Soft Smart Biopatch for Continuous Authentication‐Enabled Cardiac Biometric Systems

Biometric locking systems offer a seamless integration of an individual's physiological characteristics with secure authentication. However, they suffer from limitations such as false positive and negative authentication, environmental interference, and varying disadvantages across multiple authentication methods. To address these limitations, this study develops a soft smart biopatch for a continuous cardiac biometric wearable device that can continuously gather novel biometric data from an individual's heart sound for authentication with minimal error (less than 0.5%). The device is designed to be discreet and user‐friendly, and it employs soft biocompatible materials to ensure comfort and ease of use. The patch system incorporates a miniaturized microphone to monitor sounds over long periods and multiple dimensions, enhancing the reliability of the biometric data. Furthermore, the use of machine‐learning algorithms has enabled the creation of unique identification keys for individuals based on the continuous monitoring properties of the low‐cost device. These advantages make it more effective and efficient than traditional biometric systems, with the potential to enhance the security of mobile devices and door locks.


Introduction
A biometric identification and security system is a set of coordinated technologies that collect information about an individual compensate for possible environmental factors. [3]Currently, there is a magnitude of user resistance to Facial ID because, despite its advantages, there is some hesitancy because of privacy concerns. [4]Because it scans for particular patterns, inconvenient lighting scenarios, and random reflections, iris scanning can be fooled with a high-quality image of a known user's eye.Because this biometric is more expensive than the rest, all these disadvantages add up to it being overpriced for inferior technology.The most crucial disadvantage for all security above applications is that there is no continuous monitoring across multiple security dimensions.Behavioral biometrics have been examined for continuous authentication through smartphones, wearable devices, and facility cameras.Examples of such behavioral biometrics include keystroke and touchscreen dynamics, eye movement, walking gait, body gestures, and sounds of behaviors. [5]owever, these listed biometrics are subject to change over time, such as how an individual's eye movement dynamics and walking gait vary with time of day, stress, and other environmental factors.Studies of a user's behavioral evolution have been proposed but proved difficult due to the lack of training data sets and successful AI models. [5]Another example of behavioral biometrics is keystroke dynamics, where machine learning models extract features from a user's raw keystroke data to allow or prevent authentication. [6]Although this method improves the strength of typical passwords by discerning a user's unique typing tendencies, like other behavioral biometrics, keystroke dynamics have lower accuracy and shorter permanence than physiological characteristics such as heart sounds. [7]The advantages of biometrics are numerous and significant.The first of the advantages is that biometric data is difficult to forge, providing excellent security because only known users will be granted entry when using these systems.In addition, they do not require user memory or an actual key -biometric technology makes users themselves the key.With this type of security, there is never a situation in which a user cannot get past the verification step.Another advantage is the automatic and easy identification of biometric systems.These key advantages are the most important for this technological era, and biometrics are currently considered bleedingedge technology. [8]It can already be seen in office buildings and hospitals. [9]ere, we introduce a wearable continuous cardiac biometric (CCB) patch to retain the advantages of state-of-the-art biometric technology while resolving some of the current issues with this type of security system.CCB patch demonstrates its potential as a non-invasive and continuous biometric tool.We conduct zero-phase and bandpass filtering to uniquely distinguish each heart signal into unique waveforms in second windows for the machine learning algorithm, all integrated with the mobile application.In addition, our wireless data acquisition unit adopts a low-energy Bluetooth module and biocompatible adhesives, which are suitable for longterm biometrics systems as wearables.This study collectively meets the qualifications of other identification systems and displays novelty in its ability to monitor the heart sound biometrics continuously and remotely across large expanses of time while remaining significantly cheaper than its currently used counterparts leveraged across phone security and door lock applications.

