nature biotechnology

Article https://doi.org/10.1038/s41587-023-01800-0

A fully integrated wearable ultrasound

system to monitor deep tissues
in moving subjects

Received: 13 July 2022 Muyang Lin 1,12, Ziyang Zhang2,12, Xiaoxiang Gao1,12, Yizhou Bian1, Ray S. Wu 1,
Geonho Park 1, Zhiyuan Lou1, Zhuorui Zhang3, Xiangchen Xu1, Xiangjun Chen4,
Accepted: 21 April 2023
Andrea Kang5, Xinyi Yang4, Wentong Yue1, Lu Yin 1, Chonghe Wang3,
Published online: 22 May 2023 Baiyan Qi4, Sai Zhou4, Hongjie Hu1, Hao Huang1, Mohan Li5, Yue Gu4,6, Jing Mu 4,
Albert Yang7, Amer Yaghi1, Yimu Chen1, Yusheng Lei1,8, Chengchangfeng Lu 5,
Check for updates
Ruotao Wang1, Joseph Wang1, Shu Xiang9, Erik B. Kistler7,10, Nuno Vasconcelos5
& Sheng Xu 1,4,5,7,11

Recent advances in wearable ultrasound technologies have demonstrated

the potential for hands-free data acquisition, but technical barriers
remain as these probes require wire connections, can lose track of moving
targets and create data-interpretation challenges. Here we report a fully
integrated autonomous wearable ultrasonic-system-on-patch (USoP).
A miniaturized flexible control circuit is designed to interface with an
ultrasound transducer array for signal pre-conditioning and wireless data
communication. Machine learning is used to track moving tissue targets
and assist the data interpretation. We demonstrate that the USoP allows
continuous tracking of physiological signals from tissues as deep as 164 mm.
On mobile subjects, the USoP can continuously monitor physiological
signals, including central blood pressure, heart rate and cardiac output, for
as long as 12 h. This result enables continuous autonomous surveillance of
deep tissue signals toward the internet-of-medical-things.

With decades of development in probe fabrication1,2, circuitry design3 tomography13 and magnetic resonance imaging14, ultrasonography
and algorithm optimization4,5, medical ultrasonography can qualita- is safer, less expensive and more versatile. However, the accessibility
tively and quantitatively acquire a broad range of physiological infor- and accuracy of ultrasonography face several technical challenges.
mation from the human body6,7, including anatomical structures8, First, common ultrasound probes are bulky and wired to large con-
tissue motion9, mechanical properties10 and hemodynamics11. Com- trol systems, which limits their usage to centralized facilities. Second,
pared with other medical imaging methods12, such as X-ray computed those probes need manual placement and maneuvering and require

Department of Nanoengineering, University of California San Diego, La Jolla, CA, USA. 2Department of Computer Science Engineering, University of
1

California San Diego, La Jolla, CA, USA. 3Department of Mechanical Engineering, Massachusetts Institute of Technology, Cambridge, MA, USA. 4Materials
Science and Engineering Program, University of California San Diego, La Jolla, CA, USA. 5Department of Electrical and Computer Engineering, University
of California San Diego, La Jolla, CA, USA. 6Department of Neurosurgery, Yale University, New Haven, CT, USA. 7Department of Bioengineering, University
of California San Diego, La Jolla, CA, USA. 8Department of Chemical Engineering, Stanford University, Stanford, CA, USA. 9Softsonics LLC, San Diego,
CA, USA. 10Department of Anesthesiology and Critical Care, University of California San Diego, La Jolla, CA, USA. 11Department of Radiology, School of
Medicine, University of California San Diego, La Jolla, CA, USA. 12These authors contributed equally: Muyang Lin, Ziyang Zhang, Xiaoxiang Gao.
e-mail: [email protected]

Nature Biotechnology | Volume 42 | March 2024 | 448–457 448

Article https://doi.org/10.1038/s41587-023-01800-0

the subjects to remain motionless, introducing operator dependency. achieve a depth of ~78 mm with an axial resolution of ~330 μm for
Third, the interpretation of sonographic data requires medical profes- targeting major arteries (for example, aorta, carotid and femoral
sionals with specialized training and is labor-intensive and error-prone. arteries). The 6-MHz transducers achieve a depth of ~9 mm and an
Recent advances in point-of-care ultrasound systems15 have sub- axial resolution of ~230 μm for targeting smaller peripheral arteries
stantially reduced the device size (Supplementary Fig. 1 and Supple- (for example, radial and brachial arteries) (Extended Data Fig. 1). To
mentary Table 1). However, they either need manual operations3 or achieve the desired beam profiles, we customize three probe layouts:
require bulky rigid circuits16, because ultrasound hardware typically disc, linear array and two-dimensional array, for penetrative, wide and
requires high power and high bandwidth. The use of bulky rigid probes narrow beam, respectively (Supplementary Fig. 6 and Supplementary
and circuits creates difficulties to cover a large area and conform to Discussion 1). For electrical connection, we use anisotropic conductive
highly curved body surfaces. Emerging wearable ultrasonic probes films (ACFs) with easy attachment and detachment for repetitive use
leveraging soft structural designs can naturally conform to the skin (Supplementary Fig. 7).
and acquire deep tissue signals in a hands-free manner17–19. Alterna- The control electronics are designed as a flexible printed circuit
tively, integrating rigid ultrasound chips with soft adhesive materials board (fPCB) (Supplementary Fig. 8 and Supplementary Table 2)
can achieve a reliable interface on the human skin20. However, these for ultrasonic sensing and wireless communication. The circuitry
wearable probes all require cumbersome cables for power and data consists of an analog front-end (AFE) and a data acquisition module
transmission, which substantially limits the subjects’ mobility, mak- (DAQ module) (Fig. 1b and Supplementary Fig. 9). The AFE achieves
ing surveillance challenging during dynamic tests or daily activities. ultrasonic sensing through coordinated sequence control of multiple
Developing a fully integrated ultrasonic probe with soft front-end components (Extended Data Fig. 2 and Supplementary Discussion 2).
circuits has yet to be demonstrated21,22. Additionally, current wear- First, the sequencer initiates sensing by sending trigger signals to
able ultrasound technologies can lose track of a target tissue dur- the pulse generator and multiplexer. Then, the pulse generator reads
ing subject motion, because the device on the skin surface shifts its the trigger signals and outputs high-voltage impulses to activate the
position relative to deep tissues. Thus, they require frequent manual ultrasound transducers. Meanwhile, the multiplexer drives the arrayed
repositioning and only allow point-in-time examinations3,23. Moreo- transducers to generate ultrasound and receive echoes. Finally, the
ver, with the large amount of data generated from continuous surveil- echoes are collected by the transmit/receive switch, and then ampli-
lance, the front-end circuits and back-end processing units would be fied and filtered by the receiver circuit. After the AFE completes the
overwhelmed. Therefore, a critical milestone in the development of ultrasonic sensing process, the analog echoes are relayed to the DAQ
wearable ultrasound technology is to realize a fully integrated wireless module. The microcontroller unit samples the echoes with a built-in
system that can track a moving target and automate data acquisition analog-to-digital converter, and then the Wi-Fi module wirelessly
and processing. transmits the digitalized echoes to a terminal device (for example, a
Here, we report a fully integrated autonomous ultrasonic- computer or a smartphone), where an online machine learning algo-
system-on-patch (USoP). The USoP integrates the ultrasonic probe and rithm and an application program can process and display the signals
miniaturized wireless control electronics in a soft, wearable format, autonomously (Fig. 1c).
which overcomes the above-mentioned limitations. Multiple channels The AFE and the DAQ modules are interconnected by serpentine
of deep tissue signals acquired from the subject are conditioned and wires that allow for folding to minimize their footprint (Supplemen-
preprocessed on-board, and then wirelessly transferred to a back-end tary Fig. 10). An elastomeric encapsulation mitigates strain concen-
receiver, where they are analyzed by a customized machine learning trations and protects the circuit from irreversible deformations
algorithm. When the USoP on the skin moves relative to the target tis- (Supplementary Figs. 11 and 12 and Methods). The fully integrated
sue, the algorithm classifies the data and selects the best channel in real system can be bent, stretched and twisted (Extended Data Fig. 3)
time, yielding a continuous data stream from the target tissue. There- and can be conformally laminated on the human body (Fig. 1a and
fore, this technology allows continuous monitoring of deep tissue Extended Data Fig. 4).
signals during human motion. The fully integrated autonomous USoP The ultrasonic probes have MHz-level bandwidth, substantially
eliminates the operator dependency of conventional ultrasonography, higher than other common sensors25 (Supplementary Fig. 13). There-
standardizes the data-interpretation process and therefore expands fore, achieving high sensing bandwidths and sampling rates is critical
the accessibility of this powerful diagnostic tool in both inpatient and for the circuitry design. In this work, the DAQ module samples the signal
outpatient settings. 12 million times per second corresponding to a sensing bandwidth
of 6 MHz. The Wi-Fi module can transmit such wide-band signals at
Results a distance of ~10 m and a speed of 3.4 Mbps with zero data loss (Sup-
Design of the USoP plementary Fig. 14)26. The USoP system has a power consumption of
The USoP hardware consists of an ultrasound probe and control elec- ~614 mW. A standard 3.7-V commercial lithium-polymer battery can
tronics which are fabricated in a miniaturized, soft format (Fig. 1a). enable continuous operation for up to 12 h (Supplementary Fig. 15).
The ultrasonic probe is made of piezoelectric transducers, backing The USoP can perform tissue sensing in multiple modalities,
materials, serpentine interconnects and contact pads, similar to our including amplitude mode (A-mode), brightness mode (B-mode) and
reported structures18,19,24. This soft probe design reduced noise cou- motion mode (M-mode), to reveal the tissue structures and interface
pling, enhanced resolution, enabled gel-free acoustic coupling and movements27 (Supplementary Fig. 16, Supplementary Discussion 3,
ensured probe durability (Supplementary Figs. 2–5 and Supplemen- Fig. 1d and Supplementary Video 1). We characterized the elevational
tary Discussion 1). We design the probes with center frequencies from and lateral resolutions of these sensing modalities. In A-mode and
2 MHz to 6 MHz to achieve the desired bandwidth, axial resolution M-mode, the elevational and lateral resolutions show a degrading trend
and penetration depth. We determine the bandwidth as the −3-dB when the sensing depth increases (Supplementary Fig. 17 and Supple-
frequency band of the pulse-echo response, axial resolution as the full mentary Discussion 3). In B-mode, the elevational resolution can be
width at half maximum of the pulse-echo response and penetration defined by the transmission beam pattern, while the lateral resolution
depth as the −3-dB attenuation point in tissues. All soft probes can can be determined directly from image reconstruction (Supplementary
achieve a relative bandwidth of ~50%, which is similar to a commercial Fig. 16 and Supplementary Discussion 3). When the probes conform to
probe (Extended Data Fig. 1). The 2-MHz transducers achieve a depth skin surfaces within certain bending radius thresholds, the soft probes
of ~164 mm with an axial resolution of ~600 μm for targeting visceral offer stability in sensing. For A-mode and M-mode, the resolutions can
organs (for example, heart and diaphragm). The 4-MHz transducers be maintained with an array bending radius >6 mm (Supplementary

