System overview
Figure 1a illustrates multimodal applications of AusculPatch that can be attached to multiple sites of the human skin and capture a wide range of physiological mechano-acoustic signals for remote health monitoring and HMI. AusculPatch offers conformal skin contact by adopting an island-bridge structure applied to a flexible printed circuit board (fPCB), which is encapsulated within a stretchable silicon elastomer, Fig. 1b and Supplementary Figs. 1 and 2. The foldable serpentine interconnect in the island-bridge configuration decouples the sensing element from the bulky electronic components, such as the microcontroller and battery, thereby reducing mechanical noise arising from skin-induced perturbations (Fig. 1c).
Fig. 1: Design and architecture of the AusculPatch system.
The alternative text for this image may have been generated using AI.
a Application of the broadband AusculPatch device: (top) for remote healthcare monitoring and (bottom) for human–machine interaction. The user interface displays the physiological signals, including heart pulse waves, heart sounds, SCG cardiac, breathing patterns, and Korotkoff sounds. Created in BioRender. Dang, T. B. (2026) https://BioRender.com/s0u64su. b The exploded view with multiple layers of deposited materials of AusculPatch. c The zoomed-in photo on the cantilever acoustic sensor, from bottom to top: CPT attached to an fPCB island with a diameter of 8 mm, the couple of sensing elements on a single CPT, the cantilever sensor with 1-um gaps surrounding. d Block diagram showing a wearable, wireless, broadband AusculPatch attached to a human’s skin to capture body acoustic signals, e Image of the AusculPatch with 30° bending, and f FEA results, g Image of AusculPatch with 20% stretching on serpentine connection, and h FEA results.
Figure 1d presents the acoustic sensing mechanism of AusculPatch, incorporating a cantilever pressure-transducer (CPT) mounted on a conical air chamber, which is covered by a thin elastic membrane to facilitate the collection of body signals. CPT enables broadband detection of physiological mechano-acoustic signals by tailoring an architecture that integrates a Si cantilever, surrounded by ultra-narrow gaps (1 µm). Low-frequency components arising from skin deformation induced by pulsatile blood flow or respiratory-driven chest wall motion cause a deflection in the chamber membrane in direct contact with the skin. The narrow gap traps and suppresses air inside the chamber, which exerts pressure on the cantilever. The resulting cantilever bending leads to a resistance change in the doped Si piezoresistive layer, enabling the detection of subtle, low-frequency physiological signals that are typically beyond the detection limit of conventional MEMS microphones. In parallel, higher frequency signals such as heart sounds and vocal cord vibration can transmit through the chamber, focusing acoustic energy onto the Si cantilever through the conical structure. The high gauge factor of Si, coupled with a strategically engineered geometry with narrow legs (10 µm width and 30 µm length) and a large air pad (80 µm × 80 µm) (see Supplementary Table 1), enables a highly sensitive piezoresistive effect in response to acoustic waves. These resistance variations arising from both low- and high-frequency modes are converted into output voltages using a Wheatstone bridge, followed by analog amplification and digitalization through an onboard microcontroller for downstream processing and analysis. In addition to sensing capabilities, mechanical flexibility and stability of electrical interconnection are critical factors for continuous wear. We conducted finite element analysis (FEA) using COMSOL Multiphysics, confirming the mechanical robustness of AusculPatch. As shown in Fig. 1e–h, the serpentine interconnect sustains minimal strain of less than 5% under deformation, including 30° bending and 20% stretching, indicating the resilience of the bridge-island configuration for reliable, high-fidelity signal acquisition on complex anatomical surfaces.
Layout the fPCB and the working principle of the cantilever pressure transducer
Figure 2a shows an assembled fPCB incorporating the CPT chip and electrical circuitry of AusculPatch (see Methods). All components, including CPT, fPCB circuit, air chamber, and battery, are encapsulated inside polydimethylsiloxane (PDMS) seals (Fig. 2b), forming a flexible, compact (20 × 47 × 3.5 mm) and lightweight (3.2 g) form factor, capable of simultaneous data acquisition and wireless transmission. Figure 2c depicts details of the circuit schematic and system block diagram. Figure 2d presents bright field microscopy images of the CPT (1.5 × 1.5 × 0.3 mm). CPT is fabricated from a SOI wafer (see Supplementary Fig. 4) and integrates two cantilever elements on a single chip. One was released from the buck Si serves as the primary acoustic detector (Fig. 2d, top right), and the other fixed on the substrate functions as a temperature compensation component, respectively (Fig. 2d, bottom right). The piezoresistive layer was formed on the top ~50 nm of the 300 nm Si device layer using ion implantation, thereby maximizing its distance from the neutral mechanical axis and enhancing its strain sensitivity (Fig. 2f). The relationship between the pressure input and the corresponding resistance changes is estimated using an analytical model. Figure 2g shows the FEA simulated strain distribution across the cantilever structure under a 40 Pa pressure load, revealing a deformation of 7.7 μm at the cantilever tip and an average strain of 2.35 × 10−4 in the piezoresistive region, in strong agreement with the analytical calculation (Supplementary Figs. 5 and 6). The resulting fractional resistance change is derived from the geometry parameters of the cantilever as schematically illustrated in Fig. 2h (Supplementary Note S1):
$$\frac{\Delta R}{R}=\Delta P\times \frac{2\times {\pi }_{l}\times {h}_{p}}{{{t}_{c}}^{3}}\times \left[{{l}_{h}}^{2}+\frac{3\frac{{w}_{p}}{{w}_{h}}\times {l}_{p}({l}_{h}+{l}_{p})}{2}\right]$$
(1)
where \({{{\rm{h}}}}_{{{\rm{p}}}}\) is the distance between the sensing layer and the neutral axis, \({{{\rm{\pi }}}}_{{{\rm{l}}}}\) is the longitudinal piezoresistive coefficient. For the n-doped <100> Si formed using ion implantation (Arsenic doped) with a carrier concentration of ~1019 cm−3, \({\pi }_{l}\) is expected to be ~85 cm2/dyne37. From Eq. 1, the output of the sensors per unit pressure (Pa) is estimated at \(\frac{\Delta R}{R}/\Delta P\) ~ 6.8 − 8.4 × 10−4 Pa−1 for the diffusion depth ranging from 50 to 100 nm (Supplementary Fig. 7).