Overview of the Device Design and Functions
The CCB Patch's advantages come from its cross-disciplinary abilities and novel technologies.The device's design requires precise nanomanufacturing to deliver critical attributes for its application, such as increasing its ambulatory aspects, ensuring device longevity, and optimizing component placement for proper auscultation.In addition, the meticulous nanomanufacturing of the CCB patch opens a new avenue for remote patient cardiopulmonary auscultation (Figure 1A). Figure 1B shows an exploded picture of one of the two mechanically soft, wireless components that comprise the whole system.The device itself is a flexible patch, designed in KiCAD (Figure S1 and Table S1, Supporting Information), with copper tracts across multiple layers connecting different board sections.The separate but linked layers of the circuit are essential in manufacturing and packaging to reduce surface area, which is necessary for usage on any patient who requires continuous auscultation with little to no pain.The ambulatory nature of the device is a significant advantage, and it comes from the composing structure of the patch.Being a flexible device provides the device with that malleability and bending.An additional part of its design also allows for an increase in the device's ability to limit motion artifacts and bend through the elastomeric enclosure with an inner silicone-gel liner shown in Figure S2 (Supporting Information).The elastomer helps the skin conformality, ensuring compatibility with the delicate skin and highly curved anatomical features of users utilizing the device.The moisture vapor transmission rate (MVTR) of the other commercial medical tapes is evaluated, and the resulting permeability of the medical tape used in the CCB patch has the highest MVTR, clearly outperforming the other industry standard materials against that it was evaluated (Figure S3 and Table S2, Supporting Information).The data from the long-term monitoring testing showed consistent data points across two and a half hours' worth of recording, with a signal-to-noise ratio showing minimal change through repeated and long-term use of the device (Figure S4A, Supporting Information).During prolonged use, it is common for sweat or dead skin cells to accumulate on the surface of wearable devices.However, our CCB patch exhibited only a minimal decrease in performance, with the attenuation level decreasing from 21.07 dB (Figure S4B, Supporting Information) to 20.68 dB (Figure S4C, Supporting Information) throughout longterm wear.This means that the patch is permeable enough not to irritate over a long period of wearing the device.The before and after image where the device was placed on the skin shows that no disturbance had occurred (Figure S5, Supporting Information).The exemplary passage of all these tests indicates that the CCB patch can endure most of the environments it could experience while on a patient.The ambulatory nature of the device allows for increased quality data because of its ability to move with the user instead of the user moving the microphone.The flexibility and skin-conformal contact by elastomer layers are crucial in continuous biometric identification.The ambulatory nature, in combination with continuous recording, creates the ability to seamlessly record and identify individuals across several layers of identification and throughout an extended period.Part of the idea is to eliminate the need to re-scan at every access point where another Figure 1.Overview of the cardiac biometric system using continuous authentication.A) An illustration of a subject wearing a continuous cardiac biometric (CCB) patch with a close-up of the actual device's top and bottom views (top-left photos) and a control system for unlocking the phone and door locks using the mobile application.B) Exploded view of the patch with multiple layers of deposited materials, integrated chips, and a microphone.C) Schematic illustration of the CCB patch on power, data, hardware structure, and the real-time security system.D) A flow chart describing the entire biometrics system, from the CCB patch to the control of mobile phones and door locks via a machine learning classification algorithm.biometric would require yet another scan.Continuous recording can identify unauthorized individuals with a variation in cardiac cycles, which do not change over extended periods, making it more difficult to imitate than one-time biometric scans.To replicate the sound, one would need to have an exact replica of the heart and chest cavity because the sound is made from the valves in the heart opening and closing and reverberating through the body.Applying the circuitry to the board is quick, efficient, and, more importantly, effective.The CCB Patch utilizes a compact lithium-ion polymer battery (40 mAh), providing a 24 h operating time based on the current design, drawing 6.3 mW power.To optimize battery usage, the firmware can incorporate various power-saving techniques, such as always-on or standby mode.In standby mode, the device significantly reduces the data sampling rate when not connected to the host device.Upon triggering, it resumes normal sampling rates to conserve battery power effectively.Similarly, suppose extended authentication system usage of up to a week is desired.The CCB patch design can be modified to accommodate a larger 250 mAh battery for a week-long recording.Also, for a guided battery connection, the circuit's power pads are connected to the battery.Battery location ensures gentle placement on the curved skin of the chest by being secured atop the device.The battery enables it to auscultate cardiac activity by a Bluetooth Low Energy (BLE) system-on-a-chip (SoC) and associated set of sensors for the auscultation of cardiac activity.The analog signal acquired from the MEMS microphone travels through the pre-amplifier.Then it moves to the ADC to get converted to a digital signal to feed the BLE microcontroller to send data wirelessly via Bluetooth to mobile devices (Figure 1C).The raw signal acquired from the device will then be processed through the convolutional neural network (CNN) training with cardiac activities logging to physicians and finally fed into the secondary authentication using cardiac sounds.The extracted features will be trained in the model to authenticate the individual.This process flow is simplified in Figure 1D.The device's standard biometric qualities, ambulatory nature, remote and continuous recording, very low possibility of a change in heart sound for an individual, and the near impossibility of forgery combine to create a secure, easy-to-use biometric security system.

