected by COVID-19 using the EMG technique [5], and L. Jiang et al. developed an advising device for at-home exercise using sEMG [6].
Some disabled and paralyzed patients, however, who have both arms amputated cannot wear face masks independently. According to the latest survey reported by China Disabled Persons' Federation, in 2010, China had 20.54 billion amputation patients [7], of which 2/3 were upper limb amputees [8]. In addition, there is a nonnegligible group that has a severe impairment of both upper-and triplelimb functions [9]. However, assistance in public places increases the risk of viral infection. Furthermore, crossinfection in hospitals when nurses take care of paralyzed patients and manually wear face shields can also occur. [10] Nevertheless, patients who cannot or have difculties putting on masks by themselves arrive at hospitals without wearing masks or need to call a nurse for assistance at their home. As a result, the person can get infected if the nurse or people in the hospital are recessive carriers of COVID-19.
Many casual activities, such as dining and drinking, require that masks are constantly put on and of. However, some patients cannot be assisted by nurses all day. In some public places, directly touching the inside of a face shield by hand can contaminate this equipment, reducing its protection against infectious diseases and threatening health. We also believe that a new hand-controlled face-shielding device could be convenient for ordinary people. Hence, a noncontact wearable face-shielding-wearing system is proposed.
Heckendorn [11] and Backers Today [12] separately developed two automated face-shielding-wearing systems controlled through buttons. Te automatic face shielding of Heckendorn is promising due to its mechanical structure and portability. Furthermore, the one designed by Backers Today is convenient for drinking water. Tese designs, therefore, reduce the possibility of contaminating face shields and are conducive for people with disabilities.
Huang developed an automask utilizing an ultrasonic sensor and an sEMG sensor that allowed the wearer to actuate when sneezing by contracting the abdominal muscle while manually covering the ultrasonic sensor [13]. Tis design could detect and avoid the sneeze spreading; however, the situations in which such devices can be applied are limited. Furthermore, the mask is not suitable for patients with upper limb amputations, as it can only be actuated by contracting the abdominal muscle while covering the ultrasonic sensor by hand. Te utilization of an HC-SR04 ultrasonic distance sensor would signifcantly harm the cost efciency of the design. Te HC-SR04 module is an active measurement instrument [14] that continuously emits an echo pulse at a 40 kHz frequency [15]. Owing to the active detection limit, the power consumption of the HC-SR04 module is incessant during wearing. In contrast, the sEMG module is a passive measurement instrument whose power consumption is much lower; therefore, it is more suitable for casual wearing. Moreover, the ultrasonic distance sensor can be accidentally triggered from the outer environment, such as the clothing of the user or the impact caused by passing-by pacers. Furthermore, the signals of ultrasonic distance sensors commonly fuctuate (fuctuation phenomenon) [16], even though a threshold value is set to avoid accidental triggering. In addition, the control algorithm of the design involved no signal classifcation, increasing the risk of accidentally triggering the mechanism. Te actuation and control of the servo motor without a signal classifcation are usually achieved by manually setting the threshold of the triggering voltage amplitude. [15] However, unwanted spike signals that trigger the actuators can be generated. For example, accidental pressing on the sEMG sensors or external force-induced position shift of the electrodes would create spikes and noise. Furthermore, the fuctuation phenomenon of the ultrasonic distance sensors contributes to more accidental actuation in Taliyah Huang's design. Finally, controlling the mask based on the abdominal muscle is not feasible, as the body fat at a selected abdomen position signifcantly varies depending on the person; therefore, the EMG signal cannot be commonly applied. Diferent factors, such as epidermal body fat, signifcantly infuence the sEMG measurement, [17]. As lipids easily accumulate in the abdomen, the abdominal muscle EMG could be weak to be detected when facing thick abdominal fat tissue.
To solve these problems, we designed a portable, noncontact, and automatic face-shielding-wearing device that considers the EMG signals.