Fig. 2: Design of fPCB and cantilever-based acoustic sensor analysis.
The alternative text for this image may have been generated using AI.
a Photograph of the AusculPatch’s thin, flexible PCB. b Photograph of a flexible PCB without (top) and with encapsulation (bottom). c Circuit and block diagrams of the platform and the user interface communicating through the BLE radio. d Cantilever pressure transducer (left) with zoomed-in free-standing and fixed cantilever elements. e Block diagram of the multi-position cardiorespiratory monitoring scheme: throat and arteries auscultation (left) and chest wall auscultation (right). f–h The working principle of CPT: f figure showing a cross-sectional view of CPT, g FEA results of strain simulation of the free-standing cantilever element under 40 Pa pressure, and h 3D design parameters definition of CPT.
The resonant frequency of the Si cantilever is estimated using the Rayleigh-Ritz method, yielding a theoretical first mode at 11.9 kHz (Supplementary Note S2), in good agreement with FEA data (Supplementary Fig. 8). This prediction closely matches experimental measurement obtained using a piezoelectric excitation setup (Supplementary Fig. 9). This high first mode resonance, together with the engineered micro airgap architecture, enables broadband frequency detection in AusculPatch. The multimodal sensing capability of AusculPatch is facilitated by applying digital frequency filters, establishing an all-in-one auscultation platform (see Fig. 2e and Methods). Using a single sensing element, AusculPatch effectively consolidates the functions of multiple conventional sensors, such as 3D pressure sensors (for pulse wave), IMU sensors (for low-frequency vibration), and MEMS microphones (for high-frequency signals). Beyond its multimodal capability, AusculPatch offers a small footprint and low-power consumption (4.5 mW), outperforming multi-sensor systems reported in the literature (Supplementary Fig. 10)16,17,38,39,40. Supplementary Table 2 provides a comprehensive comparison of the measurement ranges and form factors between AusculPatch and other devices reported in the literature.
Fundamental characteristics of AusculPatch
To validate the capability of the AusculPatch for physiological signal acquisition and its suitability for long-term wear, we conducted a comprehensive series of characterizations. First, we assessed the electro-mechanical responses of AusculPatch under static pressure conditions, demonstrating both high sensitivity and durability. Next, we investigated the acoustic characteristics of the system to optimize structural parameters such as the air chamber geometry for effective broadband measurement. Finally, a thermal drift compensation strategy was developed to mitigate signal fluctuation due to ambient temperature changes, ensuring stable performance in long-duration use.
Figure 3a presents the experimental setup for evaluating the response of CPT to static pressures, where the sensor was connected to a precision pressure generator (KAL200, Halstrup-Walcher GmbH). As pressure increases from 0 to 40 Pa, the resistance of the Si cantilever increases accordingly, and returns to baseline upon pressure release, indicating reversible, low hysteresis behavior, Fig. 3b. To demonstrate the ability of CPT in capturing sub-Hz frequency, we conducted a comparative frequency sweep from 0.025 to 10 Hz, Supplementary Fig. 11. The fractional resistance change in both static pressure test (Fig. 3c) and sub-Hz frequency test (Supplementary Note S3) exhibits a linear relationship with the applied pressure, yielding a sensitivity of \((\varDelta R/R)/\varDelta P\) ≈ 7 × 10−4 Pa−1, well matching with estimations from the numerical analysis. A linear current-voltage curve confirms good Ohmic contact between the Cr/Au electrodes and the doped Si layer, indicating that resistance measurements are unaffected by applied bias, Fig. 3d. Figure 3e presents the response of CPT over 94,000 cycles at 10 Hz and under a loading amplitude of 100 Pa, showing the stable performance of the CPT (see Supplementary Note S4).
Fig. 3: Characterizations of the system.
