EchoWrist: Ultrasound Sensing Tech

• EchoWrist is a wrist-worn device that uses active ultrasound to capture rich biomechanical signals from tissue, tendon, and bone.
• It integrates custom transducer arrays and deep learning pipelines to enable high-fidelity heart-rate monitoring, 3D hand pose tracking, and grip force estimation.
• Evaluations demonstrate robust accuracy and energy efficiency, achieving near-real-time performance under diverse use conditions.

EchoWrist is a class of wrist-worn devices that leverage advanced active acoustic sensing—primarily ultrasound—for continuous, low-power monitoring and feedback of biomechanical signals and hand activity. Employing custom transducer-microphone topologies and deep learning–based inference pipelines, EchoWrist systems integrate functionalities such as high-fidelity heart-rate monitoring, 3D hand pose tracking, continuous grip force estimation, and spatially mapped haptic feedback. Distinct from purely inertial or vision-based approaches, EchoWrist exploits the propagation and reflection of ultrasonic chirps through tissue, tendon, and bone to obtain a rich representation of hand and wrist biomechanics, achieving robust, real-time performance under diverse use conditions. The EchoWrist line is documented in the literature through several notable works (Giordano et al., 2024, Mahmoodi et al., 27 Jul 2025, Lee et al., 2024, Palmer et al., 2022).

1. Sensing Architectures and Device Hardware

EchoWrist devices employ active ultrasound/acoustic emission, typically using surface-mounted piezoelectric or MEMS transducers combined with MEMS microphones, arranged circumferentially or along specific wrist axes for targeted echo acquisition. For continuous heart-rate monitoring, the system integrates:

• A low-power ultrasound pulser (e.g., STHVUP32, ±15 V drive, MCU-configurable PLL, 10–100 MHz)
• A single-element 10 MHz piezoelectric transducer coupled via silicone gel
• Analog envelope extraction chain: multi-stage high-pass/low-pass filtering with noninverting op-amp stages and peak detection, delivering >5× baseband bandwidth reduction (Giordano et al., 2024)
• Energy-efficient microcontroller (e.g., STM32L496 ARM Cortex-M4, 80 MHz, 12-bit ADC @ up to 5 Msps) handling fixed-point digital signal processing

For hand pose and grip force estimation, EchoWrist variants (EchoForce) deploy dual ultrasonic speakers (e.g., OWR-05049T-38D) and MEMS microphones (ICS-43434, SPH0641LU4H-1), positioned dorsally and volarly, or at an oblique angle for optimal tendon coverage. Typical emission schemes use linear FMCW chirps sweeping from 20–29 kHz, with a high sample rate (up to 96 kHz per channel) and minimal hardware height (<5 mm above skin) (Lee et al., 2024, Mahmoodi et al., 27 Jul 2025). All sensor wiring is routed on flexible PCBs to ultralight central processing units and LiPo battery modules (e.g., nRF52840-based, enabling 57.9 mW system power).

In wrist-based haptics, dual 1-DoF tactors are mounted ventrally/dorsally using micro-servo rack–pinion actuators (Hextronik HXT500) for normal skin indentation, with <15 mm extension and <20 g wrist mass (Palmer et al., 2022).

2. Acoustic Signal Modeling and Processing Pipelines

EchoWrist systems reconstruct biomechanical states by analyzing the propagation, reflection, and modulation of intentionally emitted acoustic signals. Core methods include:

• Raw echo model (heart rate):

$$s(t) = \sum_{n=0}^{N-1} A_n(t)\cos(2\pi f_0 t + \varphi_n(t)),\quad f_0=10\,\text{MHz}$$

Tissue pulsatility and movement modulate both amplitude $A_n(t)$ and phase $\varphi_n(t)$ over each pulse-repetition time (Giordano et al., 2024).

• Envelope extraction: Analog filtering suppresses the carrier, yielding a baseband envelope $e(t)$:

$$H(f) \approx \frac{1}{\sqrt{1 + (f/f_c)^{2n}}},\quad f_c \approx 2\,\text{MHz},\ n=2$$

This allows ADC sampling bandwidth reduction from 10 MHz to 2 MHz without loss of HR information (correlation $r(92)=0.99$, $p<0.001$).

• FMCW processing (pose/force): Chirps of the form

$$s(t) = A\cos\left(2\pi\left[f_0 + \frac{\Delta f}{T}t\right]t\right),\quad t \in [0,T]$$

are cross-correlated with received echoes to derive 1D range profiles, which are stacked over time to form 2D “echo maps” (Lee et al., 2024, Mahmoodi et al., 27 Jul 2025).

• Differential echo profiles: To suppress static tissue and highlight dynamic deformation, echo frames are differenced:

$\Delta E(t, r) = E(t, r) - E(t-1, r)$

• Feature cropping and windowing: For pose, 72×72×4 “images” are constructed; for grip force, 160×78 “echo maps” over 2 s moving windows encapsulate depth/time evolution at the skin–tendon interface (Lee et al., 2024, Mahmoodi et al., 27 Jul 2025).

