Near-infrared spectroscopy–enabled electromechanical systems for fast mapping of biomechanics and subcutaneous diagnosis

Abstract

Fast and accurate assessment of skin mechanics holds great promise in diagnosing various epidermal diseases, yet substantial challenges remain in developing simple and wearable strategies for continuous monitoring. Here, we present a design concept, named active near-infrared spectroscopy patch (ANIRP) for continuously mapping skin mechanics. ANIRP addresses these challenges by integrating near-infrared (NIR) sensing with mechanical actuators, enabling rapid measurement (<1 s) of Young’s modulus, high spatial sensing density (~1 cm²), and high spatial sensitivity (<1 mm). Unlike conventional electromechanical sensors, NIR sensors precisely capture vibrational frequencies propagated from the actuators without needing ultraclose contact, enhancing wearing comfort. Demonstrated examples include ANIRPs for comprehensively moduli mapping of artificial tissues with varied mechanical properties emulating tumorous fibrosis. On-body validation of the ANIRP across skin locations confirms its practical utility for clinical monitoring of epidermal mechanics, promising considerable advancements in real-time, noninvasive skin diagnostics and continuous health monitoring.

A wearable patch with NIR sensing and electromechanical actuation allows noninvasive, continuous monitoring of skin mechanics.

INTRODUCTION

Rapid and accurate approach to assessing the biomechanics of soft tissue at depth holds great promise in advancing biomechanics research and clinical diagnosis (1, 2). The strong correlation of tissue biomechanics with dehydration (3), fatigue (4), and other onset pathologies (5, 6) allows its clinical utility as an admission-level biomarker to screen early signs of skin anomalies. For example, skin cancer, one of the most common forms of cancer globally, presents substantial detection challenges due to its subtle symptoms at an early stage and unpredicted rapid progression (7, 8). Early detection is critical as it greatly increases survival rates, yet existing methods can be invasive and reliant on specialist expertise (7, 9). However, skin cancers, like melanoma, often develop together with fibrosis or hyperkeratosis, which can lead to notable changes in the skin’s stiffness (10, 11). The elastic modulus of common tumorous tissue (like melanoma or breast cancer) can range from 2 to 10 times the corresponding normal tissues, thus enabling its biomechanical differentiation (10, 12). Similarly, disorders like psoriasis (13), edema (14), keloids (15, 16), morphea (17–19), and scleroderma (20–23), characterized by rapid skin cell turnover and biomechanics change, pose challenges in ongoing severity assessment and monitoring. Rapid and continuous spatial mapping of skin biomechanics with depth may provide a real-time risk evaluation of those preventable and treatable diseases (24).

Current state-of-the-art methods (table S3) of measuring skin and/or subcutaneous biomechanics include ultrasonography (25–27), magnetic resonance elastography (MRE) (28–30), optical coherence elastography (31, 32), indentometry (33), cutometry (34–36), or other mechanical stimulation including torsion (37), tension, and compression (38). However, these techniques mentioned above are rife with inherent limitations. Traditional techniques such as MRE and ultrasonography are highly specialized and require hospital visits or expensive instruments, making them inaccessible and cost-prohibitive to many individuals in need. Direct analyses via bio-indentation, vacuum deformation, torsion, and tension/compression can often require biopsies to be performed, which, while effective, can be uncomfortable and invasive. In addition, some of the diagnoses can take hours or even weeks to interpret, making continuous monitoring unsustainable. Hence, there is a pressing need for accurate, accessible methods of broad-scale mapping of skin biomechanics for diagnosis in clinical practice.

Electromechanical modulus (EMM) sensing has emerged as a promising way of performing quick analysis of skin modulus (1, 39–41). Existing EMM sensors primarily use piezoelectric transducers to inject vibrational waves into the skin in a highly adjustable manner (42, 43). By measuring the strain and stress, mechanical properties including elasticity, stiffness, or viscosity can be obtained (44–46). Recent development of linear resonance actuator (LRA) provides another way to deliver mechanical stimulation to skin (1). While being similar to piezoelectric transducers using alternating current (AC), LRA usually generates vibrations with higher magnitude, allowing its interaction with subcutaneous tissue in depth. Existing studies follow a basic mechanism that uses actuators to deliver vertical vibration to the skin at controlled frequency and amplitude. As the strain sensor is closely attached to the surface of the skin, the deformation or motion of the sensor and the skin match each other, thus reflecting the mechanical properties upon mechanical stimulation. Normally, high-modulus skin shows lower deformation upon certain stimulation. The strain sensor converts the mechanical deformation into an electrical signal, which will subsequently undergo lock-in amplification prior to data acquisition. The elastic modulus and electrical signal amplitude are therefore related. In general, EMM sensing provides an easy and comfortable way to perform mechanical analysis on skins. However, these EMM sensors all require individual measurements of local strain on any sites of interest. Meanwhile, the lock-in amplification of signals at specific frequencies complicates data acquisition.