Background Knowledge
An electromyography instrument is used in EMG to record muscle bioelectric signals. Tese electrical signals can describe some characteristic behaviors of the human muscles. EMG signals can be used in many felds, such as clinical applications [18] and human-computer interaction [19]. More direct monitoring of the human body can be achieved through the analysis, processing, classifcation, and various signal analysis methods of these signals [20]. New implementations of EMG, such as muscular paralysis disease prediction using EMG, have been developed during the pandemic [21].
However, people who are inconvenienced by wearing masks often need to adapt themselves to EMG devices for a long time. Commercial electrodes are generally Ag/AgCl electrodes; however, these components should be avoided, as they cause damage to human skin [22]. As a result, we used carbon nanotubes, which are nontoxic, harmless, and pollution-free [23], as the primary material for the electrode, increasing its stronger skin afnity and allowing it to be more suitable for long-term use.
Te proposed device collects and analyzes the human body EMG signal, realizing thereafter, through machine learning, the wearing and removal of the face shield according to the human body EMG signal. As a result, it does not need to be manually controlled. Tis study used the specifc location of the muscle signal of a dual-channel sEMG detector. We also improved the electrode material in contact with the human body, resulting in a more suitable material for long-term use. Based on the corresponding signal feature processing and machine learning classifcation, noncontact automatic face shielding and wearing could be realized.

System Architecture and Methodology
Te main objective of this study was to design a portable face-shielding device based on sEMG. Te user of the proposed device could wear/undress the mask by contracting two specifc muscles, fulflling the noncontact demand for paralyzed patients with amputated upper limbs or ordinary people, therefore avoiding the accidental touch of the inner surface of the face shield. By selecting two independent muscles that do not usually form a synergic pattern, the device can be precisely actuated based on the corresponding gesture, thereby avoiding accidental triggers. To realize this purpose, at least three primary components were involved: data acquisition units, portable data processing units, and automatic face-shielding mechanisms.
A data fow chart of the designed system, composed of these primary components, is shown in Figure 1. We use a 2 MyoWare ™ Muscle Sensor (AT-04-001) [24] to achieve bichannel input of the sEMG signals, as the EMG signals of two independent muscles were detected. A hardware bandpass fltering circuit and rectifying and amplifying procedures were integrated into the MyoWare module. Te analog sEMG signal was further converted to a digital signal through an analog to digital converter (ADC) on a STM32F103ZET6 microcontroller unit (MCU), which is a portable data processing unit. Based on the received data, our algorithm yields a decision that fulflls the user demand. Accordingly, the triggering signal is sent to the automatic face-shielding mechanism via IIC based on the corresponding gesture.
To increase the dermotropic property and electrical conductivity of the electrodes, we have investigated the improvement of the electrode material. Te mechanical structural design targets a lightweight and conciseness compatible wearable mechanism. As shown in Figure 2, the model training of the EMG data fow can be divided into an ofine data training process and a real-time model training.
In the ofine model training procedure, the data were, frst, preprocessed. Afterward, serial feature extraction based on the segmented data batches was conducted. Te classifcation process was based on artifcial neural networks (ANN) in this research.
Models are obtained through ofine training, regardless of the real-time control of the mechanism and considering datasets stored on a personal computer (PC). Further development of these idealized models needs to be achieved and calibrated in the real-time training procedure. Teir parameters were, therefore, carefully adjusted based on the actual performance of the device to achieve the optimized conditions.
Our experimental setups of the data fow model training are shown in Figure 3, which indicates the key components for the feature extraction model and the classifcation neural network training. Te red arrows represent the power supply directions; the blue lines stand for data fows from the peripheral devices to the MCU. In contrast, the yellow arrows represent the opposite data fow. Both data fow represented by the yellow and blue arrows should be omitted during the real-time optimization to achieve the portability of the design. Te green and grey arrows indicate the data fow between the MCU and peripherals.
