Machine Embroidery Enclosure for Stretchable Fiber Optic Respiration Sensor

Developing an unobtrusive respiration sensor is a key task for daily health and safety monitoring, especially for individuals with potential breathing abnormalities or those performing in physically challenging conditions. This study introduces a machine‐embroidered enclosure that guides a soft and extensible lightguide to react to the strains on a compression shirt caused by the respiration of the wearer. The shape of the stretchable optical fiber changes from serpentine in rest (exhalation) to straight under strain (inhalation), which affects the light transmittance. In tests with 13 healthy adults, the respiratory volume prediction through a deep learning model trained by the light intensity and truth from a commercial spirometer shows a high correlation in both static (r2 > 0.880) and dynamic postures (r2 > 0.690). This system also accurately measures the respiratory rate under sitting and walking conditions (absolute error <1.513 BPM, SNR > 7.122 dB). The strain‐sensing capability of this sensor depends on the parameters of the embroidery, including the stitch density, stitch tightness, and embroidery shape. Durability tests confirm that the system still functions after 100 000 abrasions and 10 cold washes.


Introduction
The respiratory pattern is one of the major indicators of the physiological and emotional status of a person. [1,2]From chronic respiratory diseases and of the severe acute respiratory syndrome (SARS) of the recent pandemic to heart failure, an assessment of the respiratory pattern can contribute to accurate medical prediction, diagnosis, and treatment. [3,4]The respiratory rate, tidal volume, absence/duration of the flow, and flow rate are the most popular parameters used to define a pattern, especially abnormal respirations such as apnea, hyperpnea, and/or Cheyne-DOI: 10.1002/adsr.202300017Stokes respiration. [5]Because the breathing behavior accompanies movements of the ribcage and organs, real-time respiration tracking is also crucial for accurate medical snapshots such as in magnetic resonance imaging (MRI). [6]onitoring respiratory patterns is also of significant interest to those pursuing high performance in physical activities like elite sports because a different breathing pattern can improve exercise performance. [7]In a similar sense, professionals conducting important missions under a physically harsh environment like soldiers, astronauts, firefighters, and divers are looking for new or better methods of integrating a respiration monitoring system into their mission arena. [8]he method used to assess pulmonary function involves the measurement of direct/indirect respiratory parameters such as the airflow, chest wall movement, or intrathoracic volume. [9]Spirometry, which directly quantifies the entering and leaving airflow, requires a nose clip or tightly sealed face mask to ensure accurate measurements without air leakage.15][16] Various methods including but not limited to electrocardiography (ECG), [17] near-field coherent sensing (NCS), [18] computer vision, [19] and radar [20] has been explored, but noise control, skinfriendly device fabrication, and blind spot in vision sensors remain as tasks.As a common indirect sensor for respiratory inductive plethysmography (RIP), strain gauges around the thoracic and abdominal circumferences are used to track subtle expansions of the ribcage following the lung volume. [16]The strain measurement approach can be used in most settings from a static lab session in a hospital to sports and exercise in an unstructured environment. [21] fiber optic strain gauge is one of the promising onbody strain sensors.[24] The embedded optical guidelines react to the perturbations of the structure, which leads to the optical changes of the transmitted light. [25][36] The merits of optical sensing also include immunity to electromagnetic interference, [37] so optical fiber-based respiration sensors can contribute to accurate results and the real-time monitoring of a patient's condition during MRI with a high magnetic field and/or computed tomography (CT), which often require the patient to hold their breath for ≈15 s to remove respiratory artifacts. [26,38,39]he challenges to designing a wearable fiber optic strain gauge for respiration sensing include i) sufficient sensitivity to monitor subtle ribcage movements, ii) the minimization and distribution of the light source and detector, and iii) the prevention of discomfort from the tightness of the strain gauges.The chest circumference expands by up to 4 cm during normal breathing and 8 cm during deep breathing, which is a strain of only 4.6-9.2% in terms of the underbust girth of a U.S. female and 4.0-8.0%[42] Therefore, the fiber optic sensor has to have a high gauge factor and linearity in the low strain range (<10%).A fiber Bragg grating (FBG) can detect subtle strain changes and multiplexed information from a single fiber (e.g., 1.21 pm of wavelength shift per microstrain in case of the standard 125-micron silica fiber [43] ), but it often comes with a relatively bulky and rigid optical interrogator, which is not ideal for most daily activities. [33]][46] Lastly, elongation-or compression-based strain sensing requires either a pre-strain to avoid non-linear signal changes or a high stress than bending, which will introduce a stronger tension around the body. [35,45,47]ontrolling the fiber movement is important to ensure consistency in macro-bending methods, but most applications so far have only secured the fiber permanently using glue or stitches at a few fixed points, which has allowed the rest of the fiber to arbitrarily bend without any control or guidance. [30,44,46,48]is study reports on an unobtrusive and stretchable respiratory monitoring system consisting of an elastomeric optical fiber, called Optical Lace (OL) and an embroidered enclosure.The zigzag-based machine embroidery created a room to install the OL on the compression shirts as well as guide the OL movements by strain without introducing further mechanical loads on the substrate nor the soft fiber (Figure 1a; Figure S1, Supporting Information).The system showed a strong abrasion resistance and a resilience against multiple machine washes while performing sensitively without bulky electronics like interrogators but only an LED and a photodiode.The tensile stress on an elastic band changed the shape of the optical fiber from a serpentine pattern to a straight line without tensioning the OL itself, which increased the intensity of the light transmitted in the OL.The sensitivity to strain was adjustable by the embroidery parameters such as the density and tension of the stitches.We embedded two sensors in a compression shirt on the thoracic and abdominal circumferences to accommodate different breathing styles [1,9] and monitored the respiratory rate and respiratory volume.The signal displayed satisfactory correlation coefficients with the respiratory volume (RV) trend from a spirometer in both static and dynamic postures (0.690 < r 2 < 0.881), and a neural network model (multilayer perceptron (MLP) regressor) accurately predicted the RV (0.258 L < RMSE < 0.381 L).The respiratory rate (RR) computed by a fast Fourier transform (FFT) also showed low absolute errors between 0.017 and 1.513 breaths per minute (BPM), except under the running condition, which added significantly more motion artifacts.This study also found that the system was resilient to typical adverse scenarios such as abrasions (100 000 rubs) and washings (10 cold washes using a detergent).Lastly, while the sensor sewn into a compression garment was the main interest of the study, we compared its performance to a belt-type sensor to reveal the best option for consumers.