Simultaneously, the application of near-infrared (NIR) light in medical diagnostics has become pivotal, especially in noninvasive optical techniques like photoplethysmography (PPG). NIR light, with its deeper skin penetration compared to visible light, interacts dynamically with skin tissues, leading to absorption and scattering by components such as keratinocytes, red blood cells, and blood vessels. These interactions are crucial as they modulate the light signal in both temporal and frequency domains, establishing signal correlations with topological chromophores of soft tissue (46–48). However, current NIR- or other optics-based methods for tissue elastography remain limited in both detection depths and highly specialized instrumentation, as summarized in table S4 (49, 50).

Here, instead of relying on electromechanical sensors as primarily used in existing biomechanical sensors, our design uses a skin-interfaced NIR spectroscopy (NIRS) coupled with ERM (eccentric rotating mass) actuators to realize continuous, noninvasive assessment of soft-tissue biomechanics in depth. This device, which we name the active NIR patch or ANIRP, delivers periodic mechanical vibrations to the skin using ERM actuators and measures the vibrational frequency transmitted through a certain distance of skin using NIRS. The response frequency of skin vibration shows a strong correlation with the local biomechanical properties of the actuated site, such as elastic modulus or damping factors, thus reflecting the regional biomechanics. Such integration of NIRS with vibration actuators offers an unconventional strategy to assess the biomechanics of biological tissues, enabling evaluations of their elastic modulus via small-scale mechanical stimuli across various frequencies, depths, and spatial scales. The engineering concepts for the ANIRP allow for tunable depth profiling, averaging measurements of tissue moduli over millimeter-scale depths, and enabling biomechanics mapping over a large area. Experimental and simulation studies demonstrate the system’s quantitative measurements of tissue moduli for different substrates and conditions. The findings hold potential for diverse applications in clinical evaluations of patients with skin disorders and can be further developed for large-area mapping of elastic moduli (46–48). By integrating multiple actuators, the scanning of skin can be completed in a short time for convenient, noninvasive, and fast diagnosis of skin diseases. The rapid and accurate mechanical mapping feature demonstrated by our ANIRP greatly solves the problems of traditional direct dermatology measurement of skin, which often takes time while being uncomfortable. With the special design of multi-actuator ANIRP, the patch can realize conformal contact as a wearable device on multiple body areas. The deploy-and-use feature greatly reduces the requirement of professional operations and does not need special bio-adhesives on the device-skin interface. Designed to have multiple working modes, ANIRP can either be used for quick detection or continuous monitoring, as it can both differentiate possible lesion areas and read out minimal changes in subcutaneous muscles. Moreover, compared with its counterparts of EMM sensors, ANIRP detects mechanical vibrations by NIRS instead of strain sensing, which greatly reduces the number of sensors required in the device. The NIRS sensor shows highly robust, precise, and reliable detection of vibration frequency while being much more durable than traditional strain sensors. In addition, ANIRP is completely fabricated from low-cost components without involving microfabrication, making it promising for at-home dermatology health care.

RESULTS AND DISCUSSION

Configuration and mechanism of ANIRP

Figure 1A presents a schematic illustration of ANIRP, which consists of two parts: (i) an array of actuators to deliver active mechanical vibration to the skin and (i) one or several light-emitting diode (LED)–photodiode (PD) sensors consisting of NIR LEDs and PDs. The components are integrated by a flexible printed circuit board (PCB) and connected to the microcontroller [as demonstrated in fig. S46 by a multichannel flat flexible cable (FFC)]. Figure 1B illustrates the structure of ANIRP, which mainly consists of three layers (from top to bottom): (i) the ERM actuators, the FFC connector, and the front side circuits; (ii) the polyimide (PI) substrate with hollow structures; and (iii) the NIR LED, the PD, and the back side circuits. The structure of the PCB that holds the ERM actuators is specially designed to minimize the propagation of mechanical vibration on the device, thus ensuring minimized interference between different actuation sites. Figure S1 further describes the local structure of the ERM actuators and the localization of vibration.