Physiology in Continuous Monitoring and Mechano-Acoustics
Analyzing cardiac sounds and the characteristics that make them unique create the basis of this biometric.The opening and closing of the heart's valves produce cardiac sounds.The heart has several distinct sounds, including S1, S2, S3, and S4, [10] but the points of interest for the device are S1 and S2 for normal subjects.The S1 sound is produced by the mitral and tricuspid valves closing, as shown in Figure 2A.This period of the cycle is known as systole. [11]When the pulmonic and aortic valves shut during diastole, the S2 sound is produced, which is the louder of the two, and this is the beginning of diastole.The average duration of systole and diastole are 0.35 and 0.45 s, respectively, totaling a cardiac cycle lasting ≈0.8 s [11] shown in 4 representative participants' data from Figure 2B.The complete cardiac cycle as S1 and S2 coordinate with pressure and volume changes can be seen in Figure 2C and 2D.The cardiac sounds have a frequency range of 20 -220 Hz. [12] The device's audio resolution is subtle that specifics concerning the noises produced by the heart may be identified. [13]Those specifics include several aspects of the cardiac sound, like magnitude, shape, size relativity between S1 and S2, and standard deviations in the timing duration of a cardiac cycle.Other factors influencing cardiac sound are how the sound reverberates within the heart and chest cavity.Significant individual variation is caused by the heart's activation order, conductivity, and heart mass orientation. [14]For example, an individual with a fatty heart could have quieter heart sounds, while an individual with a large chest cavity could have a deeper frequency sound.Once the variability of an individual's heart rate is measured and quantified, the next step begins.The raw data set from the CCB Patch is first run through digital signal processing.The two filters applied to it include the zero-phase filter with a 3rd order Kaiser window for magnitude normalization and the bandpass filter set to cut off frequencies outside of the frequency range of the heart, which is 20-250 Hz.After the filters are applied, and a new, clean data file is produced, the information is sent to the machine learning program.After the precise and high-quality data is gathered using the CCB patch and cleaned through extensive digital signal processing, the computer can generate an individual profile with the specifications wired in. [15]If the algorithm notices the known characteristics in an incoming data flow, it can identify the owner of that specific data set.The heart sound in time series, with S1 and S2 visible and distinguishable, is presented in Figure 2E [9] synced with varying pressures and ventricular volumes for two complete cardiac cycles.This time-series plot is what machine learning takes after filtering away outside noise interference.The data set's purity reflects the efficacy of the device's digital signal processing and noise suppression algorithms.