System Design
The goal of the design was to create a textile-friendly enclosure to i) hold the OL firmly on the elastic band, ii) guide its movements between a serpentine shape and straight line, iii) with no or minimal limitation of the original stretching behavior of the elastic band.Integrating an optical fiber into a textile itself at the initial manufacturing phase (e.g., weaving or knitting) could achieve a durable and seamless installation of the sensor in the textile, but the optical fiber often experiences large light loss as a result of micro-bending in the midst of the woven/knitted yarn. [32]Gluing the entire fiber optic on the textile or enclosing it in a serpentine pattern with another layer of elastic textile or elastomer could restrict the stretchability of the basis elastic band.Hand embroidery and gluing to create a few anchors on the fiber do not guarantee that the unfixed parts will remain on the textile surface or behave as expected under strain. [48]herefore, this study adopted machine embroidery consisting of only zigzag stitches as the enclosure for the strain sensing OL.A zigzag stitch is one of the most common and widely used stitching methods for stretchy textiles.It does not produce significant changes in the mechanical properties of the textile under strain. [45]Unlike typical embroidery where many short stitches fill the target area, the embroidery in this study used stitches only at the boundaries of the enclosure (i.e., the needle penetrated the upper/lower edges of the enclosure shape, not in the middle), to create a space underneath the embroidery where the OL could freely move and change its shape under strain (Figure 1b).The OL enclosed in the embroidery was longer than the horizontal length of the embroidery, which caused the OL to bend into a serpentine shape with fiber peaks as guided by the five trian-gular spaces created by the embroidery.Light leaked around the peaks of the curve, which reduced the amount of light arriving on the other side.When the elastic band was strained (i.e., the wearer inhaled and the ribcage expanded), the zigzag stitched enclosure followed the horizontal expansion and straightened the OL, which increased the light intensity.