The alternative text for this image may have been generated using AI.
a–c Pressure sweep test using a pressure generator: a experiment setup, b fractional change of the resistance in real-time, and c relationship between the fractional change of the resistance and the applied pressure. Data were presented as mean ± SD. The error bars refer to the two standard deviations on the basis of four contemporary measurements from the same experimental unit (n = 4). d The I-V curve of the voltage sweep test. e Variation of sensor resistance testing at 10 Hz frequency, 100 Pa over 94,000 cycles. f Chamber design: f1 Design parameters and f2 theoretical frequency characteristics of the system. g Chamber volume characterization using a linear actuator: g1 Experimental setup and g2 Magnitude plot of resistance change v.s. frequencies. Data were presented as mean ± SD. Solid lines represent the mean magnitude across four-cycle measurements from each frequency and chamber volume (n = 4), and shaded areas represent two standard deviations in magnitude. h Membrane diameter characterization using a speaker: h1 Experimental setup and h2 Fractional change of the resistance response to different sizes of membrane diameter (8, 12, and 16 mm) at the sound frequency 40–100 Hz. Data were presented as mean ± SD. The error bars refer to the two standard deviations on the basis of 100-cycle measurements from the same sound frequency and membrane size (n = 100). i Resistance changes under temperature variation (heating up to 40 °C for 150 s then cooling down to 25 °C) of the fixed cantilever and the free-standing cantilever (top), and the free-standing cantilever without and with seven repeated cycles of 20 Pa pressure loading (bottom). j, k The thermal response of the fixed-resistor Wheatstone bridge (top) and the thermal compensation Wheatstone bridge (bottom): circuit diagram (j) and the output voltage during seven repeated cycles of 20 Pa pressure loading, with the thermal drift level indicated in the inset (k). The thermal drift for each circuit is defined as the difference in output voltage under no-load conditions between 40 and 25 °C. Data were presented as mean ± SD. The error bars refer to the two standard deviations on the basis of a seven-cycle unloading measurement (n = 7).
To further optimize the design of AusculPatch, we investigated the influence of the chamber geometry and the membrane dimensions on the acoustic sensing performance. The narrow gap surrounding the cantilever allows air leakage from the chamber under pressure variations, serving as a mechanical high-pass filter that attenuates low-frequency signal amplitudes, where the cutoff frequency is influenced by the chamber volume27. To explore this behavior, we modeled the system with the Si cantilever located between two adjacent air chambers of volumes \({V}_{1}\) and \({V}_{2}\), Fig. 3f1. Respiratory-induced chest expansion (typically with a frequency around 0.2 Hz41) deforms the elastomeric membrane, changing the volume of Chamber 2 to \({V}_{2}+\Delta {V}_{2}\). This change in the volume results in a pressure difference \(\Delta P\) between the two chambers that drives air leakage through the narrow gap at a volumetric flow rate \(q\) (\(q \sim k\Delta P\), where \({{\rm{k}}}\) is a leakage coefficient). By applying the ideal gas law (\({PV}={nRT}\)), we model the cantilever with a small airgap as a single-input (\(\Delta {V}_{2}\)) and single-output (\(\Delta P\)) system, and establish its transfer function \({G}_{1}\left(s\right)=\Delta P/\Delta {V}_{2}\) using the Laplace transform as detailed in Supplementary Note S5. The model indicates that the transfer function attains a lower high-pass cutoff frequency by increasing the volume of the upper chamber \({V}_{1}\) infinite. Experimentally, this condition can be approximated by venting the upper chamber (\({V}_{1}\)) to the ambient atmosphere. Consequently, the cutoff frequency of the system is inversely proportional to volume \({V}_{2}\):
$${{{\rm{f}}}}_{{{\rm{c}}}}={{\rm{k}}}{{{\rm{P}}}}_{{{\rm{atm}}}}\times \frac{1}{{{{\rm{V}}}}_{2}}(\frac{{{\rm{rad}}}}{{{\rm{s}}}})$$
(2)
Equation(2) suggests that increasing the volume of the air chamber V2 reduces the cutoff frequency (i.e., a larger chamber yields a lower cutoff frequency and consequently improved low-frequency performance), as validated in the following experiments. To characterize the low-frequency response of AusculPatch, we developed three cylindrical chambers with identical diameters (10 mm) but with different heights of 1 mm (\(25{{\rm{\pi }}}\,{{{\rm{mm}}}}^{3}\) volume), 2 mm (\(50{{\rm{\pi }}}\,{{{\rm{mm}}}}^{3}\) volume), and 4 mm (\(100{{\rm{\pi }}}\,{{{\rm{mm}}}}^{3}\) volume). We programmed a linear actuator (X-LSM025A-E03, Zaber) to generate sinusoidal pressure variations within each chamber over a frequency sweep from 0.05 to 10 Hz, as shown in Supplementary Figs. 13 and 14. The experimental data show that at a low differential pressure, the leakage coefficient was found to be \(k \sim 0.023{l}\times {s}^{-1}\times {{Pa}}^{-1}\) (detailed calculation is available in Supplementary Note S6 and Supplementary Fig. 12). Based on this value and Eq. 2, we calculated the theoretical cutoff frequencies of the 1, 2, and 4 mm chambers to be 0.46, 0.23, and 0.12 Hz, respectively. Figure 3g2 overlays the experimental frequency response data and the theoretical model, showing a good agreement between the measured and calculated cutoff frequencies. The data validates the leakage model in describing the low-frequency behavior of the sensor. To preserve the compact and wearable form factor of AusculPatch while maintaining sufficient sensitivity to low-frequency respiratory signals (~0.2 Hz), the chamber volume is constrained at ~160 mm³. Adopting this chamber volume results in a cutoff frequency of ~0.2 Hz, sufficiently low to capture subtle respiratory and pulse wave signals.