Nature Biotechnology | Volume 42 | March 2024 | 448–457 449

Article https://doi.org/10.1038/s41587-023-01800-0

a b
Serpentine interconnects
ACF
la
st

E
om
er p
a c ka gin g
C
AD
U/
MC
Piezoelectric
1 cm
transducer
array
r
ive
ce
Battery Re
Q Ba
Flexible M ux DA
tte
ry
circuit T/R W
S

Stretchable
probe

on
cti
Soft n ne
package co
AF
E ble Autonomous wearable
AC lda ultrasound sensing
2 cm Fc Fo
on Tissue motion image
ne
cto 0 mm
c Analog Analog Digital
d r
AFE DAQ -Fi 2 mm
impulse echo signal
Se Wi
qu 3 mm
Mux Sequencer Battery CA en
ce
Arrayed transducer

r
Pulse generator JV Anterior Pu
MCU lse
Tx ctrl wall ge
ne
T/R Booster rat HR 61 BPM
ADC or
SW
Receiver DBP 68 mmHg
Posterior Cloud
…
…

SBP 123 mmHg

Filter Wi-Fi wall
32 ch 1 cm
VGA 5 mm 2s

Fig. 1 | Overview of the fully integrated USoP. a, A photograph of the battery. A smartphone application is designed to display the data stream from
encapsulated USoP laminated on the chest for measuring cardiac activity via the USoP. From the ultrasonic data, M-mode images and physiological signals
the parasternal window. The inset shows a folded circuit. b, Design of the USoP, can be derived and displayed in real time. The smartphone can also communicate
including a stretchable ultrasonic probe, a flexible control circuit and a battery. with a cloud server for further data analysis (lower right). c, Block diagram of
The ultrasonic probe consists of a piezoelectric transducer array, serpentine the USoP showing the flow of analog impulse, analog echo and digital signals.
interconnects and an ACF (upper left). The exploded view of the circuit shows The AFE performs pulse-echo sensing to acquire ultrasonic signals, and the DAQ
two parts: (1) an AFE, including a 32-channel (ch) multiplexer (Mux), a transmit/ module samples signals and wirelessly transmits the data to a terminal device for
receive switch (T/R SW), a receiver with a variable gain amplifier (VGA) and a processing and display. d, B-mode imaging of the carotid artery (CA) and jugular
filter, a pulse generator with a transmit controller (Tx ctrl) and a booster, and a vein ( JV), while the subject is performing the Valsalva maneuver to dilate the JV
sequencer; and (2) a DAQ module including a microcontroller unit (MCU) with a (left). M-mode imaging of the pulsation pattern of CA walls (right). HR, heart
built-in analog-to-digital converter (ADC) and a Wi-Fi chip. The two modules are rate; BPM, beats per minute; DBP, diastolic blood pressure; SBP, systolic blood
connected by serpentine electrodes, which allow the entire circuit to be folded pressure.
for a smaller footprint. The circuit is powered by a commercial lithium-polymer

Fig. 18 and Supplementary Discussion 4). For B-mode, the image arti- In arterial pulse waveforms, the pulse interval reflects the heart
facts could be neglected when the bending radius of the array is >6 cm rate, and the pulse intensity can be correlated to blood pressure (Fig. 2c
(Supplementary Fig. 19 and Supplementary Discussion 4). and Supplementary Discussion 6)18. We validated the USoP results
against a clinical-grade tonometer, the noninvasive gold standard for
Physiological signal recording and validation pulse waveform recording32 (Methods). Bland–Altman analysis was
In clinical practice, A-mode and B-mode are commonly used for performed to compare the waveform-derived heart rate and blood
temporary measurements, while M-mode is for monitoring signals pressure from both devices (Fig. 2d). The 95% limits of agreement
continuously27. Additionally, M-mode is valuable for quantitatively included >95% of differences between the results from the tonometer
characterizing tissue dynamics28–30. Therefore, in this work, we focus and USoP, showing measurement consistency between these two
on the use of the USoP in M-mode. Natural physiological processes, devices. Additionally, the time difference between myocardial con-
such as circulation and respiration, can be manifested in the motion traction and arterial pulsations can be used to quantify the pulse wave
of tissue interfaces, such as myocardial contraction, arterial pulsation velocity, which correlates to the arterial stiffness of specific arterial
and diaphragmatic excursion. The USoP can quantify these interfacial segments (Extended Data Fig. 5 and Supplementary Discussion 7).
motions from multiple sensing windows in the human body (Fig. 2a, Comparing the results of the USoP with those of the tonometer sug-
Supplementary Figs. 20 and 21, Supplementary Discussion 5 and gests a mean pulse transit time difference of <0.5 ms, which results in
Supplementary Table 3). <4% error in pulse wave velocity recording, further demonstrating the
From myocardial contraction, the diameter change of the left ven- accuracy of the USoP (Extended Data Fig. 5)33.
tricle during cardiac cycles can be recorded, and therefore fractional The USoP can also measure diaphragmatic excursion as a surro-
shortening can be derived as a measure of left ventricular function gate for changes in respiratory volume. The diaphragm depth recorded
(Fig. 2b, left)31. A comparison of measurements from the USoP and a by the USoP is compared with the respiratory volume recorded by
commercial ultrasonic system shows a mean difference of ~1% (Fig. 2b, a spirometer (Fig. 2e, left, and Methods). With a linear regression
right, Supplementary Fig. 22 and Methods). model, the correlation coefficients between the diaphragmatic depth

Nature Biotechnology | Volume 42 | March 2024 | 448–457 450

Article https://doi.org/10.1038/s41587-023-01800-0

a Time (s)
0 1 2 3 4 5 6 7 8 9 10 11 12
Depth (mm)
0 Radial artery
Brachial artery 1.9
Carotid artery

20 Femoral artery 2.0

Carotid
artery
4.1
40

4.2

Abdominal aorta
Left ventricle 60
14
wall

Depth (mm) 15
80
Diaphragm
Brachial Left ventricle wall 17
artery 100
18
Abdominal Radial
aorta artery
120 60
Femoral
artery
61
140

62
Diaphragm
160 Time (s)

b d 1
Fractional shortening 35
Heart rate difference (bpm)

10
Fractional shortening (%)

= 31.44% 10 1.96 s.d.