Respiratory Volume Measurement by Conditions
The normalized light intensity of the raw signals from the OL and the respiratory volume prediction using the MLP model showed good agreement with the ground truth from the spirometer (Figure 2).The Pearson correlation coefficient (r 2 ) was the highest at 0.924 in the setting where the wearers sat and breathed deeply (Std.Error = 0.016, Figure 3a).Normal breathing while sitting had an r 2 value of 0.887 (Std.Error = 0.024).When the wearers walked and ran on the treadmill, because of the motion artifacts, the coefficient was not as good as those for the stable postures (walk: r 2 = 0.844, Std.Error = 0.045; run: r 2 = 0.872, Std.Error = 0.024).The root-mean-square error (RMSE) values of the predictions in terms of the actual respiratory volume were 0.258 L (sitting and breathing normally), 0.381 L (sitting and breathing deeply), 0.338 L (walking), and 0.367 L (running).The raw signal also showed a satisfactory correlation across the tasks (normal  breathing: r 2 = 0.880, deep breathing: r 2 = 0.896, walk: r 2 = 0.816, run: r 2 = 0.691).][51] The RMSE needs to be reduced for a more accurate prediction considering that normal breathing commonly has an amplitude of ≈1 L. However, the current prediction was based on the model trained by only two 1-minute sessions (i.e., 120 seconds in total for the individual MLP model for each participant), so it can be improved by training with a larger dataset.
There was no significant difference between the shirt-type sensor and belt-type, but as a single sensor, the belt on the upper chest showed fine performance overall.Meanwhile, the sex or body mass index (BMI) did not produce any significant bias in the results (Figure 3b,c).

Respiratory Rate Measurement by Conditions
The stretchable optical fiber enclosed by machine embroidery accurately monitored the respiratory rate when the wearer was sitting (absolute error deep-shirt = 0.017 ± 0.002 BPM, Figure 4a) or walking (absolute error shirt = 1.024 ± 0.555 BPM).However, the error was significantly high when the wearer ran (absolute error shirt = 9.35 ± 5.003 BPM, absolute error belt = 23.86 ± 6.999 BPM) as a result of the increased motion artifacts.This study used the frequency with the highest amplitude in the FFT result as the RR.Thus, when the noise caused by consistent running motions increased, it sometimes replaced the RR. Figure 4b shows the second-highest peak around 180 BPM, which was from a common running speed of 90 strides (left and right) per minute. [52]The signal-noise-ratio (SNR) also showed the amount of noise from the motion artifact (Figure 4c).While the signal was robust compared to the noise in sitting postures (SNR > 14.6 dB), active body movements added significant noise (SNR walk > 7.122 dB, SNR run > 4.106 dB).There was no noticeable difference between the shirt-and belt-type sensors, but the absolute error for respiratory rate monitoring during running was higher when using the belt.

Sensitivity Adjustment
The light source and straight-line length of the OL predetermined the level of maximum light intensity on the side of the light detector.Therefore, the sensitivity to strain depended on the methods used to decrease the light transmittance when the sensor was free of tension.The embroidery enclosure determined the condition of the OL when released.Thus, the parameters of the embroidery could be used to adjust the strain sensitivity of the sensor.The embroidery shape, number of curves, stitch tightness/density, thread type, and other factors could be adjusted to achieve the optimal sensitivity for a given application.Figure 4 shows the gauge factor (GF = |ΔV∕V|  , where V is the analog input level indicating the intensity of light and  is the strain) between the released status and fully strained status based on the embroidery parameters.
Because the light leakage around the curve depended on the curve angle (Figure S2, Supporting Information), the shape of the embroidery, which caused the fiber to curve, affected the GF, as shown in Figure 5a.When the width and height of the shape were the same, a triangle-shaped enclosure, which could create the sharpest curve angle (>0.92 rad), showed the highest sensitivity, followed by the square (>1.57rad) and round shapes (with different but larger angles than the triangle).The number of spaces accommodating the extra length of OL determined the number of curves.More curves resulted in greater light leakage, along with larger changes in the light intensity when the OL was straightened (Figure 5b).The stitch spacing, which refers to the distance between two adjacent peaks in zigzag stitches, had a negative relationship with the gauge factor (Figure 5c).A smaller spacing (i.e., denser stitches) loaded more compression from the thread on the soft optical fiber.It reduced the light transmittance but led to a greater increase in light intensity when the textile was under tension.The number of layers of water-soluble stabilizers determined the tightness of the embroidery.Less or no stabilizer created tighter stitches, which resulted in a bigger difference between the released status and strained status (Figure 5d).
The gauge factor observed in the lab tests (≈5) exceeded that of most carbon nanotube (CNT)-filled elastomeric strain gauges for the strain range of interest, [53][54][55][56][57] but fell short of the gauge factors exhibited by highly sensitive systems based on carbon black (CB), [58,59] graphene, [60,61] silver nanowires (AgNW), [62,63] polypyrrole/polyvinyl alcohol ink, [64] 2D transition metal carbides and nitrides (MXene), [65] or liquid metal. [66]In comparison to strain gauges employed for wearable respiration monitoring, which typically range from 0.9 to 522, [67][68][69] the current system's performance lies in the low-to-middle range.When compared to polymer-based optical waveguides, the gauge factor is similar to or lower than other strain gauges, [35,45,46,70] yet higher than FBG-based strain gauges [71,72] and the reported sensitivity of fiber-optic-based respiratory sensors. [48,73]Overall, the number of rooms was the most significant factor in gauge factor.However, it is also important to optimize the other parameters to maximize the sensitivity, because only a limited number of rooms can be fabricated within the relatively flat side of the front torso.