For the high-frequency characterization, we employed a speaker (Fostex, FF125WK) to generate sound at 40−100 Hz, matching the dominant frequency range of human heart sounds, as the experimental setup shown in Fig. 3h1 and Supplementary Fig. 15. The generated sounds pass through the phantom skin and device chamber and reach CPT in the form of mechano-acoustic pressure, thereby replicating the propagation of body sounds through the human chest wall42,43. As the membrane diameter determines the effective area for collecting acoustic energy, we fix the volume of the chamber at the previously determined optimal value and vary the chamber diameters of 8, 12, and 16 mm to investigate the influence of the membrane size on the obtained signal, Fig. 3h2. It is worth noting that the variation in sound amplitude across different frequencies arises from the mechanical properties of the phantom skin. Experiment results show a monotonic increase in signal amplitude with increasing membrane diameters, indicating enhanced acoustic capture efficiency. Based on these findings, we select the 16 mm membrane diameter while preserving user comfort and conformability for long-term wear.
Although the bare CPT sensor (i.e., without the air chamber) exhibits a high intrinsic resonance at ~12 kHz (Supplementary Fig. 9), integration of the acoustic chamber modifies the overall dynamic response of the system. In particular, air volume under the cantilever introduces additional damping, which can reduce the effective resonance frequency and narrow the overall bandwidth. To quantify this effect, we characterized the packaged device using the experimental setup shown in Supplementary Fig. 16. The vibration amplitude and mode shapes of the cantilever were measured using Laser Doppler Vibrometry (MSA-050, Polytec). The magnitude and phase response indicate a damping ratio of ~0.48 is associated with the acoustic packaging. Although this damping shifts the resonance frequency downward, the resulting resonant peak remains above 10 kHz (Supplementary Note S7), thereby preserving the broadband sensing capability of AusculPatch. Building on this system-level dynamic characterization, we next quantified the acoustic bandwidth of the optimized device using the setup shown in Supplementary Figs. 17 and 18a. A nearly flat band piezoelectric actuator (AE0203D18H18DF and KEMET) was used to apply a consistent, frequency-varied displacement to the chamber membrane from 1 Hz to 20 kHz, while the corresponding piezoresistive output was recorded using an oscilloscope (Analog Discovery 3, Diligent). The selected AusculPatch chamber (i.e., ~160 mm³) exhibits a resonance frequency at ~4.8 kHz and maintains a flat frequency response up to 2 kHz, as verified by the measured piezoresistive output voltages at discrete frequencies of 50, 100, 200, 500, 1 kHz, and 2 kHz (see Supplementary Note S7). These results confirm that, despite the damping introduced by acoustic packaging, the device preserves a broad operation bandwidth. Such performance enables detection of high-frequency physiological signals, including higher-order components of heart sounds, and vocal harmonics extending into the 1–2 kHz range.
It is well-known that piezoresistive sensors are not only sensitive to strain but also responsive to temperature due to the thermoresistive effect44. To mitigate this phenomenon, our thermal compensation architecture includes a free-standing cantilever for mechano-acoustic sensing and a non-released cantilever with the same shape for temperature cancellation (with the same resistance of 4.3 kΩ). These two cantilevers exhibit approximate resistance variations as the temperature increases from 25 to 40 °C, Fig. 3i (top). The Si piezoresistive sensor is then connected to a Wheatstone bridge, providing an output voltage of:
$${V}_{O}=\left(\frac{{R}_{1}}{{R}_{1}+{R}_{4}}-\frac{{R}_{2}}{{R}_{2}+{R}_{3}}\right)\times {V}_{{REF}}$$
(3)
We simulated two distinctive Wheatstone bridge designs to evaluate the performance of the thermal noise cancellation design, Fig. 3j. The first configuration connects the Si piezoresistive cantilever (\({R}_{1}={R}_{{SEN}}\)) to three fixed resistors \({R}_{2}={R}_{3}={R}_{4}=R\), with the same resistance of 4.7 kΩ, Fig. 3j (top). Meanwhile, the second configuration (i.e., the thermal compensation architecture) replaces the fixed resistor \({R}_{4}\) with the on-chip thermal compensation cantilever \({R}_{{REF}}\), Fig. 3j (bottom). In the former circuit design, because of the different responses of R4 and RSEN to temperature variations, the output voltage fluctuated under thermal effect, Fig. 3k (top). In the later design, the two Si cantilevers experienced the same local thermal effect, resulting in similar resistance variations, thereby enhancing the stability of the output voltage \({V}_{O}\). When gradually increasing the temperature of CPT from 25 to 40 °C for 300 s, the thermal compensation circuit effectively eliminates 95% of thermal drift errors, Fig. 3k (bottom). This improvement is particularly critical for long-term physiological monitoring, where precise and consistent readings are essential for reliable health assessments.