To illustrate the hardware selection, two MyoWare ™ muscle sensors (AT-04-001) were selected for the sEMG bi-channel sEMG module data processing unit peripheral mechanism feature extraction classifcation actuation data acquisition bandpass fltering  Journal of Healthcare Engineering signal detection. A 32-bit microcontroller composed of an ARM Cortex-M3 core STM32F103ZET6 [31] MCU module was selected to execute the computational tasks of this portable device, owing to its computational performance and stability. A push button, which was pressed to indicate the need for the shielding mask, was used for signal labeling. A 5 V polymer battery was needed for the power supply. Te PCA9685 module is a 16-channel PWM output motor control module [32]. In this study, we realized control of a FUTABA Corp. S3003 servo motor through IIC communication and PCA9685.

Design of Mechanism
Te mechanical design of the face-shielding device should promote function, reliability, and economic characteristics. To achieve a functional design, the basic motion form, as well as the main measurement of the device, should be analyzed based on practical applications. After completing the overall design of the device, which is usually idealistic and abstract, its reliability must be evaluated; therefore, specifc parts, such as rockers and connectors, should be selected, and mechanical calculations and analyses are necessary to ensure the feasibility of movement. Certain elements might also be modifed and optimized. Finally, the material and manufacturing process are were selected based on a limited budget.
Our face-shielding device must successfully take the face shield on and of. To simplify this situation, it can be assumed that the shield is moving around a fxed point. As this is the only movement required, the device can be designed as a planar mechanism, as shown in Figure 4(a). To further simplify the system, a rigid four-bar linkage was used as the basic structure.
As the only movement of the face shield is rotation, the mechanism has only one degree of freedom (DOF). To fulfll this requirement, a parallelogram form of the four-bar linkage was used, as shown in Figure 4(b)). Based on this adaptation, the rotary motion of the driver bar at the driven bar is duplicated, therefore creating a stable parallel relationship.
Te planar DOF is governed by the following equation: where M is the DOF, L is the number of links, and J 1 and J 2 are the number of lower and higher pairs, respectively. In the parallelogram linkage, where one bar is already fxed, L � 4, J 1 � 4, and J 2 � 0; therefore, M � 1 based on equation (1). Tis value of DOF indicates that the four-bar linkage of the parallelogram fulflls the requirements of this study. Te concept of the planar mechanism was, thereafter, further refned into a 3D mechanism by adding two identical four-bar linkages connected by an axis, therefore completing the structure that could be fxed on the neck of the subject and achieving a rotary motion. Te dimensions of the structure were based on the human face and actual movement. Figure 5(a) demonstrates the 3D version of the device constructed through SOLIDWORKS 2020. Two bases were placed on either side of the neck of the subject, and a servo motor was installed to drive rocker1, allowing it to complete the rotary motion. Te face shield was placed between connecting rod3 and rod4.
However, according to our requirements, bar2 should be fxed to restrict the DOF to one. Moreover, the joints on which the connectors should be used have not been determined yet. As the stability of the structure must be ensured during operation, the movement of the bars moving along the axle cannot be executed. As a result, based on the  reliability analysis, the joints were connected by bolts and nuts, a new base was designed to restrict the movement of bar2, and the diameters of the two axles were modifed to ensure the axial constraint. Te optimized mechanism is shown in Figure 5(b)).
In additive manufacturing (3D printing), simulations can be directly transformed into real objects, resulting in a swift process. Te simulation in SOLIDWORKS was already implemented; therefore, due to the principle of economic design, additive manufacturing was selected as our manufacturing method. Te material that has been frequently used in this process is resin, as it demonstrates suitable rigidity for the rotation of light objects, such as face shields. Te accuracy of this method, however, is not satisfactory yet, which may result in higher dimensional and geometric tolerances in each part.
Te proposed mechanism provides, therefore, a relatively efective procedure for taking the face shield on and of without requiring a complex structure. Furthermore, the material selected for the face-shielding device is light and the rigidity against the torque generated by the servo motor is sufcient, resulting in less stress for the subjects when using the device. However, the stability of the entire mechanism is not perfect owing to its simple structure, therefore requiring future improvements. Moreover, the mechanism might wear after a long-term operation due to the lack of protection.