3. Embedded Inference and Learning Approaches

EchoWrist systems embed both lightweight signal processing and more complex neural inference pipelines:

• Onboard pulse extraction (heart-rate): Pulse–echo data is differentiated along the time axis, fast-time FFTs are computed, and spectral peaks in [0.5, 2 Hz] are mapped to HR: $A_n(t)$9 All processing occurs in fixed-point Q1.15 format with CMSIS-DSP primitives, yielding 71 ms update latency and energy usage of 1.21 mJ per HR estimate (Giordano et al., 2024).
• Deep learning for pose/action/force: Modified ResNet-18 backbones are trained on the stacked echo maps (either 72×72×4 for pose, 1050×88×4 for object interactions) or 160×78 for grip force. Losses include mean joint Euclidean distance error (MJEDE) for pose or MSE for force. User-independent (“foundation”) and user-dependent fine-tuned models are both supported, with LOPO cross-validation and input data augmentation to account for remounting variance (Lee et al., 2024, Mahmoodi et al., 27 Jul 2025).
• Latency: For 3D hand pose and interaction recognition, total algorithmic latency, including sensor-to-BLE to inference, is ∼0.44–0.54 s (Lee et al., 2024).

4. Experimental Evaluations and Performance Metrics

EchoWrist designs have undergone controlled, multi-participant evaluations encompassing cardiovascular, motor, and haptic sensing domains:

• Heart-rate monitoring: In (Giordano et al., 2024), over 92 one-minute recordings from 10 adults, lateral wrist placement (proximal to the radial artery) yielded a Pearson $r=0.99$ (p<0.001) and mean error $A_n(t)$0 bpm versus synchronized ECG. Power averaged 5.8 mW (envelope filter 53%, MCU 25%, pulser 22%), enabling >7 days operation on standard smartwatch batteries. Performance was strongly anatomical-location dependent (central/medial positions: $A_n(t)$1, $A_n(t)$2).
• 3D hand pose and interaction: In (Lee et al., 2024), 12 users yielded fine-tuned pose errors of $A_n(t)$3 mm (SD 0.99), mean joint angle error $A_n(t)$4 and 97.6% interaction recognition accuracy for 12 daily classes; user-independent pose errors were higher (12.20 mm/7.37°). Robustness to background noise and minor sensor-skin height variations was empirically verified.
• Grip force estimation: (Mahmoodi et al., 27 Jul 2025) reported 9.08% mean error (user-dependent, RMSE 2.31 kg) and 12.3% user-independent accuracy in 11 adults across multiple wrist orientations and remount sessions.
• Haptic feedback evaluation: For pick-and-place VR tasks (Palmer et al., 2022), dual-tactor feedback with a linear force–displacement mapping ($A_n(t)$5) yielded distinguishable and subjectively useful feedback. Finger–tactor mappings (index→dorsal, thumb→ventral) produced slightly higher intuitiveness without changing completion times. No statistically significant mapping dependence was observed for task efficiency; visibility of the manipulated virtual object was the dominant factor.

Summary Table: Key Performance Metrics

Functionality	Metric	Value	Reference
Heart-rate (lateral)	Pearson $A_n(t)$6, mean error	0.99, $A_n(t)$7 bpm	(Giordano et al., 2024)
3D hand pose (fine-tuned)	MJEDE	4.81 mm (SD 0.99)	(Lee et al., 2024)
12-class interaction	Accuracy	97.6% (SD 0.82%)	(Lee et al., 2024)
Grip force (user-dep.)	RMSE, error rate	2.31 kg, 9.08% MVC	(Mahmoodi et al., 27 Jul 2025)
Grip force (user-indep.)	RMSE, error rate	3.11 kg, 12.3% MVC	(Mahmoodi et al., 27 Jul 2025)
Haptic feedback	Tactor displacement mapping	$A_n(t)$8	(Palmer et al., 2022)

5. Advantages, Limitations, and Anatomical Considerations

EchoWrist yields several technical advantages relative to prior inertial, vision, or EMG-based wearables:

• Non-contact, robust acquisition: Acoustic/ultrasound emission is insensitive to color, lighting, or occlusion by mild clothing, and can penetrate superficial tissue to image deep veins and tendons.
• Miniaturization and energy efficiency: <5 mm sensor profile, system power ∼1/10–1/100 that of camera-based approaches (∼58 mW vs. 3.6 W), and memory footprint of 68 kB for HR extraction (Giordano et al., 2024, Lee et al., 2024).
• Privacy: Operation in near-ultrasound (20–29 kHz) avoids recording audible speech or video imagery.
• Wearability: No electrodes, glue, or forearm straps; sensors are easily integrated into standard smartwatch or wristband form factors.
• Spatial acuity: For HR extraction, lateral wrist placement over the radial artery