30

difference (mmHg)
1.96 s.d. 8.59
30

Blood pressure
25 5
Depth (mm)

Septum 0.38
20 Mean Mean
50 LVIDs 0 0
LVIDd 0.013 0.17
15
–1.96 s.d.
70 –5
10 –0.36 –1.96 s.d.
5 –8.24
90 –10 Diastolic Systolic
Left ventricle wall
0 –1
0.5 1.0 1.5 2.0 2.5 USoP Commercial 40 60 80 100 120 140 160 40 60 80 100 120
Time (s) Heart rate (bpm) Blood pressure (mmHg)

c e 18
1.5
Systolic blood
Diaphragm excursion (cm)

Heart 3.0
pressure
Expiratory volume (l)

rate
16 2.5 Forced breathing
1.0 V = 0.47D–5.00
Volume (l)

2.0
CC = 99.9%
14 1.5
0.5

Heartbeat 1.0
Normal breathing
interval Blood 12 V = 0.28D–2.78
Diastolic blood 0 0.5
pressure CC = 99.7%
pressure
0
0 2 4 6 8 10 12 10 11 12 13 14 15 16 17 18
Time (s) Diaphragm excursion (cm)

Fig. 2 | Monitoring and analysis of tissue interface motions using the USoP. (bpm) is observed, and 135 of 142 (95.1%) data points are within 95% limits of
a, Schematics and measurement results of seven representative dynamic tissue agreement defined by ±1.96 s.d. Right: for the blood pressure, a mean difference
interfaces. b, Deriving physiological parameters from myocardial contraction. of 0.17 mmHg is observed, and 269 of 280 (96.1%) data points are within 95%
From the M-mode waveforms of the septum and left ventricular wall, the LVIDd limits of agreement defined by ±1.96 s.d. e, Derivation of expiratory volume from
and LVIDs can be used to derive the fractional shortening (left). Comparison of the diaphragmatic excursion. Simultaneous measurements of diaphragmatic
measurements between the USoP and a commercial ultrasound probe (right). excursion and respiratory volume show a similar pattern (left). The regression is
The results are averaged from 10 independent measurements, and the error on expiratory volume (V) with diaphragmatic depth (D) in normal breathing and
bars represent the standard deviation. c, Derivation of physiological parameters forced breathing. Strong linear relationships, with correlation coefficients (CCs)
from the arterial pulse waveforms, including the heart rate and blood pressure. close to 100%, can be found between the diaphragmatic excursion and expiratory
d, Bland–Altman plot showing measurement agreement between the USoP and volume in both breathing conditions (right).
a tonometer. Left: for the heart rate, a mean difference of 0.013 beats per minute

and respiratory volume under normal and forced breathing condi- performance and identify airway obstruction or lung capacity restric-
tions are 99.9% and 99.7%, respectively (Fig. 2e, right). Furthermore, tion (Supplementary Fig. 23, Supplementary Discussion 8 and Sup-
these derived volumes can be used to characterize the respiratory plementary Table 4), which can potentially be used for screening

Nature Biotechnology | Volume 42 | March 2024 | 448–457 451

Article https://doi.org/10.1038/s41587-023-01800-0

respiratory issues such as chronic obstructive pulmonary disease34 The results show that only 32 unlabeled images from a new subject
and pulmonary fibrosis35. suffice to achieve >90% classification accuracy (Supplementary Fig. 33
and Supplementary Discussion 13).
Autonomous data acquisition and analysis by machine
learning Continuous monitoring during exercise
We use the USoP with a 4-MHz 32-channel linear array probe to auton- The USoP can continuously track multiple deep tissue signals during
omously and continuously track the position of the carotid artery human motion. To test its performance, we used it on a participant
and sense its pulsations. The linear array has an acoustic aperture of during aerobic exercise, when the participant performed 30 min con-
~25.4 mm, which is sufficiently wide to accommodate the misalignment tinuous cycling followed by 30 min rest. We record the carotid blood
between the probe and the carotid artery36. Pulsation is visible in the pressure waveform while the participant moves freely (Fig. 4a and
M-mode images derived from the transducer channels directly above Supplementary Video 3). Similar measurements were also made dur-
the carotid artery, while the M-mode images from the other adjacent ing anaerobic exercise, when the participant performed high-intensity
channels show weaker or no pulsations (Fig. 3a). We train machine learn- interval training (HIIT) comprising six 1-min training sessions, sepa-
ing models to classify those M-mode images and identify whether sali- rated by six 1-min periods of resting (Extended Data Fig. 7).
ent pulsation patterns are present in the image (Supplementary Fig. 24). Upon the onset of exercising, the substantial increase in the blood
Specifically, we use a VGG13 model because it outperforms other com- pressure and heart rate suggests a boost in circulating blood, also
monly used models for medical image classification in terms of preci- known as the stressed volume (Fig. 4b,c)38. During both cycling and
sion, recall and accuracy. This model can even handle compromised HIIT, the systolic pressure increases more than the diastolic pressure,
ultrasound images and maintain the precision, recall and accuracy regulated by increased cardiac output and decreased vascular resist-
higher than 98.4% (Extended Data Fig. 6, Supplementary Discussion 9 ance (Supplementary Discussion 14). The heart rate increases mono-
and Methods), which is more robust than conventional logistic models tonically during both types of exercise and decreases during resting,
(Supplementary Fig. 25 and Supplementary Discussion 9). Based on as anticipated39. As cycling progresses, the blood pressure gradually
the arterial wall patterns in the M-mode images, this model predicts stabilizes at a relatively elevated level, resulting in narrow distributions
probability scores for each of the 32 channels and, therefore, gener- of both systolic and diastolic pressures in the histogram (Fig. 4d, top).
ates a probability profile of the position of the artery (Supplementary These results imply that the systemic vascular resistance decreases to a
Discussion 10). The channel with the highest probability is determined physiologically determined steady state to support prolonged muscle
as the center of the artery (Supplementary Fig. 26), and its channel data activity40. This is in stark contrast to HIIT, during which blood pressure
are used for generating the pulse waveforms (Fig. 3b). fluctuates, resulting in wider distributions of both diastolic and systolic
We record human head motion using inertia measurement units blood pressures (Fig. 4d, bottom). In both cycling and HIIT, resting
(Supplementary Fig. 27 and Methods) and simultaneously image the allows blood pressure to gradually decrease toward the baseline.
carotid artery to quantify its displacement. The head can yaw at a larger We derive the vascular responses to exercise by calculating the
angle than it can roll and pitch, and yawing generates the largest arterial augmentation index (AIx)41,42 (Supplementary Figs. 34 and 35 and Sup-
displacement (~19 mm) (Supplementary Fig. 28). The USoP generates plementary Discussion 15). In both cycling and HIIT, the AIx increases
M-mode images from all channels with head yawing. The VGG13 model with exercise and recovers with resting; when the exercise is sufficiently
identifies the M-mode images containing arterial pulsations, deter- long, as in the case of cycling, the AIx stabilizes (Fig. 4e). The increase
mines a moving sub-aperture to follow the carotid artery (Supplemen- in the AIx during exercise may have two causes: vessel stiffening43 and
tary Fig. 29), selects the optimal channel from the probability profile vasodilation42,44. We measure the change in the arterial stiffness index
(Supplementary Video 2) and generates continuous pulse waveforms before, during and after exercise (Supplementary Fig. 36 and Methods).
autonomously (Fig. 3c). In contrast, without the model, a fixed chan- The results suggest a negligible change (<0.34%) in the stiffness index45.
nel with head yawing loses track of the pulsation waveform once the Additionally, such a negligible change in the stiffness index leads to
artery is outside its sensing aperture (Fig. 3c). The model prediction a central blood pressure error <1.58 mmHg after calibration, which
remains reliable at a head yawing rate <60° s−1 (Supplementary Fig. 30 proves the reliability of the blood pressure recordings during exercise
and Supplementary Discussion 11). At yawing rates beyond this limit, (Supplementary Fig. 36 and Supplementary Discussion 16). Therefore,
the pulse waveform becomes distorted but is quickly restored when the increase in the AIx is primarily driven by vasodilation rather than
motion stops (Supplementary Fig. 31). changes in arterial stiffness. The vasodilation takes place mainly in the
Machine learning algorithms may encounter generalization prob- skeletal muscle involved in the exercise to support an elevated demand
lems when tested on images outside the training pool. For example, for oxygen and thus blood flow42,46; activating larger muscle groups
images from a new subject may have distinct brightness, contrast results in greater vasodilation and increased blood flow, and thus a
and arterial wall patterns, which would result in different luminosity higher AIx (Supplementary Fig. 37).
distributions (Fig. 3d). We enhanced the generalization of the VGG13 We estimate the stroke volume from the pressure waveforms using
model by using domain adaptation with a minimal entropy correla- a pulse contour method (Supplementary Fig. 38 and Supplementary
tion alignment model37 (Fig. 3e and Supplementary Discussion 12) to Discussion 17)47. The cardiac output is then calculated as the product of
transfer the machine learning network to new image datasets without stroke volume and heart rate. Similar patterns in the stroke volume and
additional labeling. The use of domain adaptation allows the model to heart rate are observed in both cycling and HIIT (Fig. 4f). The measured
generalize to different subjects. A t-distributed stochastic neighbor cardiac output increases as the exercise intensifies, and the heart rate
embedding visualization of the subject distributions shows that images increases together with the cardiac output. Initially, the stroke vol-
from different subjects are unified after domain adaptation is applied ume increases before plateauing as end-systolic volume approaches
(Supplementary Fig. 32 and Methods). Model generalizability is demon- the mechanical limits of the heart48 and the increase of end-diastolic
strated through cross-validation among ten subjects (Supplementary volume begins to be limited by the shorter filling times at higher heart
Table 5). We train the classification model on each subject and then rates49. In the high cardiac output region (for example, >15 l min−1), the
validate it on the nine other subjects. Without domain adaptation, stroke volume plateaus, and the increase in cardiac output is mainly
the model only has an average accuracy of 63.23% on new subjects attributed to the increase in heart rate50. Compared with cycling, HIIT
(Fig. 3f, left). After domain adaptation, this accuracy increases to produces a greater increase in stroke volume and a higher maximum
96.13% (Fig. 3f, right). We also investigate the minimum data required cardiac output, indicating that HIIT may be a more effective training
to be collected from a new subject for successful domain adaptation. modality for enhancing cardiac functions51,52.