Reliability and durability
The reliability and durability of the current sensor using OL and machine embroidery were demonstrated through tensile, abrasion, and washing tests to simulate potential adverse scenarios to which a textile-based respiratory sensor could be subjected over its lifespan.An extension-controlled cyclic tensile test confirmed the consistency in signal changes for up to 500 cycles of strain (Figure 6a).The low drift implied that respiratory volume monitoring using this current sensor is promising (Figure 6b).The signal trends by strain were consistent under different tensile extension speeds, which showed the capability of the sensor to accurately detect an abnormal breathing pattern such as apnea, hypopnea, and/or hyperpnea (Figure 6c).Meanwhile, the sensor had a consistent signal level under 60 s of strain loading (Figure 6d), which could be useful for monitoring breath-holding behaviors. [74]Because strain sensors for respiratory monitoring need to be tightly fitted to the torso, the effect of abrasion on the signal was examined, assuming that the sensor would be worn as innerwear.The signal noise caused by abrasion with a pressure of 9 kPa was not significant (Figure 6e), especially when considering that common garment layers will not create much pressure on the innerwear.
In terms of durability, the embroidery enclosure made of nylon thread suffered only minor damage after 100 000 abrasions.None of the threads and only a few fibers were broken (Figure 7a).The tensile tests showed that even though the sensitivity decreased under strains larger than 10%, the sensor still functioned, and the gauge factor was sufficient to monitor respiration (Figure 7b).However, the other two types of common embroidery threads, polyester and rayon, could not withstand more than 5000 abrasions (Figure S3, Supporting Information).Ten cold machine washes using detergent did not cause any noticeable changes in the embroidery threads or OL (Figure 7c; Figure S4, Supporting Information).In addition, there was no degradation in the sensitivity to strain over the repetitive washing (Figure 7d).