Background.
Te standard sEMG electrode commonly used in experiments is usually composed of gold (Au), silver (Ag), or silver-silver chloride (Ag/AgCl). Ag/AgCl is the most widely used material, as it has a lower baseline-noise interface than that of the other electrodes. Most Ag/AgCl on the market is present in the form of gel, for better maneuverability. Tis material can also be used in the surface coating of the electrode to reduce surface impedance and improve signal quality [33].
However, Ag/AgCl electrodes have several problems. Te proposed equipment must be designed considering that people can wear it for a long time; however, the long use of an Ag/AgCl electrode can easily stimulate the skin and afect (2) the signal quality after drying [34]. As a frequent replacement of new electrode sheets and stimulation of human skin are not considered in our design, an electrode that can be worn for a long time was developed [35].
Owing to the good dispersibility of carbon nanotubes (CNTs) and the fexibility of polydimethylsiloxane (PDMS), CNT/PDMS composites are considered good fexible conductors. Te nanotubes dispersed in the electrode were in good contact with each other, forming a conductive path. Conductivity tests showed that the electrical performance depends on the concentration of carbon nanotubes. Furthermore, electrode materials are safe for biomedical applications, as they are not easily afected by sweat on the surface of the human body [36].

Material Parameters.
Te CNT used in the electrode was manufactured by Suzhou Tanfeng Tech. Inc., China. Te purity and length of the carbon nanotube were over 95 wt% and 3-12 μm, respectively. Te multiwall CNT was manufactured through vapor deposition. PDMS was manufactured by Shenzhen Xinwei New Material Co. Ltd. by using 909 potting glue.

Fabrication of Electrodes.
Te electrode was composed of an insulating layer, conductive flm, and button electrode, as indicated in Figure 6.
Te CNT was dispersed into the PDMS precursor (viscosity: 4000 ± 500,909-A, Xinwei New Material Co., Ltd.) and mixed through dispersion. A crosslinking agent (viscosity: 100 ± 10,909-B, Xinwei New Material Co., Ltd), whose ratio was 10 : 1, was solidifed for 12 h at room temperature. When this agent could not solidify, the button electrode was embedded in the colloid and the solidifcation process occurred for 24 h. A 10 : 1 ratio of PDMS was used on the periphery, forming an insulating layer to facilitate contact between the electrode and the skin. Table 1.
As shown in Table 2, the specifcations of the electrode sheet must be defned to enable the usage of the equipment for a long duration. Te diameter of the conductive electrode part of the CNT/PDMS electrode sheet used in the experiment was 10 mm, which is a small part of the entire electrode. Te conductive electrode (main body of the electrode), whose length and width were 36.5 mm and 30.2 mm, respectively, was composed of a PDMS solid gel. Te overall thickness of the electrode sheet was 2 mm.

Electrical Test and Specifcation.
Te impedances of the electrodes were measured using an electric meter. Table 1 shows a test comparison diagram of commercial Ag/AgCl electrodes and CNT/PDMS electrodes considering the same specifcations.
Te transparent material in Figure 7 is the PDMS patch, and the black component corresponds to the CNT/PDMS electrode. Te button electrode was connected on top of the CNT/PDMS to obtain an electrical signal from the device.

Electrode Positioning and Muscle Selection
Te muscles were selected in this study by mainly two criteria: (1) if they are independent of each other; therefore, they ought not to collaborate to execute the same gesture, to reduce the possibility of accidentally actuating the mechanism; (2) if the selected gesture pattern is easy to achieve. Furthermore, face shielding should not afect the feasibility and regularity of the selected pattern.