Fig. 1. Design and operational mechanism of the active NIRS patch (ANIRP).

(A) Schematic illustration of ANIRP. (B) The exploded view of ANIRP, including three layers [(i) the actuator layer of eccentric rotation mass (ERM) actuators; (ii) the substrate layer flexible polyimide (PI) with conductive patterns and cable connections; and (iii) the sensor layer of light emitting diodes (LEDs) and photodetectors (PDs)]. (C) Schematic illustration showing the forced vibration of the skin induced by an ERM actuator from the ANIRP. The propagating mechanical waves (highlighted by blue bands) modulates the sensing signal in a way associated with the underlying skin modulus. (D) Schematic illustration with a cross-sectional view on an adjacent pair of an LED and a PD capturing the frequency of mechanical waves propagated along the skin from the actuator. The mechanical wave deforms the skin, affecting the light pathway within, thus leading to temporal variations in collected photocurrent (I_t). (E and F) The outlook of the wearable ANIRPs of different configurations for biomechanics sensing. (E) A large ANIRP of 16 ERM actuators (0.5 cm) under a hollow mesh design. (F) A medium ANIRP of 4 ERM actuators (1 cm) under a conformal patch design. Scale bars, 1 cm.

Figure S2 describes the low cost and easy fabrication of ANIRP. The fabrication starts from the double-sided copper patterning of the PI substrate, realized by IR laser ablation and wet etching. Then, the hollow structure on the PI substrate is created by ultraviolet (UV) laser ablation. After that, the components (the LED, PD, ERM actuators, and the FFC connector) are soldered onto the prepared circuits. The customization of our ANIRP allows for the integration of multiple ERM actuators from 1 to more than 10, different dimensions of actuators from 0.5 to 1.4 cm (diameter), and a relatively high actuator density (up to 1 cm⁻²), meeting the requirements of various scenarios.

Figure 1 (C and D) demonstrates the working mechanism of the ERM actuator and the LED-PD sensor. An ERM actuator generates vibration from the rotational movement of its eccentric mass, which is described by the following simplified model

$$F_{\text{centripetal}}=-F_{\text{centrifugal}}=m{\mathrm{\omega }}^{2}r$$ (1)

where F_centripetal and F_centrifugal are the magnitude of the centripetal and centrifugal forces, m is the mass of the rotational object, ω is the angular velocity of the rotation, and r is the equivalent radius of the rotation. For the direct current (DC) motor in an ERM actuator (as shown in fig. S3A), the angular velocity, or spinning rate, is directly controlled by the input power. As the centripetal force affects the movement of the rotational object, the vibration amplitude of an ERM actuator cannot be adjusted separately (51). When an ERM actuator is mounted to a rigid, stationary substrate, the surface provides any centripetal force required for rotation. Thus, the actuator vibrates at its intrinsic angular velocity (ω₀) under a certain input power, as illustrated in fig. S3B. When it is mounted to a damping substrate, the dynamic deformation of the substrate changes both the centripetal force and the equivalent radius of rotation, resulting in a decrease in the resulting vibration velocity (ω_damped) from the intrinsic velocity. Therefore, when the ERM actuator is under constant input power, the resulting vibration frequency is related to the mechanical properties of the substrate. We also analyze the mechanism from the perspective of energy consumption (note S1). The torque output (T) of the DC motor inside the ERM actuator is determined by the input power (V_INI) and internal resistance (R)

$$T=\frac{1}{\mathrm{\omega }}(V_{\text{IN}}I-I^{2}R)$$ (2)

In the meantime, the output torque balances the Coulomb friction (B_c) and the viscous substrate damping friction (B_vω) under steady rotation, where ω is the angular velocity. In a simplified model, the viscous substrate damping coefficient (B_v) is proportional to deformation volume (V_d)

$$T=B_{\mathrm{v}}\mathrm{\omega }+B_c=k\bullet V_{\mathrm{d}}\mathrm{\omega }+B_{\mathrm{c}}$$ (3)

The volume of deformation (V_d) is strongly correlated to the modulus of the substrate under the same level of actuation where substrates with lower modulus have a higher V_d, while substrates of higher modulus or stiffness have lower V_d. Therefore, a decrease in Young’s modulus results in a larger V_d and therefore a lower ω. When the modulus is high enough, there is little difference in V_d upon actuation, thus showing little damping effect on ω or the vibration frequency from its intrinsic (undamped) value. This is also observed in our calibration (Fig. 2C).