Conclusion
In this study, machine embroidery was used to install a stretchable light guide called OL on an elastic textile to create a wearable respiratory monitoring sensor.The embroidery with enclosed spaces guided the OL to form a serpentine shape when the textile was free of tension (i.e., exhalation), and to react to the ribcage expansion during inhalation by being straightened.The intensity of the light passing through the fiber changed in correspon-  dence to the respiratory volume trend, demonstrating a high correlation even when the wearer was running (r 2 = 0.872), with the assistance of a neural network.The sensor was used to accurately measure the respiratory rates of 13 human participants (absolute error < 1.513 BPM) when the participants were sitting or walking.The machine embroidery used in the current study consisted only of zigzag stitches.Thus, the enclosure did not restrict the original stretchability of the elastic textile.Furthermore, the parameters of the embroidery, including its shape and stitch density/tightness, could be used to adjust the sensitivity of the OL to strain.The sensor was durable and reliable against abrasions, repetitions, and machine washes.
The primary population who could benefit from this sensor would be those with health risks in daily life due to abnormal respiratory symptoms such as hyperventilation and/or hypoventilation.Because this shirt-type sensor could be used with innerwear, the real-time respiratory monitoring results could be shared with the wearer, caregivers, and medical professionals while the patient is continuing their daily routine.Those who put importance on their physical performances would be another major target group, including but not limited to athletes, military personnel, astronauts, and the general public who love sports.This study only investigated RV and RR, but the raw signal could be processed into key health performance indicators according to the interest of the population, to contribute to improvements in physical activities.Because respiration is an indicator of an individual's emotional condition, those interested in their mental health and/or meditation would be able to gain insights using the current sensor by associating the readings with other sources like heartbeats.
The machine embroidery enclosure is easy to fabricate, costeffective, fast, textile-friendly, and durable enough to launder.Installing a strain sensor on a stretchy textile while preserving its original elasticity and that of the textile is often challenging, but the embroidery consisting of only a zigzag pattern allowed the OL to be embedded on the textile surface while still behaving like a part of the textile.The computerized design and fabrication made the parameter adjustment easy and yielded consistent results, unlike most manual garment construction techniques.A significant benefit of the OL is its stability against sweat, which contributed to the breathability, softness, and comfort of the final product.The current study did not add any cladding to the OL, but cladding would increase the durability and stability of the OL against abrasion, moisture, and external shock.
While the machine embroidery made of zigzag stitches showed an advantage in terms of elasticity, if the embroidery was large enough, the OL sometimes stuck out between the threads.This rarely occurred during the actual usage for respiratory monitoring.Rather, most instances were after machine washing.Extensible stitch types other than zigzag stitches may be able to prevent this issue.While the OL and embroidery enclosure were resistant to water, the other electronic components like the light source, photodetector, and others on the circuit board were still susceptible to moisture.In this study, the light source and photodetector were permanently attached to the OL.In future developments, modular design to create detachable, but also stable and consistent connections between the OL and electronics will be an important task to guarantee the sensor reliability over time and convenience for maintenance.The current study predicted the respiratory volume through calibration using the data from a commercial spirometer.Therefore, an easier way to quickly calibrate the respiratory volume should be developed and examined with the current sensor (e.g., a smartphone application predicting respiratory volume through breathing sounds).Lastly, the current system aimed to monitor everyday activities such as sitting, walking, and running with a limited sampling rate (100 Hz), but for more advanced features to instantly respond to abnormal events such as obstructive or paradoxical breathing behaviors, a more elaborated experiment design and a higher sampling rate will be necessary according to the feature of the target symptoms.
The present research introduced a machine embroidery enclosure that supported the shape transitions of a stretchable light guide as a result of strain.Although we used it to track ribcage movement for respiratory monitoring, its potential is open to most textile-based wearable or robotics applications involving strains.The embroidery enclosing the OL will be able to measure human joint movements or the stretching behaviors of soft systems.Furthermore, the enclosure is not limited to optical fiber, but could be used for any other non-stretchable but soft wearable electronics in the form of a fiber or tube, including sensors, wires, actuators, and/or batteries, to make them flexible and stretchable on textiles.