Te position and orientation of the muscle sensor electrodes signifcantly afect the strength of the signal. Owing to the diferential detection mode of the sensors, two detection points and a reference position should be defned. Te electrodes should be placed in the middle of the muscle body and aligned with the muscle fbers [25]. Meanwhile, the reference electrode should be placed on a separate section of the body, such as the bony portion of the elbow or a nonadjacent muscle. Due to the size of the MyoWare ™ muscle sensors (AT-04-001), the sensors of the system are closely positioned (C-position). As a result, the diferential voltage between the two points in the middle of the muscle body could be detected [37].     Figure 8 shows the positions of the electrode placement based on the muscle. Te green box indicates the correct positioning of the electrodes (midline of the muscle belly between the innervation zone and myotendinous junction) [24].
After comprehensive consideration of the two major factors afecting the positioning of the electrodes, the gesture pattern was defned based on the zygomaticus and bicipital muscle of the arm. In particular, these muscles were selected as they can be easily contracted and their gestures can be executed by diferent users, such as patients with upper limb amputation and ordinary people. Te gesture pattern was defned as a signal trigger when both muscles simultaneously contracted. Te detailed placement of the electrodes is shown in Figure 9. To preserve the accuracy and repeatability of the experiment, the contour of the electrode was marked by using a luminous pen after each test.
After attaching the electrodes to the surface of the muscles, a bichannel ADC function was completed on the MCU utilizing a timer interrupt and setting a sampling rate of 100 Hz. Te sampled data were stored in an array for further processing.
A fxed-length segmentation method [38] was applied in this study. Te windowing length was restricted to the performance of the STM32 MCU. Te coding process of the MCU was conducted using Keil MDK v5.26 Software. Only 64, 256, and 1024 points of fast Fourier transformation (FFT), which is a method to analyze the power-related characteristics of signals from the aspect of the frequency domain, were allowed on the F1 series of the STM32 [39]. Time-frequency analysis is an important key to EMG signal feature extraction [40,41]. Due to the selected sampling rate and the requirement for a quick response, the FFT of 64 points was used in this study. To date, a windowing of 64 data points per batch has been set for feature extraction, indicating that the response time for pattern recognition is approximately 80 ms, based on a calculation machine cycle (MC).
Te acquired voltage amplitude data of the two inputs and labeling data of the push button were sent from STM32F103ZET6 to the PC using a universal asynchronous receiver/transmitter (UART) and stored in text fles (.txt). Te fles were later processed using MATLAB, as shown in Figure 10, where the vertical and horizontal axis indicates the amplitude and the timeline, respectively. Te peaks in the fgure are spiky as this dataset was composed of over 15000 points of data, resulting in a contraction in the horizontal display of the graph. Te data sent to MATLAB were raw data that were not binned based on 64 batch sizes, as the segmentation of the data was performed via MATLAB2019a in this simulation stage.
During the feature extraction step of this study, the model was frst simulated on a PC using MATLAB2019a to test the proposed concept. Afterward, the model was coded again using the C language and executed on MCU.
In this study, 16  Equation (2) indicates the IEMG defnition [42], where n is the batch size of the segmented data (64) and X i represents i th data in the dataset. Te IEMG is the sum of the absolute value of the EMG signal, which can be treated as a signal power estimator [38].
Te MAV, MMAV1, and MMAV2 can be defned by equation (3) [43], equation (4), and equation (5) [44], respectively, where w i is the i th modifed weight of the mean absolute value, which is an important characteristic of the sEMG signals that are based on the time domain.
MMAV1 � 1 n n i�1 w i X i , Te MAVS, which indicates the variation between two adjacent batches, can be correctly represented by equation (6) [38], where k represents the k th batch in the dataset.
Te RMS can be described by equation (7) [39]. Tis quantity is modeled as an amplitude-modulated Gaussian random process whose RMS is related to the constant force and nonfatigue contraction [38].
Te VAR can be expressed by equation (8) [42]. Tis quantity indicates the diference between the signals in the dataset.
Te WL is defned by equation (9). Tis value, which is the cumulative length of the waveform over a segment, is infuenced by the waveform amplitude, frequency, and duration [43].