Fig. 2. Validation of skin-modulus measurement of ANIRP and its operational workflow.

(A) The calibration curve of the intrinsic vibration frequency with varying driving power (in voltage) of the 10-mm ERM actuator. (B) The calibration of the elastic modulus of PDMS samples of varying mixing ratios. (C) The calibration curve of the response frequency of PDMS samples, using a 10-mm ERM actuator and the driving power of 2.5 V. (D) The combined spectrograms (5 s for each sample) corresponding to (E) and (F), with a unified color bar. The color presents the power spectral density (PSD) in dB/Hz, reflecting signal intensity. (E and F) The 3D plot of the normalized fast Fourier transform (FFT) signal of the PDMS response frequency calibration from different perspectives. (G) The workflow of biomechanics mapping based on multi-actuator ANIRP: (i) An ERM actuator creates the active mechanical vibration on the skin, which is captured by the LED-PD sensor; (ii) digital filters and FFT are applied to the amplified signal and transformed it into the frequency domain signal or a spectrogram for real-time analysis; (iii) the response frequency is detected and analyzed; and (iv) the corresponding elastic modulus of the testing point is obtained from the calibration curve and marked on the skin.

Upon deployment, the ERM actuator creates a forced vibration in the skin. This vibration consists of transverse waves and longitudinal waves and propagates along the surface of the skin (movies S1 and S2) (51). The frequency of the forced vibration is mainly determined by the local Young’s modulus of the skin and the related subcutaneous tissues right under the actuator. The tissue right in contact with the ERM actuator provides different damping effects determined by the local modulus, resulting in different response vibration frequencies. By changing the position of the actuator, the vibration frequency changes due to the biomechanical difference of skin, thus capturing local skin and subcutaneous biomechanical properties (Fig. 1C). Meanwhile, the frequency does not change as the vibration propagates through the surface of the skin. The frequency and intensity of the vibration signal are characterized in a series of experiments by moving the sensor away from the vibration source. The intensity of the vibration signal decays fast within 5 cm but stays quite stable at a farther distance, while the frequency of the signal shows no obvious change. A relative increase in the 180-Hz noise intensity as distance increases can be found in the normalized frequency domain plots and the spectrograms, indicating a higher signal-to-noise ratio (SNR) at a closer distance. When the mechanical wave reaches the sensor, the sensor collects the vibration signals based on the periodic changes in the light path from the IR LED to the PD (Fig. 1D). Two possible detailed mechanisms (distance changing and curvature changing) are discussed in fig. S4. In the distance-changing mechanism, the thickness of the air gap between the sensor and the skin changes in the vertical direction. When the sensor moves far away from the skin, the optical current gets weaker because of a longer light path. A higher portion of the IR light is absorbed or scattered before reaching the PD. Thus, the photocurrent fluctuates with the mechanical vibration. In the curvature-changing mechanism, the distance between the sensor and the skin remains constant. Instead, the local curvature of the skin under the sensor changes as the skin vibrates. When the curvature radius gets smaller as the local skin bends upward, the LED and PD are facing more to each other, thus having more IR light received by the PD. Figure 1 (E and F) shows the actual outlook of ANIRP and the way an ANIRP can be deployed on the forearm. A piece of Tegaderm film is covered from the top of the device to ensure a fixed position between the device and the skin.