Nature Biotechnology | Volume 42 | March 2024 | 448–457 452

Article https://doi.org/10.1038/s41587-023-01800-0

a b
32 channel images

y Classification
r arra
Linea
CA nCA
channels channels
(14–20) (1–13, 21–32)

Score
CA probability
Artery profile 0
0 Channel 32

Pulse wave
generation
Tissue

nCA image CA images nCA image

c Channel number Channel number Channel number Channel number Channel number
1 32 1 32 1 32 1 32 1 32
1
Score
0
–80° –40° 0° 40° 80°

–80° Yawing –40° Yawing 0° Yawing 40° Yawing 80°

360
Carotid pulse

280
(µm)

200
120
40 Ch no. 5 Ch no. 8 Ch no. 16 Ch no. 23 Ch no. 29
0 2 4 6 8 10 12 14 16 18

360
Carotid pulse

280
(µm)

200
120
40 Ch no. 16
0 2 4 6 8 10 12 14 16 18
Time (s)

d e f
Training images New images Training subject Training subject
No. 1 Bright pixels (training) (new) 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
10
5 100%
1 1
10
5
36k
4 2 2
10 Conv. Conv.
6k 3 3
Validation subject

Validation subject

layer layer
Pixel number

10
3 4 4
10
4
50%
5 5
No. 2 2
10 6 6
240 250
FC layer FC layer
3 7 7
10
8 8
Geodesic
Classification 9 9
covariance
loss
10
2 distance loss 10 10
0 50 100 150 200 250
Luminosity

Fig. 3 | Autonomous and continuous blood pressure recording in a moving from the training subject (no. 1) and a new subject (no. 2), showing different
subject. a, Left: schematic cross-sectional view of a soft 4-MHz linear array image patterns (left). The histograms of the two CA images show a substantial
sensing the carotid artery. Right: representative M-mode images of difference in luminosity distribution (right). Inset, subject no. 2 has ~6 times
channels with beam penetrating or not penetrating the carotid artery, more white pixels than subject no. 1, indicating thicker arterial walls.
classified as carotid artery (CA) or noncarotid artery (nCA) images, respectively. e, Schematic diagram showing the workflow of the minimal entropy correlation
b, Flow diagram showing the process of autonomous CA detection and pulse alignment model, consisting of two encoders with five convolutional (Conv.)
waveform generation. c, Recording in a moving subject using the USoP with layers and three fully connected (FC) layers. The classification loss and geodesic
and without an autonomous algorithm. The algorithm can reliably track the CA covariance distance loss are used to align features extracted from the training
position with head yawing from −80° to +80°, corresponding to a ~19-mm CA image set (source domain) and those from a new image set (target domain).
displacement. Prediction scores of different transducer channels for tracking f, Model generalizability validation on 10 subjects. The classification model is
the CA at each yawing position and corresponding B-mode images collected by trained on each subject and validated on the remaining subjects. Without domain
a commercial ultrasound machine (top). By actively selecting the best channel adaptation, the matrix plot shows an average classification accuracy of only
to follow the CA motion (for example, no. 5, no. 8, no. 16, no. 23 and no. 29), 63.23% on new subjects (left). After domain adaptation, the average classification
continuous pulse waveforms can be recorded (middle). In contrast, without the accuracy is boosted to 96.13%, showing the improved generalization of the
auto-selection algorithm, a fixed channel (for example, no. 16) results in signal classification model (right).
loss during motion (bottom). d, Two representative M-mode images recorded

Nature Biotechnology | Volume 42 | March 2024 | 448–457 453

Article https://doi.org/10.1038/s41587-023-01800-0

a
i. Forward ii. Turned iii. Bowed iv. Raised

Cuff
Smartphone

b Training Rest Roll Yaw Pitch Time (min)

0 10 20 30 40 50 60
i ii iii iv
Rotation rate
(50° s–1)

140
Blood pressure

120
(mmHg)

100

80

60

140

120
Heart rate
(bpm)

100

80

60

c 50
Roll Yaw Pitch
(mmHg) rate (° s–1)

i ii iii iv
pressure Rotation

–50 Cuff: 150/66 mmHg

150 Cuff: 148/65 mmHg Cuff: 148/65 mmHg
Cuff: 128/61 mmHg
120
Blood

90
60
120 Cuff: 113 bpm
Cuff: 105 bpm
Heart rate

Cuff: 90 bpm
(bpm)

90
Cuff: 68 bpm 2s
60

d e 55 f
20 mmHg Cycling Cycling 140 Cycling
16% 150
50
47 mmHg
45 120 135
12%
40 100 120
8% 35 105
Stroke volume (ml)
Heart rate (bpm)

80
Value percentile

4% 30 90
15.6% higher
25 10 min 60
AIx (%)

75
0% 55
HIIT HIIT 140 HIIT
16% 150
50 i ii iii iv v vi
45 120 135
12%
38 mmHg 40 100 120
8% 55 mmHg 35 105
80
30 90
4%
25 2 min 60
75
0%
40 80 120 160 Time 4 6 8 10 12 14 16 18 20
Blood pressure (mmHg) Cardiac output (l min–1)

Discussion it is critical to monitor deep tissue signals directly. More importantly,

While most existing wearable devices capture signals on or near the deep tissue physiology is constantly changing. To identify potential
skin surface53–56, such signals are often manifestations of physiological risk factors for a disease, capture its early onset, or evaluate its pro-
processes in deep tissues25. Therefore, in many clinical applications, gression, obtaining longitudinal data over the course of days, weeks