Te threshold crossings (TCs) are defned as the instances of the recorded data that crosses a threshold value during one batch.
Equation (10) represents the WAMP [39], which is the cumulative length of the waveform over a segment. Tis quantity indicates the number of times that the diference of consecutive amplitudes exceeded a predetermined threshold, ℓ.
Equation (11) represents the SSI. Tis quantity is the cumulative length of the waveform over a segment. It also represents the accumulation of the bathed data power [38]. Equations (12) and (13) indicate the FMN [38] and FMD [44], respectively. Tese quantities use FFT to calculate some properties regarding the power spectrum density (PSD). However, the Pwelch values must be obtained for the calculation of the PSD in MATLAB. Tis is a difcult task to be executed using an MCU, as similar results are difcult to be obtained. Hence, in this study, we obtained a resemblance [45] between the Pwelch method and the PSD calculated using FFT through equations (14) and (15), respectively. Tis procedure was added to the MCU using the built-in FFT library from the STM32 DSP.
Te MFMN and MFMD were proposed by Phinyomark et al. [44]. MFMN is calculated as the sum of the product of the amplitude spectrum and frequency divided by the sum of the spectral intensity. MFMD is defned as half of the sum of the amplitude spectrum [38]. As a result, the amplitude spectrum of the signal can result in a volatile FFT.
Te frequency ratio (FR) [46] is calculated in equation (16), where fft max and fft min , respectively, represent the maximum and minimum values of the dataset after the fast Fourier transformation.
Te fag label (FL) shown in equation (17) indicates the decision after evaluating the received data batch. When the dataset includes more than three sampled data points that are labeled, the entire batch is considered for the demand of the decision to trigger the actuator. Tis feature was omitted in practical scenarios, and it is only available for real-time calibration of the neural network classifer.
Te feature extraction was frst built using MAT-LAB2019a considering a raw dataset stored in txt fles. Each fle contained more than 15000 samples of data. Sixteen feature extractions were applied for each channel of the sEMG signal, resulting in 32 features and one labeling data. Tese data were stored in csv fles and were used, thereafter, in the classifcation stage.
After the ofine data training procedure, the feature extracting model was coded again using the C language and downloaded into the MCU. Te FFT was obtained by using the ARM_math and ARM_cFFT_Radix4 libraries [47]. 32 features were extracted from the two input channels. Te output of the feature extraction was validated using a COM helper application and a JLINK debugger.
Te Softmax classifer has outstanding performance when treating a multiclassifcation problem, as all types of features are normalized according to the number of classes. Furthermore, this classifer can clarify positive features [27]. In the binary classifcation problem in this study, the device is required to distinguish the triggering signal from the background noise, as the same triggering gesture pattern is used for the masking and unmasking actions. In this case, a Softmax classifer is adequate to solve this problem, as it yields fewer parameters than that of the other classifers, owing to its simplicity. A large number of parameters would slow the computation speed and consume much fash memory of the MCU. In addition, the increase in complexity of the neural network results in the increase in difculty to rewrite the model on a portable device.
We constructed the binary Softmax classifer using the 2.3.1 framework backend on TensorFlow 1.13.1. Te IDE program on which the algorithm was developed was Pycharm 2020.3-Python version 3.7. No GPU was required, as the network structure and data type were not complicated.
For dataset recording, the person that tested the device was placed in a comfortable position and the electrode was frmly attached to the selected zygomaticus and bicipital muscles. A series of random movements of the limbs and changes in the countenance were performed while periodically executing the triggering pattern (simultaneous contraction of the zygomaticus and bicipital muscles). Whenever the targeting pattern was executed, the subject was requested to press the push button at the same time that he labels the corresponding data batch. Troughout the ofine model training stage, the features were extracted using MATLAB and stored in csv fles, which were, thereafter, used in the training of the neural network. However, during the real-time optimization of the model, the features were directly extracted by the portable data processing unit and sent to the PC using UART. Tese data were again stored in txt fles, which were, thereafter, converted into csv fles for calibration. As a result, a more realistic dataset was acquired, and the Softmax classifer could be calibrated to achieve its best performance.