Workflow of ANIRP applications and related calibration

Establishing the relation between Young’s modulus and response frequency is the key to measuring biomechanics. We first calibrate the power curves (fig. S5, A and B) and the intrinsic vibration frequency (without damping) of both the 1- and 0.5-cm ERM actuators by a piezoelectric sensor (Fig. 2A and fig. S5, C and D). The 1-cm ERM actuator functions stably above 2.0 V, while the 0.5-cm type actuator functions stably above 2.5 V. A positive relation is observed between the input power (V) with the intrinsic vibration frequency (Hz). Overall, the 0.5-cm actuator has a higher intrinsic vibration frequency than that of the 1-cm actuator. Figures S6 and S7 present the related spectra and 5-s spectrograms corresponding to fig. S5 (C and D). Digital filters are applied to the signals to remove artifacts and the 60-Hz noise. Then, the 0 to 400 Hz frequency domain spectra are normalized to 0-1 and then stacked with offset for comparison. The spectrograms are 5-s slices from the signals processed by the same digital filters, with a unified color bar. The highest peak in the range of 0 to 400 Hz is recognized as the response signal, which is the dominant signal in most cases. We also observe that the spectrum gets more peaks when the ERM is provided with a low power (such as a 1-cm ERM actuator powered at <1.0 V). This is mainly because of the relatively low vibration amplitude and the consequent low signal intensity of the response frequency. In such cases, the signal is not as prominent as it is when the vibration is strong, compared to noise and interference, indicating that a minimum input power is required for the application. The input power to the ERM actuator determines the intrinsic vibration frequency while affecting the response frequency under certain damping environments (note S1).

Figure S5 (E and F) presents the relation between input power and the response frequency when other factors are controlled. A 1-cm actuator (fig. S5E) and a 0.5-cm actuator (fig. S5F) are tested on a piece of polydimethylsiloxane (PDMS) sample (15:1 mixing ratio, 5 mm thick). The response frequency is strongly affected by the type of actuator and the input power, and it is important to control these factors. Here, we choose 2.5 V DC as the standard input for the 1-cm ERM actuator in the following calibration and measurements and 3.0 V DC for the 0.5-cm ERM actuator unless specified, both at a power around 150 mW. Figures S8 and S9 present the related spectra (normalized) and 5-s spectrograms corresponding to fig. S5 (E and F).

Then, we fabricate a series of PDMS artificial skin models [mixing ratio from 5:1 to 40:1 (w/w), 1:500 (w/w) pigment loaded, and 3 mm in thickness] and measure their response frequencies by a single actuator ANIRP (type 1). The pigment loaded into PDMS helps simulate the absorption and scattering of light in the skin. In the meantime, we characterize the Young’s modulus of the PDMS models by nanoindentation (Fig. 2B). Figure S10 describes the experimental details of nanoindentation. Three steps of the indentation (loading, holding, and unloading) constitute a complete cycle of indentation hysteresis, as shown in fig. S10 (A to C). The fitting of the loading section provides the estimation of Young’s modulus by the following equation

$$E=\frac{3(1-v^2)F}{4R^{0.5}{\mathrm{\delta }}^{1.5}}$$ (4)

In this equation, E is the Young’s modulus of the measured local surface, v is the Poisson’s ratio of the sample (typically around 0.5 for PDMS) (52–54), F is the force on the cantilever, R is the radius of the cantilever tip, and δ is the indentation. Figure S11 shows all the hysteresis and fittings. Higher mixing ratio samples (lower modulus) show greater hysteresis area, indicating a greater sample viscosity and stronger damping effect. We also use another type of elastomer, Ecoflex 00-30 and 00-31 instead, which can be even lower in Young’s modulus (fig. S12).

Figure 2 (E and F) presents the response frequency of the PDMS samples stimulated by a 1-cm ERM actuator at 2.5 V. As a result, Fig. 2C reveals the calibration between the response frequency and Young’s modulus. The calibration curve indicates high sensitivity at a lower range of Young’s modulus (<300 kPa) while showing saturated behavior when the modulus goes higher. This is attributed to the capping effect from the intrinsic frequency of the actuator, which is measured at ~213 Hz. An exponential approximation fitting curve is presented, with a max value of 213 Hz. The fitting is calculated as below, with an adjusted R² of 0.98

$$f/\text{Hz}=213-209.18*\text{exp}[-0.00875*(x/\text{kPa}+31.805)]$$ (5)

Figure S13 presents the related spectra (normalized) and 5-s spectrograms corresponding to Fig. 2 (E and F). Figure 2D shows the combined spectrograms corresponding to Fig. 2 (E and F). It is also observed that the harmonics of the 60-Hz noise get stronger when the sample’s modulus decreases, resulting in a lower SNR. This indicates that the mechanical wave attenuates faster when propagating on a softer and more viscous sample.