Mechanical Characterization of the Wearable Patch
The results of the mechanical testing performed on the CCB patch are shown in Figure 3.The bend/stretch testing on the device showed that it had no significant variation of resistivity due to applied stress with various motions from subjects wearing the device shown in Figure 3A.To simulate stretching and bending, cyclic testing has been done with measuring the resistance change between each end of the circuit to ensure data reliability during the actual monitoring (Figure S6, Supporting Information).Figure 3B shows the stretch testing results, and Figure 3C shows bending results from the cyclic mechanical tests using a digital force gauge (EMS303, Mark-10) and a multimeter (DMM7510, Tektronix).Less than 20 mΩ fluctuations in resistance fluctuation were observed over the 100 cycles.For cyclic stretching testing, the total resistance change was ≈0.53 mΩ, and for cyclic bending tests, it was 0.64 mΩ.Computational finite element analysis also captures the stretching and bending results from Figure 3D and 3E with 180 degrees bending and 20% stretching, [16] showing less than 1.5% of strain for every section of the circuit.The CCB also maintained the highest pressure over the extended period compared to commercial medical tapes, meaning it maintained the best skin contact (Figure S7, Supporting Information).The CCB patch performed better than expected in the waterproof testing, with the data quality.At the same time, underwater being only marginally noisier, the excess noise ceased as soon as the device was pulled out (Figure S8, Supporting Information).The device is conformal to the skin, maintaining its functionality under various conditions.The device accurately detects radial pulses even after water flows and water permeates beneath the skin.Moreover, its performance remains consistent across a range of body temperatures.This demonstrates its robustness and reliability, suggesting its potential for diverse applications.

Machine Learning Architecture and Performance
For pre-processing cardiac data from the CCB patch, bandpasses of 20 Hz to 250 Hz and zero-phase filtering (Figure 4A) are done to reduce noise in the signal and preserve time and phase shift issues that is crucial in time-series-based ML training.Details of the zero-phase filtering are discussed in Figures S9 and S10 (Supporting Information), showing the actual power spectrum of the zero-phased cardiac data.Figure 4B shows the 1 s window of the example cardiac signal from the CCB patch with raw, band-passed, and zero-phase filtered graphs.While bandpass eliminates the DC offset caused by the device, it cannot entirely reduce the baseline noise, which zero-phase filters out with a phase shift of zero for all frequencies.In the CNN architecture, data from 20 subjects in our lab dataset were utilized.The dataset was divided into three segments: 60% for training (40 075 epochs), 20% for validation (13 625 epochs), and 20% for the test (13 625 epochs).During each training iteration, the CNN network parameters' weights were adjusted based on the model's training validation accuracy.Most of the hyperparameter values, such as learning rate, kernel size, filter count for each convolutional layer, and units for each dropout, were determined using a random search approach.Ultimately, the model with the highest validation accuracy was selected as the optimal model.The performance of this top model was assessed based on the prediction accuracy of the test dataset.Consequently, the CNN architecture employed (3,1) pool size of 2D-max pooling, 25 filters, and (10,1) kernel size for the two 2D-convolutional layers in the initial block.Batch normalization was incorporated to mitigate overfitting.For the second block, 25 filters and sequential kernel sizes of (10,25), (10,50), and (10100) were used for the three individual convolutional cells (Conv_1,2,3), along with a (3,1) pool size of 2D-max pooling (refer to Figure 4C and Figure 4D).A series of layers in the machine learning model was applied, as illustrated in Figure 4E, Figure S11 and Table S3 (Supporting Information).The model was trained to ensure that each participant's heart sound waves had different patterns and forms for each S1 and S2 pair segment in the time series.Because the participants' average beats per minute were about 60, 120 samples with the input of 2 s were sent to each participant's class. [15]he confusion matrix of each participant's different waveform trained in the model is shown in Figure 4D, demonstrating that the machine correctly detects each participant's heart sounds. [1]s demonstrated by the proposed biometric accuracy or correct recognition rate of 99.55%, the heart sounds biometric offers a solution to all these difficulties, albeit having an error rate of 0.45%.As compared with other devices in Table 1, this work shows the best performance in accuracy.The suggested heart biometric outperformed the estimated error rate of fingerprint, signature, and voice recognition.There is an extremely little chance that a  The capacity to continuously acquire an individual's heart sound is the most significant and evident benefit because it creates identification across many layers of security.The goal is to eliminate the requirement to re-scan a biometric at every entry point necessitating a new scan.The CCB Patch is an efficient, adaptive, unobtrusive, and accurate technology that analyzes an important biometric in the body -heart sounds.