Nature Biotechnology | Volume 42 | March 2024 | 448–457 454

Article https://doi.org/10.1038/s41587-023-01800-0

Fig. 4 | Continuous monitoring during exercise. a, Photographs showing a reflections. d, Histograms of the diastolic and systolic pressures during cycling
subject cycling while the carotid pulsation waveform is measured by the USoP and HIIT. During cycling, the variations in diastolic and systolic pressures are
with different head positions, including (i) forward, (ii) turned, (iii) bowed and 20 mmHg and 47 mmHg, respectively. During HIIT, the variations in diastolic
(iv) raised. The USoP can transmit data to the cloud server for processing and the and systolic pressures are 38 mmHg and 55 mmHg, respectively. e, Changes
smartphone mounted on the bicycle displays the results. Inertia measurement in augmentation indices during cycling and HIIT. The AIx first increases and
units are used to record the head motion. An automated cuff on the upper then plateaus during cycling, and then recovers during resting (top). The AIx
arm acquires brachial blood pressure levels for reference. b, Head motions fluctuates during HIIT, coinciding with the training–rest cycles (bottom).
recorded by the inertia measurement units during cycling. The carotid blood Notably, the AIx was substantially higher during (ii) and (iv) training sessions,
pressure waveforms and heart rate are recorded simultaneously using the USoP. indicating greater arterial vasodilation. Average augmentation indices are
The maximum increases in diastolic and systolic pressures are 17 mmHg and calculated from 50 independent pulse waveforms every minute. The error
45 mmHg, respectively. c, Enlarged view of the head motion, blood pressure bars represent the standard deviations of the recorded augmentation indices.
waveforms and heart rate recorded during the (i)–(iv) motion periods in b. The f, Cardiac response to cycling and HIIT. In both exercise scenarios, the stroke
carotid diastolic pressures measured by the USoP agree well with the brachial volume first increases and then plateaus while the heart rate continues to
pressures measured by the cuff. The carotid systolic pressures measured by increase. Cycling has a smaller increase in stroke volume than HIIT. The maximum
the USoP are ~10 mmHg lower than the cuff brachial values, due to lower distal cardiac output measured during HIIT is 15.6% greater than during cycling.

or even months is key. This calls for a tool that enables long-term deep unknown transducer locations when conformed to dynamic and curvi-
tissue surveillance, processes the data stream in real time and remains linear skin surfaces. A-mode and M-mode using single transducers with-
accurate during human motion. out beamforming are not affected, but unknown transducer locations
Medical ultrasound is one of the most widely used methods for cause phase aberration and compromised beamforming for B-mode
deep tissue sensing, but due to the complex equipment and the require- imaging. Potential solutions include applying additional shape sensors
ment for an operator, traditional ultrasound exams offer point-in-time to map the transducer locations in real time62, or developing iterative
measurements only. In fact, a critical barrier that prevents traditional contrast optimization algorithms to compensate the phase distortion
ultrasound from long-term use is its operator dependency57,58. Even with of a deformed array63. Second, the long-term wearability of the USoP
standardized exam procedures, results reported using conventional should be further improved. Incorporating highly integrated chips
ultrasonography strongly depend on operator skill. When mishandled, with multilayered soft circuitry64 could further enhance the mechani-
it may generate compromised or even erroneous results (Supplemen- cal compliance of the system. Combining wearable power-harvesting
tary Fig. 39 and Supplementary Discussion 18). devices65 could extend the battery life of the USoP. Replacing silicone
Recent advances in wearable ultrasonography have shown the adhesives with more durable and permeable adhesives could help
promise of capturing deep tissue signals over the long term. Soft, enhance skin integration under skin deformation and perspiration66.
wearable ultrasonic probes17–19,24, as well as rigid ultrasound chips inte- Third, the USoP can potentially be applied to other tissue targets,
grated with soft adhesives20, have demonstrated hands-free ultrasound particularly in high-risk populations where continuous monitoring
signal acquisition. However, removing the requirement to handhold is critical (Supplementary Discussion 20). Fourth, the cloud com-
the probe is only the initial step toward continuous operation, and puting resources necessary for machine learning processing limit
three further technical barriers remain. First, these probes have to be the accessibility in remote and undeveloped regions. On-board data
wired to a central processing station, which largely limits the wearing analytics based on power–performance balance optimization and arti-
subject’s mobility. Second, existing wearable ultrasound devices face ficial intelligence-on-a-chip technology may be a possibility67. Finally,
challenges with measurement continuity and reliability in moving through strategically tuning the ultrasound controlling parameters
subjects, because the device on the skin shifts in position relative to such as activation frequency and pulse profile, this technology could
the target tissue. Third, wearables generate new challenges for manual enable more intriguing wearable diagnostic and therapeutic applica-
data processing because any clinicians will be overwhelmed by the tions, including anatomic imaging20,24, functional imaging19,68,69 and
continuous data stream. ultrasound stimulation70.
The fully integrated USoP addresses these three barriers and
makes continuous surveillance of deep tissue signals possible. First, Online content
the USoP eliminates wire connections by connecting the device and the Any methods, additional references, Nature Portfolio reporting sum-
back-end processing system wirelessly, which allows for large-range maries, source data, extended data, supplementary information,
subject mobility. Second, the USoP uses machine learning-based algo- acknowledgements, peer review information; details of author contri-
rithms to automate the data acquisition and channel selection in real butions and competing interests; and statements of data and code avail-
time. To our knowledge, no previously reported wearable device can ability are available at https://doi.org/10.1038/s41587-023-01800-0.
autonomously track a moving target. Third, deep learning-enabled
data post-processing relieves the human burden and enables potential References
scale-up. Together, these innovations open up many new possibilities. 1. Savoia, A. S., Caliano, G. & Pappalardo, M. A CMUT probe for
For example, patients can be monitored as they conduct their natu- medical ultrasonography: from microfabrication to system
ral daily activities, which can provide rich information that is more integration. IEEE Trans. Ultrason. Ferroelectr. Freq. Control 59,
clinically relevant59. Responses to high-risk activities such as during 1127–1138 (2012).
an intense workout can be captured for more rigorous diagnostics60,61. 2. Shung, K. K., Cannata, J. & Zhou, Q. Piezoelectric materials
Continuous monitoring over days or weeks of the dynamic changes of for high frequency medical imaging applications: a review.
the cardiovascular system in response to stressors can benefit a broad J. Electroceram. 19, 141–147 (2007).
range of populations, from athletes who need training optimization, 3. Rothberg, J. M. et al. Ultrasound-on-chip platform for medical
to cardiac rehabilitation patients who require safety measures, and to imaging, analysis, and collective intelligence. Proc. Natl Acad. Sci.
general high-risk populations for cardiovascular risk stratification and USA 118, e2019339118 (2021).
prediction (Supplementary Discussion 19). 4. Brattain, L. J., Telfer, B. A., Dhyani, M., Grajo, J. R. & Samir, A. E.
Future developments of this technology can be focused on the Machine learning for medical ultrasound: status, methods, and
following areas. First, the soft ultrasonic probes face challenges of future opportunities. Abdom. Radiol. 43, 786–799 (2018).