Blood pulse wave measurements
With the device design and architecture optimized in the previous section, we demonstrate the applications of AusculPatch in detecting blood pulse waves. We conducted benchtop validation using an artificial hemodynamic setup, consisting of a linear actuator driving a water-filled syringe that pumps water through elastomeric tubing (inner diameter 4 mm) embedded beneath a phantom skin layer (Ecoflex 00–31). Flow was modulated between 0.5–2 Hz, simulating physiological heart rates at 30–120 bpm. AusculPatch was then placed above the artificial vessel to capture the pulsatile signal and wirelessly transmitted the record to a nearby user interface via BLE. The signal obtained from AusculPatch exhibited consistent repeatability over multiple pressurization–depressurization cycles, confirming its reliability in detecting low-frequency pulsatile flow patterns (Supplementary Fig. 19).
Upon validation using the lab-built setup, we performed on-body measurements by attaching AusculPatch to healthy subjects on the wrist and the neck (Fig. 4a, b) using 3 M double-sided medical adhesive tape. The overlaid waveforms in both measurements illustrate the similarity and stability of the heart pulse waveforms collected from the carotid and radial arteries, yielding key physiological features such as systolic (SP), tidal (TP), diastolic peaks (DP), and a dicrotic notch45 (Movie S1). Supplementary Fig. 20 overlays the raw signal with the extracted pulse wave component (<10 Hz). The solid agreement between the original trace and the extracted pulse waveform indicates that the CPT captures the high-fidelity pulse morphology, including the systolic peak and dicrotic notch with minimal interference from the attenuated high-frequency components.
Fig. 4: Measuring body acoustic signals from arteries.
The alternative text for this image may have been generated using AI.
a On-throat measurement: photograph of the device attached to the throat (left), 30 s pulse wave recording, and overlaid waveforms of pulse wave (right). b On-wrist measurement: photograph of the device attached to the wrist (left), 30 s pulse wave recording, and overlaid waveforms of pulse wave (right). c–e Pulse wave measurement from carotid artery using AusculPatch, benchmarking with blood flow velocity measurement using ultrasound probe: c The ultrasound probe and AusculPatch was attached closely over human artery on the side of the neck, d Ultrasound image of the carotid artery depicting blood flow, shown in red, representing the flow from the heart to the upper body, and e Comparison between the blood flow velocity and the pulse waveform. Created in BioRender. Dang, T. B. (2026) https://BioRender.com/s0u64su. f, g Pulse wave measured from ten participants, benchmarking with measurements from an ECG bioamplifier: f Overlaid waveforms of pulse wave measured from ten participants, and g Correlation plot of HR measurement between AusculPatch and reference ECG illustrating linear curve fit with r2 = 0.95 and Bland–Altman plot with 0.15 bpm difference. h Continuous measurement of HR and IBI of AusculPatch in comparison with reference ECG measurement, with the indication of irregular heartbeat detected from the IBI plot. i–k Using AusculPatch to capture Korotkoff sounds from the brachial artery and assist BP estimation: i Experiment setup, AusculPatch was positioned below the cuff to detect Korotkoff sounds during arterial compression, thereby assisting in blood pressure estimation, an Omron device was used for validation, j Korotkoff sounds were detected using AusculPatch, along with the measurement of cuff pressure, determining the SBP and SDP values, and k Box plots showing the comparison between BP estimated with AusculPatch and reference BP measured using the Omron device (N = 5 participants, 5 trials per participant; total n = 25). Boxes show range between the 25th and 75th percentiles, whiskers show range between the 5th and 95th percentiles, and midline indicates the median of each dataset. Created in BioRender. Dang, T. B. (2026) https://BioRender.com/s0u64su.
To benchmark the low-frequency performance of AusculPatch, we developed wireless prototypes integrating commercial surface-mount IMUs (BMI270) and MEMS microphones (ICS-40212), both of which have been employed in previous studies (Supplementary Fig. 21)17,46. As shown in Supplementary Fig. 22, IMU-based devices, despite being designed for low-frequency measurements, failed to capture subtle physiological signals such as pulse waves. In contrast, the MEMS microphone was able to detect the pulse events but lacked sufficient resolution to accurately resolve the pulse waveform. These results further confirm that while AusculPatch offers clear advantages for detecting multiple classes of physiological mechano-acoustic signals compared to conventional MEMS microphones, it also enables selective sensing and extraction of signals of interest through the appropriate choice of attachment locations.