Security Applications of the Patch for Mobile Devices and Door Locks
Figure 5A shows the overall system of our security system using our CCB.Data transmission structure is simple: Wearable devices on the user's sternum, mobile devices, and wireless door lock systems.In this study, three devices work sequentially for secure biometric authentication: A lock with an RF receiver mounted with a linear actuator to unlock the lock, an RF remote integrated with the BLE development kit (Figure 5B).Finally, a mobile device working as the primary device to compute the authentication system, as shown in Figure 5C.An example in Video S1 (Supporting Information) demonstrates this security application.The system applied security functions through Android Keystore API in Jetpack android studio library for the encryption and decryption (or store and access) of biometric data input/output stream, allowing for the secure processing of the data.CNN model is embedded in the Android application with Tensor-Flow.The CNN model can extract the S1 and S2 data in real-time for the user's feature detection and authentication (Figure S12, Supporting Information).After applying the CNN model, the new user's processed data will be registered for authentication.
For the performance and secure file stream structure of the CCB system, we developed a signal classifier mechanism to evaluate CNN's continuous input cardiac data.The classifier is based on sequential matching and anomaly detection that can comprehensively predict both class labels and the similarity of signal features.App backend decrypts the model and compares the user signal features (Figure S12, Supporting Information).The signal classifier output will be wirelessly communicated as metadata to the control device, as shown in Figure 5C.Encoded metadata of continuous packets from mobile device advertise through read/writable BLE structure on individual UUID characteristics.Since the output from the mobile device is a decision-making signal, it is encoded securely.A key pair decoder can only decode it in the control device firmware.Inside the control device, authentication of model output is used as decision-making signal data, as illustrated in Figure 5D.The firmware structure in the microcontroller decodes the signals and validates the model classification.With an embedded validation algorithm, the control device records the success log to the internal board and sends a control signal to the wireless door lock.Also, the control device sends a writable byte signal to the mobile device for an additional record logging of successful authentication shown in Figure 5E (details in Figure S12, Supporting Information).Lastly, the wireless door lock receives the paired control device's signal and controls the motor driver to lock or unlock the door.The CCB interface shows a secure hardware and software interface applicable to wide-ranging applications.Video S2 (Supporting Information) shows the successful door unlock demonstration of the registered user.In addition, because the second user was not one of the 20 training data on the ML algorithm, the CCB system identified as not a registered individual to attempt door unlock.If we expand the sample size, the accuracy will rise and keep ≈100% accuracy.The presented work can find additional applications using wearable cardiac-biometric systems (Figure S13, Supporting Information), including confidential file transfer, bank security, 2-factor authentication, and secured remote patient monitoring.Applications for mobile devices and door locks.A) An illustration of the CCB user unlocking the phone with the registered cardiac data as well as unlocking the door using the mobile application.B) Actual photos of the door lock system with a circuit integrated with a remote control that sends the signal to the door lock actuator.C) A schematic flowchart of the control system and logistics of the door unlocking system.D) A detailed schematic flowchart of the control device in the hardware.E) Screenshots of registered and unregistered users in the mobile application.