We further characterize the response frequency by using a 0.5-cm actuator at 3.5 V to provide a higher stimulating frequency, as shown in fig. S14. By increasing the input power or switching to a different actuator that vibrates at a higher frequency, the capping frequency can be further increased and therefore avoided in the interested range of measurement frequency. The normalized magnitude spectra and spectrograms are shown in fig. S14 (B and C).

Figure 2G presents the completed workflow of the biomechanical spatial mapping by ANIRP. When an ANIRP with multiple ERM actuators is deployed on the arm, the ERM actuators will be sequentially switched on for a short time (typically <5 s). A short actuation time per location can enhance the efficiency of mechanical mapping. Meanwhile, a shorter actuation time is also helpful in reducing heat generation on the skin. For large-area sequential mapping, providing stable contact, the actuation duration on each location can be ~1 s for efficiency, as our sampling rate (10 to 30 kHz) is much higher than the response frequency (100 to 300 Hz). To determine the minimum actuation time, we characterize the working stages of the ERM actuator. As illustrated in fig. S15, the ERM actuator undergoes acceleration, stable vibration, and deceleration. On the basis of the analysis of FFT spectra, we confirm that the ERM actuator reaches stable vibration frequency in 0.3 s. The recording duration of the stable vibration stage determines the frequency resolution on the FFT spectrum, while 0.5 to 1 s provides an acceptable resolution of 2 to 1 Hz⁻¹. Therefore, the minimum actuation time, which is the sum of acceleration and stable vibration stages, can be less than 1 s. A certain DC potential (1.2 to1.4 V) is provided to the IR LED continuously, labeled as a red region in the figure. The photovoltage signal after amplification is recorded, processed, and converted to the frequency domain signal by fast Fourier transform (FFT). Then, the characteristic peak on the frequency domain can be detected and compared to the calibration curve, which relates the response frequency to Young’s modulus. A marker of the biomechanical information is added to the corresponding location of the testing point. In the meantime, a real-time spectrogram can be plotted to track the modulus changing by muscle movements.

Optical simulations and performance analysis of the LED-PD sensor

To comprehensively analyze the performance of the EMM sensing by ANIRP, we first conduct a series of optical simulations to better validate the vibration sensing by LED-PD, which is fundamental. The optical simulation is carried out by Monte Carlo eXtreme (MCX) Lab (55–58). As shown in Fig. 3A, we first build up a three-layered skin model, illustrated in detail by fig. S16. The skin model includes three layers: (from top to bottom) epidermis (0.1 mm), dermis (1 mm), and subcutaneous muscle (1.59 mm). The corresponding parameters of 900-nm NIR light and the layers (thicknesses, scattering coefficients, absorption coefficients, and anisotropy factors) are provided in table S1. Then, we define the sensor by placing the light source (representing LED) and the detectors (representing PDs). We use a zenith gaussian (z-gaussian) beam to simulate the angular distribution of light coming from a typical surface-mounted LED. The two detectors we defined are placed next to the light source at a small distance (which varies) symmetrically. More details of the simulation settings are included in the experimental methods. Figure 3B shows the flux diagram of the optical simulation by slices of cross sections corresponding to Fig. 3A. The slices of the skin model start from the center to the edge of the model. We conducted three major simulations: (i) by moving the LED-PD sensor away from the epidermis; (ii) by separating the LED and PDs; and (iii) by increasing the emitted photon number from the LED. For the interpretation of the results, we are mostly interested in the photons received by the detectors, since it is directly related to the photocurrent generated from the PD when exposed to light. Figure 3C shows the received photon ratio versus sensor-to-skin distance. By changing the distance, we mimic the z-direction component (perpendicular) of the vibration of the skin. As the sensor moves away from the epidermis by a short distance (0 to 1 mm), the photocurrent, which is proportional to the received photon number, decreases sharply (fig. S16E). Therefore, we confirmed the vibration sensing function of the LED-PD sensor that the photocurrent changes periodically as the z-direction vibration happens in this simplified model. We also observe an exponential decay of the received photons when separating the LED and PDs, from 1 to 5 mm in simulation ii (fig. S16F). This provides a crucial reference to the design of the sensor that the LED and PDs should be close enough to ensure the signal intensity. Simulation (iii) (fig. S16G) shows an expected linear relation between emitted and received photons when other factors are controlled. To examine the confounding effect of skin tones on the ANIRP sensitivity and accuracy, we carried out an on-body test with various skin tones. Figure S17 shows experimental results on two types of skin tones (one is dark, and the other is light), indicating no significant difference in signal intensity during the modulus measurement of skin.