Nature Biotechnology | Volume 42 | March 2024 | 448–457 455

Article https://doi.org/10.1038/s41587-023-01800-0

5. Liu, S. et al. Deep learning in medical ultrasound analysis: a 29. Gamble, G., Zorn, J., Sanders, G., MacMahon, S. & Sharpe, N.
review. Engineering 5, 261–275 (2019). Estimation of arterial stiffness, compliance, and distensibility from
6. Powers, J. & Kremkau, F. Medical ultrasound systems. Interface M-mode ultrasound measurements of the common carotid artery.
Focus 1, 477–489 (2011). Stroke 25, 11–16 (1994).
7. Moran, C. M. & Thomson, A. J. Preclinical ultrasound imaging—a 30. Testa, A. et al. Ultrasound M-mode assessment of diaphragmatic
review of techniques and imaging applications. Front. Phys. 8, kinetics by anterior transverse scanning in healthy subjects.
124 (2020). Ultrasound Med. Biol. 37, 44–52 (2011).
8. Jensen, J. A. Medical ultrasound imaging. Prog. Biophys. Mol. Biol. 31. Prada, G. et al. Echocardiographic applications of M-mode
93, 153–165 (2007). ultrasonography in anesthesiology and critical care.
9. Price, D., Wallbridge, D. & Stewart, M. Tissue Doppler imaging: J. Cardiothorac. Vasc. Anesth. 33, 1559–1583 (2019).
current and potential clinical applications. Heart 84 (Suppl. 2), 32. Stabouli, S. et al. Comparison of the SphygmoCor XCEL device
ii11–ii18 (2000). with applanation tonometry for pulse wave velocity and central
10. Sigrist, R. M., Liau, J., El Kaffas, A., Chammas, M. C. & Willmann, blood pressure assessment in youth. J. Hypertens. 37,
J. K. Ultrasound elastography: review of techniques and clinical 30–36 (2019).
applications. Theranostics 7, 1303 (2017). 33. Elliot, C. A., Hamlin, M. J. & Lizamore, C. A. Inter-operator
11. Poelma, C. Ultrasound imaging velocimetry: a review. Exp. Fluids reliability for measuring pulse wave velocity and augmentation
58, 3 (2017). index. Front. Cardiovasc. Med. 7, 72 (2020).
12. Kasban, H., El-Bendary, M. & Salama, D. A comparative study of 34. Hoesein, F. A. M., Zanen, P. & Lammers, J.-W. J. Lower
medical imaging techniques. Int. J. Inf. Sci. Intell. Syst. 4, limit of normal or FEV1/FVC <0.70 in diagnosing COPD: an
37–58 (2015). evidence-based review. Respir. Med. 105, 907–915 (2011).
13. Kalender, W. A. X-ray computed tomography. Phys. Med. Biol. 51, 35. Johnson, J. D. & Theurer, W. M. A stepwise approach to the
R29 (2006). interpretation of pulmonary function tests. Am. Fam. Physician 89,
14. Katti, G., Ara, S. A. & Shireen, A. Magnetic resonance imaging 359–366 (2014).
(MRI)—a review. Int. J. Dent. Clin. 3, 65–70 (2011). 36. Limbu, Y. R., Gurung, G., Malla, R., Rajbhandari, R. & Regmi, S.
15. Díaz-Gómez, J. L., Mayo, P. H. & Koenig, S. J. Point-of- R. Assessment of carotid artery dimensions by ultrasound in
care ultrasonography. N. Engl. J. Med. 385, 1593–1602 non-smoker healthy adults of both sexes. Nepal Med. Coll. J. 8,
(2021). 200–203 (2006).
16. Kenny, J.-É. S. et al. A novel, hands-free ultrasound patch for 37. Morerio, P., Cavazza, J. & Murino, V. Minimal-entropy correlation
continuous monitoring of quantitative Doppler in the carotid alignment for unsupervised deep domain adaptation. In
artery. Sci. Rep. 11, 7780 (2021). International Conference on Learning Representations (2018).
17. Hu, H. et al. Stretchable ultrasonic transducer arrays for 38. Magder, S. Volume and its relationship to cardiac output and
three-dimensional imaging on complex surfaces. Sci. Adv. 4, venous return. Crit. Care 20, 271 (2016).
eaar3979 (2018). 39. White, D. W. & Raven, P. B. Autonomic neural control of heart rate
18. Wang, C. et al. Monitoring of the central blood pressure during dynamic exercise: revisited. J. Physiol. 592, 2491–2500
waveform via a conformal ultrasonic device. Nat. Biomed. Eng. 2, (2014).
687–695 (2018). 40. Laughlin, M. H., Korthuis, R. J., Duncker, D. J. & Bache, R. J. in
19. Wang, C. et al. Continuous monitoring of deep-tissue Comprehensive Physiology (ed. Terjung, R.) 705–769 (Wiley,
haemodynamics with stretchable ultrasonic phased arrays. Nat. 2011).
Biomed. Eng. 5, 749–758 (2021). 41. O’Rourke, M. F. & Mancia, G. Arterial stiffness. J. Hypertens. 17,
20. Wang, C. et al. Bioadhesive ultrasound for long-term 1–4 (1999).
continuous imaging of diverse organs. Science 377, 42. Plowman, S. A. & Smith, D. L. Exercise Physiology for Health Fitness
517–523 (2022). and Performance (Lippincott Williams & Wilkins, 2013).
21. The Ultrasound Monitoring Patch (Pulsify Medical, 2016); 43. Salvi, P. Pulse Waves: How Vascular Hemodynamics Affects Blood
https://pulsify-medical.com/ Pressure 19–68 (Springer, 2017).
22. Project Ulimpia (Penta, 2021); https://penta-eureka.eu/ 44. Munir, S. et al. Exercise reduces arterial pressure augmentation
wp-content/uploads/2022/02/Penta_Project-Ulimpia_Impact_ through vasodilation of muscular arteries in humans. Am. J.
Summary-18_11_2021.pdf Physiol. Heart. Circ. Physiol. 294, H1645–H1650 (2008).
23. Baribeau, Y. et al. Handheld point-of-care ultrasound probes: 45. Antonini-Canterin, F. et al. Arterial stiffness and ventricular
the new generation of POCUS. J. Cardiothorac. Vasc. Anesth. 34, stiffness: a couple of diseases or a coupling disease? A review
3139–3145 (2020). from the cardiologist’s point of view. Eur. J. Echocardiogr. 10,
24. Hu, H. et al. A wearable cardiac ultrasound imager. Nature 613, 36–43 (2008).
667–675 (2023). 46. Brett, S. E., Ritter, J. M. & Chowienczyk, P. J. Diastolic blood
25. Lin, M., Hu, H., Zhou, S. & Xu, S. Soft wearable devices for pressure changes during exercise positively correlate with serum
deep-tissue sensing. Nat. Rev. Mater. 7, 850–869 (2022). cholesterol and insulin resistance. Circulation 101,
26. IEEE Standard for Information Technology–Telecommunications 611–615 (2000).
and Information Exchange Between Systems Local and 47. Scolletta, S., Biagioli, B. & Giomarelli, P. in Anaesthesia, Pain,
Metropolitan Area Networks-Specific Requirements—Part 11: Intensive Care and Emergency A.P.I.C.E. (ed Gullo, A.) Ch 21
Wireless LAN Medium Access Control (MAC) and Physical Layer (Springer, 2007).
(PHY) Specifications (IEEE, 2016); https://standards.ieee.org/ 48. Stöhr, E. J., González-Alonso, J. & Shave, R. Left ventricular
ieee/802.11/5536/ mechanical limitations to stroke volume in healthy humans during
27. Carovac, A., Smajlovic, F. & Junuzovic, D. Application of incremental exercise. Am. J. Physiol. Heart. Circ. Physiol. 301,
ultrasound in medicine. Acta Inform. Med. 19, 168 (2011). H478–H487 (2011).
28. Feigenbaum, H. Role of M-mode technique in today’s 49. Vieira, S. S. et al. Does stroke volume increase during an
echocardiography. J. Am. Soc. Echocardiogr. 23, 240–257 incremental exercise? A systematic review. Open Cardiovasc.
(2010). Med. J. 10, 57 (2016).

Nature Biotechnology | Volume 42 | March 2024 | 448–457 456

Article https://doi.org/10.1038/s41587-023-01800-0

50. Glynn, A. J. et al. The Physiotherapist’s Pocket Guide to Exercise 62. Lane, C. J. The inspection of curved components using flexible
E-Book: Assessment, Prescription and Training (Elsevier, 2009). ultrasonic arrays and shape sensing fibres. Case Stud. Nondestr.
51. Guiraud, T. et al. High-intensity interval training in cardiac Test. Eval. 1, 13–18 (2014).
rehabilitation. Sports Med. 42, 587–605 (2012). 63. Hunter, A. J., Drinkwater, B. W. & Wilcox, P. D. Autofocusing
52. Hannan, A. L. et al. High-intensity interval training versus ultrasonic imagery for non-destructive testing and evaluation
moderate-intensity continuous training within cardiac of specimens with complicated geometries. NDT & E Int. 43,
rehabilitation: a systematic review and meta-analysis. Open 78–85 (2010).
Access J. Sports Med. 9, 1–17 (2018). 64. Huang, Z. et al. Three-dimensional integrated stretchable
53. Gao, W. et al. Fully integrated wearable sensor arrays for electronics. Nat. Electron. 1, 473–480 (2018).
multiplexed in situ perspiration analysis. Nature 529, 509 (2016). 65. Yin, L. et al. A self-sustainable wearable multi-modular E-textile
54. Kim, J. et al. Battery-free, stretchable optoelectronic systems for bioenergy microgrid system. Nat. Commun. 12, 1542 (2021).
wireless optical characterization of the skin. Sci. Adv. 2, e1600418 66. Baik, S. et al. Bioinspired adhesive architectures: from skin
(2016). patch to integrated bioelectronics. Adv. Mater. 31, 1803309
55. Huang, X. et al. Materials and designs for wireless epidermal (2019).
sensors of hydration and strain. Adv. Funct. Mater. 24, 3846–3854 67. Isozaki, A. et al. AI on a chip. Lab Chip 20, 3074–3090 (2020).
(2014). 68. Gao, X. et al. A photoacoustic patch for three-dimensional
56. Kim, J. et al. Miniaturized battery-free wireless systems for imaging of hemoglobin and core temperature. Nat. Commun. 13,
wearable pulse oximetry. Adv. Funct. Mater. 27, 1604373 (2017). 7757 (2022).
57. Naqvi, J., Yap, K. H., Ahmad, G. & Ghosh, J. Transcranial Doppler 69. Wang, F. et al. Flexible Doppler ultrasound device for the
ultrasound: a review of the physical principles and major monitoring of blood flow velocity. Sci. Adv. 7, eabi9283 (2021).
applications in critical care. Int. J. Vasc. Med. 2013, 629378 70. Li, S. et al. Stretchable electronic facial masks for sonophoresis.
(2013). ACS Nano 16, 5961–5974 (2022).
58. Koski, J. M. et al. Assessing the intra- and inter-reader reliability of
dynamic ultrasound images in power Doppler ultrasonography. Publisher’s note Springer Nature remains neutral with regard to
Ann. Rheum. Dis. 65, 1658–1660 (2006). jurisdictional claims in published maps and institutional affiliations.
59. Banegas, J. R. et al. Effectiveness of blood pressure control
outside the medical setting. Hypertension 49, 62–68 (2007). Springer Nature or its licensor (e.g. a society or other partner) holds
60. Cosson, E. et al. Detecting silent coronary stenoses and exclusive rights to this article under a publishing agreement with
stratifying cardiac risk in patients with diabetes: ECG stress test the author(s) or other rightsholder(s); author self-archiving of the
or exercise myocardial scintigraphy? Diabet. Med. 21, 342–348 accepted manuscript version of this article is solely governed by the
(2004). terms of such publishing agreement and applicable law.
61. Miyai, N. et al. Blood pressure response to heart rate during
exercise test and risk of future hypertension. Hypertension 39, © The Author(s), under exclusive licence to Springer Nature America,
761–766 (2002). Inc. 2023