Conclusion
This study presents the development of wearable soft CCB patches for recording heart sounds and incorporating biometric locking mechanisms.We collected heart sounds from twenty individuals and processed them through a two-stage signal processing pipeline to remove interference and artifacts.Additionally, we explored the potential of the CCB patch system to predict heart signals using deep learning techniques, achieving a 99.55% accuracy.These findings demonstrate the superior practical usability and security of the CCB patch compared to existing biometric systems.Integrating biometric capabilities in the wearable patch offers a reliable and secure alternative to traditional passwordbased authentication, enhancing safety in virtual reality transactions and digital security applications.By leveraging biometrics, the CCB patch eliminates the need for strong user memory or physical keys, providing resistance to replication.The fast, handsfree, and remote identification capabilities of the CCB patch open new possibilities for various applications.Future research will focus on enhancing durability, integrating the patch's biometric security features with its health monitoring capabilities, enabling seamless integration into daily activities, and enhancing overall well-being.The ability of CCB to leverage a deep learning strategy and make predictions of heart signals based on obtained information holds substantial promise, particularly for individuals with underlying heart disease.Recognizing its significance, our system can potentially influence the diagnosis and monitoring of cardiac conditions.

Experimental Section
Device Fabrication: A flexible printed circuit board (FPCB) was designed using kiCAD software for the wearable stethoscope sensor.The entire device is 53 mm × 25 mm at its widest points.The copper tracts of the FPCB end on the board's surface at a copper pad, and the nanocircuitry was placed on its respective set of copper pads and melded on using solder paste and a hot plate (Figure S2A, Supporting Information).This meticulous process was completed to ensure the longevity of the device's functioning.A layer of a soft elastomer gel was used as a base adhesion layer for the wearable stethoscope to the skin.A 20-gram mixture of the Eco-flex gel A & B was poured onto the 150 mm diameter petri dish and spin-coated at 1000 rpm for 5 s.This ensures that the even layer touches the skin without deformation to the circuit.The rough boundary was cut out from the cured gel layer in the petri dish, and the integrated circuit board was placed on top of the gel layer.A hole was sliced in the middle of the microphone island to allow sound to permeate through the elastomer layer (Figure S2B, Supporting Information).Another layer of a soft silicone elastomer was created by mixing a 1:1 ratio of Eco-flex 30 A & B and pouring it on top of the entire circuit board, covering the gel layer's edges to encapsulate the device entirely.Lastly, high-tack silicone gel was used on a fabric layer, spin-coated at 1000 RPM for 30 s cured on a 60 °C hot plate for 15 min.The fabric was cut out in a circle on top of the encapsulated microphone island for better pressure applied to the microphone (Figure S2C,D, Supporting Information).
Hardware Design: Several electronic components in the design were utilized for the CCB patch.Passive components such as resistors, capacitors, and inductors were used as well as the BLE SoC (nRF52832, Nordic Semiconductor), ADS1292 Analog-Front-End chip from Texas Instruments, ICS40212 MEMS microphone chip from TDK Invensense, TS472 audio amplifier chip from ST Electronics, and finally TPS63001 switching voltage regulator chip to supply voltage to the whole circuit.