Experimental Section
Materials and Fabrication: A thermoplastic elastomer manufactured by Crystal Tec, Inc. was used as a thin, soft, and stretchable light guide (n = 1.54,D = 1.2 mm) without cladding. [45]An LED (B07QXR5MZB; BO-JACK) and a photodiode (SFH-229; Osram Opto Semiconductors) transmitted/detected a visible red light (≈650 nm) compatible with the fiber. [35]he lens of the LED and the photodiode were drilled to create a 1.2 mmdiameter hole to plug in and glue the optical fiber.A 2 kΩ potentiometer was used to adjust the brightness of the LED to an appropriate level, and an I-V converter consisting of an OP-AMP (LM324A; Texas Instruments), a 1 MΩ resistor, and a 4.7 nf capacitor delivered analog signals from the photodiode to the Arduino Nano 33 microcontroller using bluetooth low energy (BLE) (Figure S5, Supporting Information).A portable battery powered the microcontroller and whole system, and the microcontroller delivered the signals wirelessly to another Arduino Nano 33 BLE connected to a laptop through BLE.
The vector graphic software Inkscape and its add-on package Ink/Stitch supported the design of the embroidery.The dimensions of the embroidery design can be found in Figure S6 (Supporting Information).A piece of elastic (width = 25 mm, thickness = 1.5 mm) on top of six layers of a non-woven water-soluble stabilizer (thickness = .16mm) was loaded onto the embroidery frame holding the textiles tightly.A programmable embroidery machine (Brother PE770) was used to create the embroidery on the elastic band according to the design, using an all-purpose polyester thread (D = 0.06 mm, Dual Duty All Purpose Thread, Coats & Clark).When the machine was doing the embroidery, the fiber optic connection to the LED/photodiode was aligned on the elastic band manually.Thus, the embroidery enclosed the OL without penetrating it.
After dissolving the water-soluble stabilizers in warm water, the sensorembedded elastic band was dried and sewn on the under-chest or waist circumference of a sleeveless compression shirt.There were three shirt sizes per sex, and their measurements can be found in Table S1 (Supporting Information).To prevent excessive strain on the sensor during donning/doffing, the shirt was cut in half at the center of the back, and a zipper opening was added.In addition to the shirt-type sensors, belt-type sensors were fabricated by connecting length-adjustable buckles at each end of the elastic band, without sewing it on a shirt.
Uniaxial tensile test: Following ASTM D5035-11 (Standard Test Method for Breaking Force and Elongation of Textile Fabrics, Strip Method), [75] a tensile testing machine (Instron 5566, Instron, Norwood, MA) tested the sensor's sensitivity to the strain of the textile ( , where L is the current textile length under the elongating force, and L 0 is the original length).The size of the specimen was modified to 25 (width) × 200 (length) × 1.5 (thickness) mm, considering the dimension of the sensor-embedded shirt.A gripper held each end of an elastic band.After preloading 0.1 MPa at a speed of 2 mm min −1 , the specimen was elongated to 25% strain at 300 mm min −1 .The end strain and extension speed were modified to examine the effects of the extension speed (100 and 500 mm min −1 ) and cycles at different strain levels (2.5, 5, and 20%, 500 mm min −1 ).The microcontroller collected the signal from the photodiode at 100 Hz.
Human Participant Test and Data Processing: The human participant tests were conducted following the protocols approved by the Institutional Review Board (IRB) of the University.A total of 13 healthy adults participated in this study (sex: 7 females and 6 males; age: 36.5 ± 12.1 years, height: 171.6 ± 11.4 cm, weight: 69.6 ± 13.0 kg, BMI: 23.5 ± 2.9).After providing consent, each participant donned a sensor-embedded sleeveless compression shirt in their preferred size, on top of one layer consisting of a t-shirt.The fit of the shirt was adjusted to place the sensors at the under chest and waist levels.In addition, the researcher helped the participant don two belt-type sensors on the upper chest and waist levels (Figure S7, Supporting Information).The belt for the waist was placed right above or below the sensor embedded in the shirt.The microcontroller and power source were attached to the back of the participant's waist using Velcro.The microcontroller read signals and sent them at 100 Hz to the other microcontroller connected to a laptop through BLE, but the frequency of the received signals varied from 60 to 100 Hz according to the environment and body movements.
Each participant was asked to perform four tasks: 1) sit comfortably and breathe normally 2) sit comfortably and breathe deeply, 3) walk on a treadmill at a preferred speed, and 4) run on a treadmill at a preferred speed.There were three sessions for each task, and each session took ≈60 s.To collect the ground truth data, the participants used a hand-held portable spirometer (Spirotel, MIR International, France) connected to a disposable filtered mouthpiece, along with a nose clipper to ensure that air only flowed through the mouth.The filter in the mouthpiece decreased the overall airflow by < 1.5 cm H 2 0 L −1 s −1 at 14 L s −1 , and it was necessary to protect the participants during the pandemic. [76]The mouthpiece, filter, and nose clipper were discarded, and the spirometer was sanitized using 3% hydrogen peroxide immediately after each experiment. [77]The spirometer collected the respiratory data at 10 Hz for 60 s.
The light signals from the OLs were normalized within each session and smoothed by averaging over a five-point moving window.The sum of the normalized light intensity from the chest and waist was used as the main signal for each type of sensor: shirt or belt.The frequency with the highest amplitude retrieved from the fast Fourier transform (FFT) for each session (60 s) was used as the respiratory rate of the session.After joining the two datasets (the OLs and spirometer) based on the corresponding timestamps, the correlation coefficient was computed for each session.All the data processing and analysis described so far were done in MATLAB.
The neural network MLP Regressor model from the Scikit-learn library was used in Python. [78]The normalized signal from the shirt (i.e., the sum of the signals from the sensors installed at the chest and waist levels of the shirt) from t -10 to t -1 (about 1 second time window) was provided as the input to reduce the effect of the noise in prediction. [79]The respiratory volume from the spirometer worked as the output after normalization.Because each participant had a different fit with the pre-manufactured compression garment, the training and testing were done for each participant.Among three sessions for each task, data from two sessions (≈120 s) were used to train the model, and the other session (≈60 s) was used for the model test.The neural network trained the model until convergence (or max.1,000 iterations) with the Adam optimizer.The hidden layer size was 64 × 64 × 64, and a rectified linear unit (ReLU) performed as an activation function.
Abrasion Test: The abrasion resistance of the embroidery and the strain-sensing capability of the system were examined using the M235 Martindale abrasion tester, according to the modified ASTM D4966-12 Standard Test Method for Abrasion Resistance of Textile Fabrics (Martindale Abrasion Tester Method). [80]The belt-type sensor (OL enclosed by the embroidery on elastic, without being sewn on a shirt) was fabricated using three different threads (the polyester used for all the other tests in this study, rayon, and nylon).The sample size for each thread type was three.The samples were attached to the bottom of the moving textile holder with the OL and embroidery facing downward, instead of clamping the sensor into the holder, to prevent damage to the OL.The moving holder rubbed the sensor surface onto the standard rugged woven textile (D = 13 cm) clamped on the other side.The weight connected to the top of the holder added a pressure of 9 kPa to the rubbing surfaces, and the speed of the abrasion was 47.5 rpm.Because the diameter of the bottom of the holder (5 cm) was shorter than the length of the embroidery (10 cm), the abrasions only affected the center of the sensor (Figure S8, Supporting Information).
A portable microscope (USB2-MICRO-200X; Plugable Technologies, Redmond, WA) was used to obtain close-up images of the embroidery surface before the experiment, and after 5000, 10 000, 20 000, 50 000, 75 000, and 100 000 abrasion cycles.In addition, following the same uniaxial tensile testing protocol except for the specimen dimension (25 × 100 × 1.2 mm) and preload (0.025 MPa), the strain sensitivity was examined every time after taking close-up images.
Machine Washing Test: Following the AATCC LP1-2018e Laboratory Procedure for Home Laundering: Machine Washing, the washability of the system was tested.The standard normal/cold washing procedure includes 27 °C water, machine washing for 16 min, a final spin at 660 rpm for 5 min, 66 g of 1993 AATCC Standard Reference Detergent, and laundering ballast to meet the total laundry weight level of 1.8 kg.A Vortex M6 (SDL Atlas, Rock Hill, SC) standard washing machine took ≈40 min for one cycle of washing in total.Four sensor samples embedded on compression shirts were laundered along with the ballast, and each was hung on a hanger and dried for approximately 1 h before going through the next cycle of washing.The strain sensitivity of the samples was tested following the same uniaxial tensile test protocol, before the laundering and after 5 and 10 washes.