Nature Biotechnology | Volume 42 | March 2024 | 448–457 457

Article https://doi.org/10.1038/s41587-023-01800-0

Methods Spirometer test of respiratory function

Materials for device fabrication The respiratory function was tested with a clinical spirometer (SP-250,
Silicone elastomer (Ecoflex-0030) was bought from Smooth-On as the Schiller) on a healthy participant in a static sitting position. The spirom-
encapsulation material of the device. Ultrasonic transducers (PZT-5H) eter recorded the volume and flow speed of the participant during
were purchased from DeL Piezo Specialties. ACF cables were purchased inhalation/exhalation. During the test, the participant wore a nose clip
from Elform. Double-sided fluid-resistant medical silicone adhesives to avoid air leakage and followed the testers’ instructions to breathe
(2477P) were purchased from 3M. in the desired patterns (that is, deep or quick exhale). The spirometer
data were sent to the computer, where they were processed and then
Finite element analyses of fPCB deformations returned to a display to show the test results. The expiratory volumes
Commercial software ANSYS 2022 R1 was used to predict the bending were analyzed and plotted with MATLAB R2019b and Origin 2017.
deformation of the fPCB and the elastic stretchability of the elastomer
package. Twenty-node 3D solid elements (SOLID 186) and implicit static Machine learning model training and validation
analyses were adopted to ensure the convergence of the simulations. Classification models, including MobileNetV2, ResNet, VGG11 and
Bonded definitions were exploited between contact regions without VGG13, were trained on the same dataset to compare their perfor-
friction. An ideal elastic–plastic constitutive relationship was used mance. The training dataset was collected from a healthy participant
to define the copper layer, where the von-Mises stress reached the and contained 3,021 M-mode images labeled as ‘carotid artery image’
yield strength at 357 MPa across any width (corresponding to the Cu and 2,427 images labeled as ‘noncarotid artery image’. The training
yield strain of 30% (ref. 71)). When bending the fPCB, the maximum process was solely image-based using the labeled M-mode images.
principal strain of 0.57% occurred on the circuit components (Supple- During performance validation, the unlabeled M-mode images were
mentary Fig. 12). When stretching the elastomer package, we assumed used as input, and the classification models output the probability
the human skin elastically yielded at <30% strain. Linear elastic proper- of the image containing carotid artery pulses. The machine learning
ties were used to model the human skin, Ecoflex and mold for circuit algorithms were designed with the integrated development environ-
components, where the elastic modulus E and Poisson’s ratio v are ment PyCharm Community Edition 2022.2.
ESkin = 400 kPa and vSkin = 0.48; EEcoflex = 69 kPa and vEcoflex = 0.49; and
Emold = 23 GPa and vMold = 0.3 (refs. 64,72–74). Head motion recording
We measured the head motion using a pair of inertia measurement
Human test protocols and specifications units (LSM6DS3) mounted on the head and torso (Supplementary
The bio-interface excursions, blood pressure waveforms and respira- Fig. 27). The actual head rotation could be recorded by calculating the
tory volumes were measured on healthy participants. All human tests difference between the head unit and the torso unit.
were approved by the University of California, San Diego, Institutional
Review Board protocol 803942. The participants all gave voluntary Visualizing the domain distribution with a t-distributed
consent for ultrasonic measurements. stochastic neighbor embedding algorithm
We used t-distributed stochastic neighbor embedding, a dimension
Ultrasonic test with clinical systems reduction algorithm75, to visualize the distribution of the M-mode
Two clinical ultrasonic probes were used to collect images and data for image dataset. It embedded high-dimensional data into a lower dimen-
benchmarking the USoP in this work—Verasonics Vantage 64 and But- sion data (two-dimensional in this work) and created clusters among
terfly IQ. On the Verasonics Vantage 64 system, a phased array probe similar data points. Before domain adaptation, because there were
P4-2v was used to measure myocardial contraction and fractional short- domains leading to two groups of similar data points, t-distributed
ening based on the left ventricular internal diameter at end-diastole stochastic neighbor embedding created two clusters. After the domain
(LVIDd) and left ventricular internal diameter at end-systole (LVIDs). adaptation, the difference between two domains was eliminated, and
Fractional shortening was calculated by: thus t-distributed stochastic neighbor embedding could only create
one cluster.
LVIDd − LVIDs
Fractional shortening = × 100%
LVIDd
Arterial stiffness index measurement
The arterial stiffness index β could be calculated using the systolic/
On the Butterfly IQ system, a capacitive-micromachined ultra- diastolic pressure and corresponding arterial diameters76:
sound transducer probe was used for collecting B-mode images from
Dd ln( ps /pd )
radial, brachial, carotid and femoral arteries, as well as the abdominal β=
Ds − D d
aorta, heart and diaphragm, to indicate the position and movement
of these tissue interfaces.
where ps and pd are the systolic and diastolic pressures, respectively,
Carotid artery blood pressure measurement using tonometer measured by using a blood pressure cuff, and Ds and Dd are the cor-
We used an Food and Drug Administration-approved tonometer responding systolic and diastolic arterial diameters measured by the
(SphygmoCor) to directly record the carotid blood pressure wave- USoP. The measurement results of β in Supplementary Fig. 36 were
forms. Based on the pressure waveforms, heart rate and the AIx were collected on a healthy subject (the same person tested in Fig. 4).
calculated by the software SphygmoCor CvMS V9; heart rate was cal-
culated as the reciprocal of beat-to-beat intervals, and the AIx was Reporting summary
calculated as described in Supplementary Fig. 35. During the validation Further information on research design is available in the Nature Port-
test, the blood pressure waveforms were measured intermittently while folio Reporting Summary linked to this article.
the subject performed multiple cycling sessions to stimulate changes
in blood pressure. The tonometer was handheld adjacent to the USoP Data availability
to measure the same common carotid artery simultaneously while the The data and material resources supporting the findings of this study
subject was sitting still. These measurements were repeated to record are available within the article and supplementary materials. The raw
~140 cycles of arterial pulsation on a healthy participant for Bland–Alt- data can be found in a publicly accessible repository77 (https://doi.org/
man analysis of blood pressure, heart rate and AIx. 10.6084/m9.figshare.22631047.v4).