Mechanical Study:
The following mechanical tests were performed on the device to ensure its performance across all types of environments: bend/stretch testing, SNR testing, contact pressure testing, waterproof testing, permeability testing, and irritability/duration testing.The stretch/bend testing was conducted using the ESM303, Mark-10 machine; it stretched the device while resistivity was recorded to see if there was any effect on the device's efficacy.The same was done for the bend test.Additionally, the device's attachment material was tested for its ability to maintain a high enough level of pressure in skin contact and deliver an acceptable signal-to-noise ratio (SNR).The waterproof testing was the simplest of all the tests.It entailed starting the recording, submerging the device, taking it back out, and seeing if it could continue to record afterward (Figure S8, Supporting Information).Finally, the permeability testing was conducted by filling uniform containers to the brim with water and sealing them using different types of medical coverage, with an open container as a control.Water evaporates faster out of the containers filled by a medical range with higher permeability, so after seven days untouched, the material with the most increased permeability would be the emptiest (Figure S3, Supporting Information).Finally, the irritability/duration testing was conducted by collecting cardiac data across 2.5 h and examining the skin where the device was placed (to evaluate irritability) and the quality of the data collected with a signal-to-noise ratio (Figures S4 and S5, Supporting Information).
Data Collection & Filtering: The CCB Patch was mounted slightly to the left of the sternum for optimal cardiac sound collection and was tested on 20 different participants. [6]The duration for each session was ≈60 s.The subject was stationary and seated for the collection process.The patch acquires raw cardiac signals and sends buffered data wirelessly to a computer or mobile device.The signals are synched with timestamps and saved through files such as CSV, and ready to be analyzed.First, the raw signal undergoes signal processing to produce a pure form of the training and CNN classifier data.After signal processing and smoothing for feature extraction, the preliminary digital sound data is pre-processed for machine learning.Lastly, the signal can be analyzed as an individual profile in the CNN model.
Classification of Heart Sounds: Digitized sound data is recorded at 4 000 samples per second, and a professional human analyzer segments the corresponding data for training.The pieces of data were then fed into a machine learning algorithm (TensorFlow by Python), and profiles were created for everyone.In their profile, the machine was trained with information about their heart sound, how it can vary across time, and what was specific about it to them. [17]To maintain a balanced dataset during the training process, the number of samples for each class to be approximately equal was adjusted, thus minimizing classification bias.The model's architecture was devised through a series of iterative refinements, informed by prior research. [18]The model was designed based on the input data of a CNN architecture, specifically for time-filtered signal data.The CNN architecture inputs consisted of 2 s raw signal data, having a size of 8 000 × 1.To pre-process the dataset of CNN architecture, Fast Fourier Transform (FFT) and Min Max Scaling were used to extracting features.The chosen non-linear activation function was the Leaky Rectified Linear Unit (Leaky ReLU).The CNN architecture was optimized using the ADAM optimizer (learning rate = 0.001, ß1 = 0.9, ß2 = 0.999), with a batch size of 64.To avoid overfitting, early stopping was implemented by randomly excluding 30% of the data from the training set and designating it as the validation set at the onset of the optimization phase.If the validation loss ceased to improve, the learning rate was decayed by a factor of 5.The training was terminated if there were two consecutive instances of learning rate decay without any observed improvement in the network performance on the validation set.The CNN architecture incorporated two distinct convolutional blocks.The initial block consisted of a pair of 2D convolutional layers, batch normalization, and a 2D max pooling layer.The subsequent block featured three separate convolutional cells (Conv_1,2,3).Each individual convolutional cell (Conv_N) contained a convolutional layer, a batch normalization layer, a Leaky ReLU layer, and a max pooling layer.The Convolutional Neural Network's final output was a 20 × 1 vector, which was subsequently passed through a softmax layer, resulting in the predicted class corresponding to one of the twenty participants.S11 and S12, Supporting Information).The primary(master) key was generated for individual app releases and recorded for additional protection.The output sends wirelessly as metadata with byte algorithms that can only be decoded to the control device.The metadata received on the control device board as a hex-type with nRF52832 and BLE platform.The control device receives only correct peripherals signals (matching specific service UUID and byte decoder) from a mobile device and decodes the output with the received secure key.The embedded assessment algorithm in nRF52832 evaluates the class model and makes authentication decisions.Lastly, the control device transmits the control signal to the door RF receiver integrated with the linear actuator (Actuonix).For security purposes, the control device records the log of the successful trials on board and returns the success signal to a mobile device by writable characteristic UUID.This allows future investigation of tracking authentication history.The paired control device can only initiate the final wireless door lock system.The door locker receives specific critical information to unlock the door.
Human Subject Study: The human pilot study involved multiple healthy volunteers; the study followed the approved IRB protocol from the Georgia Institute of Technology (#H21038).All participants agreed and signed the consent form to allow the experiment procedure.