Comparison of pulse waves recorded from AusculPatch with blood flow velocity obtained from ultrasound scan at the carotid artery further confirms high temporal fidelity with minimal delay of our platform using a setup shown in Fig. 4c. The cross-sectional view in Fig. 4d depicts oxygenated blood pumped from the left ventricle toward the upper body in red color, and a contrast flow direction in blue color. We benchmarked the timing of pulse wave events \({\Delta {{\rm{t}}}}_{{{\rm{p}}}}\) (from the wave foot to the first systolic peak in the AusculPatch’s pulse waveform) and \({\Delta {{\rm{t}}}}_{{{\rm{f}}}}\) (a similar interval obtained from the ultrasound scan’s flow velocity waveform). \({{{\rm{ED}}}}_{{{\rm{p}}}}\) and \({{{\rm{ED}}}}_{{{\rm{f}}}}\) were defined as the intervals from the wave foot to the incisura of the pulse and flow waveforms, corresponding to the aortic valve closure event, Fig. 4e.
We found strong temporal associations between \({\varDelta t}_{f}\left(282.0\pm 19.5{{\rm{ms}}}\right)\) with \({\varDelta t}_{p}(285.0\pm 26.4{{\rm{ms}}})\), and \({{ED}}_{f}(89.1\pm 5.2{{\rm{ms}}})\) with \({{ED}}_{p}(91.6\pm 2.0{{\rm{ms}}})\), well agreed with other reports47 (see Supplementary Fig. 23). The experiment indicated that systolic upstroke and end-systolic timing were preserved across modalities, demonstrating the precision of AusculPatch in palpation.
Figure 4f presents overlaid pulse waveforms measured from the wrist of ten participants over 10 min. We compared the extracted HR with a reference device using a commercial ECG bioamplifier (ADInstruments, PowerLab) worn on the right arm (RA), left arm (LA), and left leg (LL), see Fig. 4g and Supplementary Fig. 24. HR measurements from both devices show strong agreement, with a linear fit (r² = 0.95) and Bland–Altman analysis indicating a mean difference of 0.15 bpm and a standard deviation of 1.5 bpm, Supplementary Fig. 25. We further assessed the influence of breathing activity on the HR when the participants performed two cycles of the Valsalva maneuver (30 s of breath-holding), see Fig. 4h and Supplementary Note S9. The comparison reveals a high degree of similarity between the heart rate and inter-beat interval measurements obtained from AusculPatch and ECG. Additionally, the precise heartbeat detection of AusculPatch enables the possibility of revealing irregular heartbeat events, characterized by alternating fast and slow beats. This situation can occasionally occur in healthy individuals, even without underlying heart conditions. However, the frequent presence of irregular heartbeat can be considered a symptom of heart failure with Arrhythmias48.
In addition to pulse waves, the high sensitivity of AusculPatch allows the detection of subtle flow-induced arterial vibrations such as Korotkoff sounds. The Korotkoff method remains a gold standard for noninvasive BP measurement49, employing a stethoscope to detect sounds generated by turbulent blood flow when cuff pressure on the brachial artery falls between systolic (SBP) and diastolic (DBP) blood pressure. The versatility of AusculPatch facilitates direct measurement of Korotkoff sounds, enabling BP estimation when positioned over the brachial artery. We used an Omron BP monitor (HEM7144T1, Omron) as a reference device and to supply pressure to the bladder cuff. The cuff pressure was measured using a reference pressure sensor (ABP2, Honeywell), while AusculPatch was positioned in close contact with the cuff to record the Korotkoff sounds, Fig. 4i. During cuff inflation and deflation, both cuff pressure and Korotkoff sounds were simultaneously recorded using the pressure sensor and AusculPatch, respectively (Fig. 4j and Supplementary Movie S2). BP estimated with AusculPatch in five healthy subjects (Fig. 4k) closely matched the values obtained with the Omron monitor, with a mean difference of 1.76 mmHg and standard deviation of 5.48 mmHg (Supplementary Fig. 26), demonstrating the multimodality of our device for peripheral artery palpation.
Cardiorespiratory measurements on the human chest wall
Attaching AusculPatch onto the chest wall enables monitoring of various cardiorespiratory signals, including chest movement caused by the respiration activities, SCG signals, and the heart sounds produced by the closure of heart valves (Fig. 5a). Figure 5b illustrates the breathing pattern and the computed respiration rates of a healthy individual over three states of breathing, showing the ability of AusculPatch in detecting various breathing states such as rapid breathing (25.5 bpm), breath-holding, and normal breathing (12.6 bpm). The algorithms for determining the respiration rate and heart rate are presented in Supplementary Figs. 27 and 28. Each breath cycle is represented by the upward and downward slopes of the waveform, corresponding to the expansion and collapse of the chest wall during the respiratory cycle (Fig. 5c).
Fig. 5: Cardiorespiratory measurements.