Nature Biotechnology
Article https://doi.org/10.1038/s41587-023-01800-0

Code availability authors and should not be interpreted as necessarily representing the
The codes used in M-mode image classification and domain adaptation official policies or endorsements, either expressed or implied, of the
can be found in a publicly available repository (https://github.com/ AFRL or the US Government. This research was partially supported
JackLin95/Autonomous-Ultrasound-CodeData.git). by the National Institutes of Health (NIH) (grant no. 1 R01 EB033464-
01). The content is solely the responsibility of the authors and does
References not necessarily represent the official views of the NIH. All biological
71. Jang, K.-I. et al. Self-assembled three dimensional network experiments were conducted in accordance with the ethical
designs for soft electronics. Nat. Commun. 8, 15894 (2017). guidelines with the approval of the Institutional Review Board of the
72. Lall, P., Goyal, K., Leever, B. & Marsh, J. Thermo-mechanical University of California, San Diego.
deformation in flexible-board assemblies during reflow and
post-assembly usage. In Proc. 2018 7th IEEE Intersociety Author contributions
Conference on Thermal and Thermomechanical Phenomena in M.Y.L., Z.Y.Z., X.X.G. and S. Xu designed the research. M.Y.L., Z.Y.Z.,
Electronic Systems (ITherm) 26–31 (IEEE, 2018). Y.Z.B., R.S.W., G.P. and Z.Y.L. performed the experiments. Z.Y.Z.,
73. Li, C. H., Guan, G. Y., Reif, R., Huang, Z. H. & Wang, R. K. Y.Z.B. and Z.R.Z. designed the signal processing algorithms. M.Y.L.,
K. Determining elastic properties of skin by measuring Z.Y.Z. and S. Xu analyzed the data. M.Y.L., Z.Y.Z. and S. Xu wrote the
surface waves from an impulse mechanical stimulus using paper. All authors provided constructive and valuable feedback
phase-sensitive optical coherence tomography. J. R. Soc. on the paper.
Interface 9, 831–841 (2012).
74. Park, Y. et al. Wireless, skin-interfaced sensors for compression Competing interests
therapy. Sci. Adv. 6, eabe1655 (2020). The authors declare no competing interests.
75. Van der Maaten, L. & Hinton, G. Visualizing data using t-SNE.
J. Mach. Learn. Res. 9, 2579–2605 (2008). Additional information
76. Hayashi, K., Handa, H., Nagasawa, S., Okumura, A. & Moritake, Extended data is available for this paper at https://doi.org/10.1038/
K. Stiffness and elastic behavior of human intracranial and s41587-023-01800-0.
extracranial arteries. J. Biomech. 13, 175–184 (1980).
77. Lin, M., Zhang, Z. & Gao, X. A fully integrated wearable ultrasound Supplementary information The online version contains
system to monitor deep tissues in moving subjects. figshare supplementary material available at https://doi.org/10.1038/
https://doi.org/10.6084/m9.figshare.22631047.v4 (2023). s41587-023-01800-0.

Acknowledgements Correspondence and requests for materials should be addressed to

We thank Z. Wu, W. Qiu and X. Tian for guidance on the ultrasonic Sheng Xu.
sensor and circuit design; A. Abdal for helping with mechanical
simulation; and J. Li for sensor fabrication. The material was partially Peer review information Nature Biotechnology thanks Pranav
based on research sponsored by the Air Force Research Laboratory Rajpurkar, Michel Maharbiz and the other, anonymous, reviewer(s) for
(AFRL) under agreement number FA8650-18-2-5402. The US their contribution to the peer review of this work.
Government is authorized to reproduce and distribute reprints for
Government purposes, notwithstanding any copyright notation Reprints and permissions information is available at
thereon. The views and conclusions contained herein are those of the www.nature.com/reprints.

Nature Biotechnology
Article https://doi.org/10.1038/s41587-023-01800-0

Extended Data Fig. 1 | Characterizing bandwidth, axial resolution, and probe in this work. The 2 MHz probe is used for deep organ (for example, heart
penetration of the stretchable ultrasonic probes. a, Pulse-echo response and and diaphragm) sensing. The 4 MHz probe is used for deep major artery (for
bandwidth of the probes with three frequencies. The full width at half maximum example, carotid, femoral, and abdominal aorta) sensing. The 6 MHz probe is
(FWHM) is labeled to show the axial resolution of each probe. The 2 MHz, 4 MHz used for shallow peripheral artery (for example, radial and brachial) sensing. d,
and 6 MHz can achieve 604 μm, 333 μm and 229 μm resolution, respectively. Transmission beam intensities as a function of penetration depth in tissues of the
Three probes could achieve a relative bandwidth of ~50% to their center probes with different frequencies. The intensity decay was measured in water,
frequencies at -3 dB. b, The pulse-echo response of a commercial ultrasound and then converted into tissue decay with an attenuation factor of -0.3 dB/cm/
probe with a center frequency of 3 MHz, which could achieve a relative MHz. Based on the penetration threshold of a -3 dB drop in intensity, the 2 MHz,
bandwidth of 42.3%. c, Tissue targets to be sensed by the stretchable ultrasonic 4 MHz and 6 MHz can penetrate 164.0 mm, 77.7 mm and 9.2 mm, respectively.

Nature Biotechnology
Article https://doi.org/10.1038/s41587-023-01800-0

Extended Data Fig. 2 | Schematics and control sequence of ultrasonic sensing. multiplexing and pulse-echo sensing, which shows the time sequence of the
a, Block diagram and signal transmission lines between the functional modules. receive (Rx) enable, trigger, high-voltage (HV) pulse, clock (CLK), reset (RES),
The control circuit includes two parts: the AFE and the wireless DAQ module. The digital input (Din), and latch enable ( LE ) signals. c, Signals acquired by an
AFE consists of a multiplexer (Mux), a transmit/receive switch (T/R SW), a oscilloscope showing the control sequence of the pulse-echo sensing and
receiver, a sequencer, and a pulse generator. The DAQ module consists of a transducer multiplexing. d, Signals acquired by an oscilloscope showing the
microcontroller (MCU) with an on-chip analog-to-digital convertor (ADC), and a input sequence to the shift register for multiplexing and driving the transducer
Wi-Fi transmitter. The dashed lines are for digital signal transmission and the elements. All figure panels share the same color encoding scheme.
solid lines are for analog signal transmission. b, Simulated control sequence for

Nature Biotechnology
Article https://doi.org/10.1038/s41587-023-01800-0

Extended Data Fig. 3 | Deformation of the packaged USoP. a, 90° bending, b, 90° twisting, and c, 20% uniaxial stretching of the packaged USoP. d, A zoom-in view of
the stretched interconnects.

Nature Biotechnology
Article https://doi.org/10.1038/s41587-023-01800-0

Extended Data Fig. 4 | Skin integration of the conformal USoP device. integration of the device before and after exercise. The USoP could maintain
The soft patch could conform to multiple curved body parts, including robust adhesion to the skin after the subject performs intensive exercise
a, forearm, b, brachium, c, neck, d, lower chest, and e, abdomen. f-g, Skin and sweats.

Nature Biotechnology
Article https://doi.org/10.1038/s41587-023-01800-0

Extended Data Fig. 5 | Pulse wave velocity (PWV) measurements. a, Schematic pulses were recorded to calculate average PTT values. The error bars represent
illustration of the pulse wave propagation paths in this study. Five paths were the measurement standard deviations. d, PWV calculated across five arterial
investigated, including the heart to the abdominal aorta (H-Ao), the heart to segments using the USoP. e, PWV mapping under normal conditions and
the carotid artery (H-CA), the heart to the femoral artery (H-FA), the heart to cold pressor test. The average PWV along each path was calculated from five
the brachial artery (H-BA), and the brachial artery to the radial artery (BA-RA). independent measurements. The error bars indicate the standard deviations
b, Pulse waveforms collected by synchronized USoP pairs. The pulse transit of the measured values. The PWV increases from heart-proximal to heart-distal
time (PTT) was defined as the delay between the diastolic feet of the ventricular branches. There is a regional increase of PWV in H-BA and BA-RA segments owing
contraction and arterial pulses. c, The average PTT values by the USoP and the to cold-induced vasoconstriction.
tonometer, showing consistency for both H-BA and BA-RA. Ten consecutive

Nature Biotechnology
Article https://doi.org/10.1038/s41587-023-01800-0

Extended Data Fig. 6 | The validation metrics of four models on ideal and c, The precision, recall and accuracy validated on ideal images. d, The receiver
compromised image datasets. a, The images used for validation including operating characteristic curves validated on 460 images with a mix of ideal and
ideal carotid artery images and compromised images (for example, noise compromised images, suggesting the best model VGG13 has an area under the
coupled images, artery shifting images and artery missing images). b, The curve value of 99.4%. e, The precision, recall and accuracy validated on mixed
receiver operating characteristic curves validated on 460 ideal images, ideal and compromised images.
suggesting the best model VGG13 has an area under the curve value of 100%.

Nature Biotechnology
Article https://doi.org/10.1038/s41587-023-01800-0

Extended Data Fig. 7 | Continuous monitoring during high-intensity the 12 min training and rest. The carotid blood pressure waveforms and heart
interval training (HIIT). a, Photographs showing the participant performing rate are recorded simultaneously and continuously using the USoP. The systolic
HIIT. Six training sessions, including (i) touch shoulder push-ups, (ii) cycling pressure increased ~25 mmHg between training sessions and rest sessions, while
Russian twist, (iii) push-up rotations, (iv) burpees, (v) side kick through and the diastolic pressure experienced less fluctuation. c, Zoomed-in view of the head
(vi) hand-release push-ups. b, The head motions are recorded by the inertia motions, continuous blood pressure waveforms and heart rate recorded during
measurement units, which show the rolling, yawing and pitching rates during the training sessions.

Nature Biotechnology