Figure 2 .
Figure 2. Physiology in continuous monitoring and mechano-acoustics of the cardiac system.A) An illustration of the heart valves and the phonocardiogram (PCG) principle in physiology.B) Four representatives raw PCG data of the time sequence on S1 and S2 peaks.S1: mitral and tricuspid valve closure, S2: closure of the semilunar (aortic and pulmonary) valves.C) Graph illustrating the pressure change in heart physiology.Order of occurrence: 1. Isovolumetric contraction, 2. Ejection, 3. Isovolumetric relaxation, 4. Rapid intake, 5. Diastasis, 6. Atrial systole; with important events: i. A-V valve closes, ii.Aortic valve opens, iii.Aortic valve closes, iv.A-V valve opens.D) Graph of a normal ventricular volume in a cycle of occurrences in heart physiology.E) Time series PCG data of systole and diastole synced with cardiac pressure and ventricular volume.

Figure 3 .
Figure 3. Mechanical characterization of the flexible patch.A) An illustrated diagram of a user with the CCB patch in motion where the device stretches, bends, and compresses.B) The result of the resistance change of cyclic stretching with 20% stretch through 100 cycles.C) Resulting resistance change during the cyclic bending of 180°with a 3-mm radius for 150 cycles.D) Finite element analysis of the 90°and 180°bending with the actual photos of the circuit board encapsulated with the zoomed-in views of various areas with the most strain.E) Finite element analysis of the 20% stretching with zoomed-in views of various areas with the most strain.

Figure 4 .
Figure 4. Machine learning architecture and results of the 20-participant data using the CCB patch.A) A pre-processing overview of the flowchart for filtering stages of the raw CCB data.B) 1-second window plots of raw, band-passed, and zero-phase filtered data.C) Flowchart of the overall cardiac biometric classification.D) Confusion matrix of the trained model from the 20-participant data with an accuracy of 99.55%.E) CNN-based machine learning architecture for cardiac biometric classification.

Figure 5 .
Figure5.Applications for mobile devices and door locks.A) An illustration of the CCB user unlocking the phone with the registered cardiac data as well as unlocking the door using the mobile application.B) Actual photos of the door lock system with a circuit integrated with a remote control that sends the signal to the door lock actuator.C) A schematic flowchart of the control system and logistics of the door unlocking system.D) A detailed schematic flowchart of the control device in the hardware.E) Screenshots of registered and unregistered users in the mobile application.
Biometric Authentication: A real-time biometric authentication platform was developed to process real-time CNN models in a mobile device.A secure and fast deep learning framework with Tensorflow (v2.8) and Tensorflow lite processes the output of the CNN model.TF was used for deploying the designed dataflow structure to compute the output array by processing acquired signal data from CCB. TF lite serialized toolkit was embedded in the Android Kotlin Application (Kotlin v 2022.1.7)and facilitated the implementation of the CNN algorithm.The signal processing system used in ML training was converted to C++ for the fast framework of real-time signal processing to handle real-time input.The custom classifier mechanism comprehensively analyzes segments sequentially and detects S1, S2, feature detections, and anomalies of the input filtered signal along with CNN classes (Figure S11, Supporting Information).With user input of the lock & unlock button, the extracted cardiac data was evaluated with a comparison of registered authentication data and generates a signal classifier output.The output layer of the CNN model was developed for user biometric data in real-time.The feature extracted cardiac data assessed through trained and registered authentication data inside the app.The platform utilizes security procedures with Android Keystore API in Jetpack android studio library (Compose v1.1) for model encoding/decoding, critical pair matching, and modeling loading process (Figures

Table 1 .
Comparison of the performance of various biometric systems.
person's heart sound will alter throughout a recorded time.Heart sounds are difficult to fake unless an unauthorized user attempts to clone a heart identical to another person's and set it within a chest cavity that reverberates similarly.Compared to the issues that facial recognition software, fingerprint identification, and other biometric security measures have brought up, measuring an individual's heart sound may be a far less invasive alternative.