The alternative text for this image may have been generated using AI.
a Photo of the device mounted on the chest wall (Mitral area). Created in BioRender. Dang, T. B. (2026) https://BioRender.com/s0u64su. b Measured chest wall movement indicating the breath pattern over three conditions of breathing: rapid breathing, breath-holding, and normal breathing, lasting 20 s for each state. c Zoomed-in plot of two normal breathing cycles. d SCG waveform captured from the chest wall with the indication of the MC peak, AO peak, AC peak, and MO peak, showing strong temporal associations in cardiac timing events compared to accelerometer signals. e Heart sounds waveform captured from the chest wall, including two major cardiac sounds (S1 and S2), with high similarity in the time-domain waveform compared to stethoscope measurements. f A drawing illustrating four conventional heart auscultation positions. Created in BioRender. Dang, T. B. (2026) https://BioRender.com/s0u64su. g Recorded heart sounds and SCG signals from AusculPatch at four locations: A, P, T, M areas with the SNR at 12.17, 16.83, 16.70, and 14.02 dB, respectively. h, i HR measurement results from AusculPatch on five participants in 10 min. h Correlation plot of HR measurement between the wearable device and reference stethoscope illustrating linear curve fit with r2 = 0.95 and Bland–Altman plot with 0.1 difference. i Box plots showing HR distributions recorded within 6 min using AusculPatch (n = 201, 303, 234, 259, 271, corresponding to subjects #1 to #5, respectively). Boxes show range between the 25th and 75th percentiles, whiskers show range between the 5th and 95th percentiles, and midline indicates the median of each dataset. j 10 min continuous measurement of HR, BR, IBI, and breath intervals using AusculPatch.
The SCG signal, which reflects the micromovements of the chest wall induced by the pumping of blood with each heartbeat, occupies an inaudible frequency range of below 100 Hz and is crucial for assessing cardiovascular health through cardiac mechanical events, i.e., altered ventricular stiffness50. Figure 5d demonstrates the capability of AusculPatch to capture the SCG signal with clear peaks detected, including the closing of the mitral valve (MC), the opening of the aortic valve (AO), the closing of the aortic valve (AC), and the opening of the mitral valve (MO). These peaks exhibit strong temporal associations in cardiac timing events compared to accelerometers, which cannot be achieved using commercial microphones (Supplementary Fig. 36). The high-frequency components of SCG, heart sounds (~20–100 Hz), provide valuable insights into cardiac conditions, i.e., heart valve failure and heart wall stiffness. Figure 5e and Supplementary Movie S3 show the recorded heart sounds, highlighting the two primary cardiac sounds, S1 (corresponding to the closure of the mitral valve (MV), tricuspid valve (TV)) and S2 (corresponding to the closure of the tricuspid valve (TV), pulmonic valve (PV)), with the SNR above 14 dB. The waveforms of heart sound signals from AusculPatch and a commercial stethoscope (Littmann CORE Digital Stethoscope, Black Edition 8480) show a high similarity in the time-domain signals.
For the assessment of heart valves, AusculPatch was positioned over standard auscultation areas of the participants, including the mitral area (M), tricuspid area (T), pulmonic area (P), and aortic area (A), corresponding to the optimal locations for detecting sounds from MV, TV, PV, and AV, respectively (Fig. 5f). Figure 5g, Supplementary Fig. 29, and Movie S4 show that the heart sounds recorded from the M area exhibited a louder S1 than S2, while recordings from the P area showed the opposite pattern. Recordings from both locations using AusculPatch achieved an SNR above 12 dB, indicating sufficient quality for the potential recognition of murmurs and S3, S4 heart sounds51.
Figure 5h, i shows the HR recorded with AusculPatch from five participants, benchmarked against readings obtained from a commercial stethoscope (see Supplementary Fig. 30). The results demonstrate a strong agreement in HR measurement with a linear fit of r2 = 0.95. The mean difference and standard deviation of HR were −0.35 bpm and 4.10 bpm, respectively. Figure 5j presents AusculPatch data from a participant during 10 minutes of continuous recording, including HR, RR, IBI, and breathing interval. The data reveals temporal variation in cardiorespiratory signals of the subject, with irregular intervals falling outside the 95% confidence range. Cardiac signals measurements presented in Supplementary Figs. S31–S35 demonstrate signal stability in AusculPatch even under noisy conditions. Specifically, under 90 dB white noise and conversational noise conditions, AusculPatch preserves clear cardiac lub-dub sounds without significant artifact (Supplementary Figs. 31 and 35c). Quantitative analysis confirms that AusculPatch maintains a signal-to-noise ratio (SNR) exceeding 16 dB across these scenarios. The superior noise cancellation of the AusculPatch is attributed to two primary mechanical design features: (1) Acoustic propagation path with a high-aspect-ratio cavity beneath the cantilever (150 µm × 150 µm × 300 µm cavity) (Fig. 2f) that directs body-coupled signals rather than omnidirectional ambient sound, and (2) pressure-driven sensitivity that utilizes a 1-µm gap surrounding the cantilever to enhance responsivity at low-frequency physiological mechano-acoustic signals. In contrast, microphone-based devices exhibit pronounced susceptibility to external noise, with heart sound recordings severely distorted by white noise, thereby compromising the quality of physiological sound assessment (Supplementary Figs. 32–34). This demonstration represents a key advantage of our wearable platform for home-based monitoring, enabling the shift toward decentralized healthcare.