Figure 1 .
Figure 1.Design overview and sample images.a) Sensor design and principle sensing the strain of the textile caused by respiration.b) Embroidery design (top), a close-up image of the fabricated sensor (middle), and the sensors embedded on a compression shirt (bottom).

Figure 2 .
Figure 2. Examples of respiratory volume predictions using raw signal readings as inputs for the MLP model and the ground truth from a commercial spirometer (n = 1).a) Sit and breathe normally.b) Sit and breathe deeply.c) Walk on a treadmill.d) Run on a treadmill.

Figure 3 .
Figure 3. Respiratory volume measurement (n = 13).a) Pearson correlation coefficients between sensor type and the ground truth and RMSE of the prediction."Both" was the sum of the sensors on the chest and waist, and "Prediction" was based on the signals from the shirt and MLP model.b) Correlation coefficient comparison by sex and posture, in case of using "Shirt-Both".c) Correlation coefficient and RMSE distribution by BMI and sensor type, in case of sitting-breathing normally.

Figure 4 .
Figure 4. Respiratory rate measurement.a) Respiratory rate from a commercial spirometer and absolute error from the sensor by posture and sensor type (n = 13).b) An example of FFT analysis of the signals from the shirt-type sensor by posture (n = 1).c) Signal-to-noise ratio by posture and type of sensor (n = 13).

Figure 5 .
Figure 5. Gauge factor by embroidery parameter adjustment.a) Embroidery shape (n = 3).b) Number of the enclosing rooms deciding the number of curves of OL (n = 3).c) Stitch spacing that decides the density of stitches (n = 3).d) Number of water-soluble stabilizers between the embroidery and the elastic band, deciding the tightness of the embroidery (n = 3).

Figure 7 .
Figure 7. Durability tests.a) Sensor after 100 000 abrasion cycles.Only a few fibers were broken in the red-marked circle area (n = 3).b) Tensile test result over the abrasion cycles.c) Sensor after 10 cold machine washes.d) Tensile test results over the laundry cycles (n = 5).
Statistical Analysis: The light intensity was normalized based on each trial both in human and lab testing.The plots indicate the trend of mean when n > 1 with the error bars showing the standard error.The sample number is in the caption of each figure.