Te accessed data were shufed and, thereafter, divided into a training dataset and the test dataset using the K-fold method from the sklearn library. Meanwhile, the label value Journal of Healthcare Engineering was converted into a one-hot code using the built-in function of Keras.
Te structure of the constructed binary Softmax classifer is shown in Figure 11. Overall, the classifer was composed of one input layer, one hidden dense layer, and one output layer. Te input layer had a 1 × 32 input dimension and received a fattened array of 32 features related to the bichannel sEMG input, which could be fed to the network in single frames or catches. Te hidden dense layer of the classifer outputted an array of data size 1 × 64. A drop-out method, whose ratio was 0.02, was introduced to the network to perform a regularization task, preventing the overftting of the classifer. Te output dimension of the network was the same as the class number (2).
For the neural network training process, 40 epochs of 50 batches for each step were implemented. A categorical loss function was used for the loss assessment. Te Adam algorithm was implemented for optimization.
Te performance analysis of the classifer is shown in Figure 12. Figure 12(a) shows the training and validation accuracy of the network. Te average accuracy of 97.7% was achieved, and no signifcant diference was noticeable between the accuracies of the training and validation models, indicating that no overftting or owe ftting occurred. Figure 12(b) displays the loss of the network during the training and test stages. Both losses converge to a single value.
Te weights and biases of the trained model were, afterward, converted into array metrics. Te copied network of the Softmax classifer was reconstructed using the C language, and the parameters were included. Te resulting neural network was repeatedly calibrated considering realtime features sent from the MCU, which were updated to increase the accuracy.
Nevertheless, not all 16 features used in this study should be used during real clinical applications. Te FMN shows no signifcant variation during the test, while the variation of the other features was noticeable. Accordingly, some of the 15 discussed features should be changed in clinical applications.

Experiment and Results
Based on the mechanical structure 3D printed and real-time models downloaded, a functioning prototype of the proposed device was built, as indicated in Figure 13.
Te mechanism and circuitry of the device were welted on a 5 mm acrylic board using hot-melt adhesives. Actuation of the mechanism was accomplished using a pair of FUTABA S3003 servo motors. Two sets of 5 V batteries were afxed on the board due to the connection plugin shape variance. In the left corner of Figure 13(c), the JLink device, CH340 UARTmodule, and push button were detached using the portable device. Te white ropes shown in Figure 13 were used to fasten the device to the subject.  To test the accuracy of the actual prototype, the accuracy was defned by [26] ACC � T p T p + F N + F P × 100%, where T P , F P , and F N denote the number of correct recognitions, incorrect recognitions, and missed gestures. Te subject wore the portable device for 1 min. Te results are listed in Table 3. Five tests were conducted considering a recording length of 1 min. Teir overall average accuracy was 95.3%. During the test, the subject was asked to perform daily actions, such as drinking or walking, to simulate casual activities. Te total number of gestures was not confned; therefore, the experiment could simulate casual activities within a fxed time interval. Based on the test results, therefore, the proposed portable face-shielding device was suitable for further implementation.

Conclusion
A portable face-shielding device utilizing sEMG was proposed, assembled, and proven to be feasible. Te purpose of this research was to help disabled and paralyzed patients to independently wear face shields, reducing the risk of contamination due to the direct touch inside the face shield by the hands of ordinary people. Furthermore, new electrode materials using CNT/PDMS were proposed and tested to evaluate their dermotropic properties and electrical conductivity. Te proposed device shows promising prospects for future implementation during the COVID-19 pandemic. Further product transitions are encouraged to target wear suitability and audience generalization.

Data Availability
Te raw/processed data and program source code required to reproduce these fndings are as follows: https://github. com/JasonLvernex/BMI_automatics-faceshielding-device.

Conflicts of Interest
Te authors declare that they have no conficts of interest.