Supplementary Fig. 38 summarizes the measured cardiac sounds of a subject in daily activities. Throughout the day, the subject performed different physical activities: commuting, working in the office, having meals, moving, and doing exercises within a time frame of from 9 a.m. to 6 p.m. (Supplementary Movie S5). Overall, compared to sitting, moving and eating elevated the subject’s cardiac activities (heart rate and sound amplitude). Specifically, food and caffeine intake can increase the heart rate because of the activation of digestion and adrenaline, while walking and exercising accelerate the heartbeat to supply more oxygenated blood52,53. Detailed zoom-in cardiac sounds demonstrate that AusculPatch can reliably capture heart sounds even with light exercises such as climbing stairs and squatting, maintaining overall SNR above 8 dB.
Voice recognition and human–machine interfaces
Owing to its wide bandwidth and soft form factor for skin attachment, we further demonstrate the utility of AusculPatch as a wearable throat sensor for capturing vocal cord vibration. This feature holds promise for sleep quality assessment, speech recognition, communication support, and enhanced human–machine interfaces, particularly for patients with voice disorders due to laryngeal cancer, amyotrophic lateral sclerosis, or stroke, Fig. 6a54,55,56.
Fig. 6: Voice recognition and human–machine interaction.
The alternative text for this image may have been generated using AI.
a Block diagram of the remote-control system using voice recognition. b Plot figures showing signals recorded from the human throat consist of physiological signals (e.g., respiration and cardiac signals) and human vocal cord vibration, from top to bottom: time-domain raw signals, frequency-domain spectrogram of the raw signals, and the extracted vocal cord vibration from the collected data. c Block diagram of the voice recognition model and d a confusion matrix illustrating the training result. e A photo of the human–machine interaction, using AusculPatch and a voice recognition system to control a robot arm. f, g Remotely command a robot arm using AusculPatch to assist routine tasks, f writing/drawing, and g food-collecting.
Speech recognition based on vocal cord vibration
We collected a total of 1200 data samples of vocal cord vibration for training and testing by attaching AusculPatch onto the throat surface of participants and asking them to generate 20 commonly used words from five groups (food, number, machine command, shape, and color)57. The trachea point was selected due to its easily identified location and good sound signal quality, where vocal sounds can be separated from the physiological signals (i.e., cardiac signals, chest movement) by applying a 100 Hz high-pass filter, Fig. 6b.
Our machine learning (ML) model for voice recognition consists of three layers: manual feature extraction, automatic feature extraction, and classification, Fig. 6c. A 1D sound signal recorded from AusculPatch (i.e., the input data in the first layer) was extracted and converted into a three-channel 2D image. Here, the speech event was recognized by employing a high-pass filter and an envelope detector to identify the vocal sound area. Three features were extracted from the speech event, with two from the high-frequency regime revealing the information of the frequency domain and the vocal sound duration, and one from the low-frequency regime corresponding to the throat movement. We realized that the throat movement feature played an important role in training the model and improving the accuracy of the prediction, highlighting the importance of low-frequency detection. This is because the human throat skin serves as an acoustic filter, eliminating high-frequency features from the original vocal sound data, resulting in a similar perception between some words, such as “one” and “no” (Supplementary Figs. 39 and 40). The movement of the throat caused by the activity of the tongue and mouth during pronunciation is specific to each word, thereby supporting speech recognition. Our ML model, evaluated using the fivefold cross-validation technique, can recognize these 20 classes with an accuracy of 91.5% (using a combination of vocal sound and throat movement) (Fig. 6d), significantly improved in comparison to that trained with only the vocal cord vibration signal, with an accuracy of 88% (Supplementary Fig. 41).
Human–machine interaction using AusculPatch
With its capability to capture vocal vibration signals and convert them into speech patterns, AusculPatch represents a promising tool for human–machine interaction. To demonstrate this feature, we developed a voice-controlled system using vocal sounds detected using AusculPatch and implemented it into a UR10 robot arm (Universal Robot), Fig. 6e. The signal obtained from AusculPatch was transmitted wirelessly to UR10’s controlling computer, which was then converted into a specific command. The integration of AusculPatch, ML, and Robotics enables intuitive support for tasks involving writing, illustrating, or other routine activities. For example, in a writing task, when a user spoke the names of food items (e.g., “orange,” “banana,” and “apple”), the UR10 robot accurately recognized the word from AusculPatch’s acoustic signals and wrote it down on a whiteboard or alternatively generated a drawing of the detected object via an online search. The inference time of less than 300 ms allows these actions to be performed in real-time. To illustrate the system’s ability to assist with routine tasks, we further performed a food-collecting experiment. Different types of fruits were randomly placed on a table, and a user spoke the name of their preferred fruit. If the specified food was present on the table (detected via computer vision), UR10 was commanded to grab it and place it on a dish. The entire process of wireless control shows a good response of the robot arm to human command through our vocal sound recognition system using AusculPatch, Fig. 6g and Supplementary Movie S6.