Fig. 3. Computational analysis and design optimization of ANIRP.

(A) Schematic illustration of a three-layer skin model (from top to bottom: epidermis, dermis, and subcutaneous muscle) used in the optical simulation. The cross-sectional planes of flux analysis are labeled with red slices. (B) Simulated flux diagram corresponding to cross-sectional planes shown in (A). From left to right, the simulated planes are located from 0 to 15 mm away from the center of the LED. (C) Simulation results on the received photon ratio as a function of sensor-to-skin distance. As the sensor moves away from the skin, the number of received photons decreases to 0 at around 1 mm, which means that the sensor is too far away from the skin. (D) Measured response frequency and signal intensity from ANIRP as a function of the thickness of artificial skin. (E) Measured response frequency and signal-to-noise ratio (SNR) as a function of the voltage bias applied to the IR LED. (F) Measured response frequency and SNR as a function of the distance between the ERM actuator and the LED-PD sensor.

Apart from the optical simulations, we design and conduct a series of experiments of ANIRP on artificial skin (PDMS) to better understand its sensing performance. We analyze the effect from sample thickness/weight, NIR light intensity, and sensor-to-actuator distance. Figure 3 (D to F) shows the result of both response frequency and signal intensity. In these experiments, we calibrate the signal intensity based on a specially designed sensor, containing two LEDs and one PD. We conclude that none of these factors will significantly change the response frequency of the measurement. The results indicate that the response frequency is determined by the contact region between the ERM actuator and the skin instead of the entire body (fig. S18) and does not change as the mechanical wave propagates along the skin (fig. S19). We also confirmed by experiments that the vibration frequency measured by the LED-PD sensor is directly correlated with vibrations of the targeted skin regions rather than the vibrations in the ANIRP (fig. S20).

From the light intensity experiment, an extremely weak peak can still be manually picked up from the signal without the existence of NIR light. This signal may be contributed by the environmental light, but easily immersed in much stronger noises. However, when the operating bias of the LEDs rises to 1.2 V, the signal of response frequency becomes dominant. Figure S21 (C and D) shows the corresponding spectra and spectrograms. Red arrows label the harmonics of 60-Hz noises, and green arrows label the response frequency. It can be observed that the SNR becomes much larger when the bias of the LEDs reaches 1.4 V. Although the bias can be higher (up to 1.6 V), this is good enough to provide signal quality while not generating too much heat. An alternative way to increase the light intensity is to increase the number of LEDs, which can be easily designed and fabricated.

The distance between the actuator and the sensor matters much to signal intensity, as the mechanical vibration attenuates when propagating. However, the propagation of vibration can be up to 10 cm level, as demonstrated by movie 2. Figure S19 shows the calibration of actuator-sensor distance under a low operating bias of the LED (1.2 V). The signal can still be observed 7 cm away from the vibration source. This provides reference to the design of our device that multiple ERM actuators may share one sensor at their adjacency. Figure S19 (C and D) shows the corresponding spectra and spectrograms. Red arrows label the harmonics of 60-Hz noises, and green arrows label the response frequency.

Figure S22 shows the effect of hair on real skin measurements. The experiment is conducted on the forearm of a male adult volunteer, as shown in fig. S22A. To minimize the effect of systematic differences in the skin’s modulus on the hairless and hairy locations, we choose two locations on the front side of the arm, which are close enough, but differ largely in hairs. The measurements are repeated five times on each. Both frequency and intensity are not significantly affected by the existence of hair (P = 0.2046 and 0.4788 separately, α = 0.05, and n = 5).

ANIRP for subcutaneous lesion detection

Figure 4 demonstrates the spatial mapping performance of ANIRP. For spatial mapping by active actuation, we develop two types of ANIRP consisting of four ERM actuators (1 cm for type 1 and 0.5 cm for type 2), as shown in fig. S1. To deliver mechanical actuation to specific locations, confining the vibration from the ERM actuators is important. Therefore, we introduce the hollow structures to the ANIRPs to isolate each ERM actuator and minimize the interference. We then fabricated a series of artificial skin models (samples 1 to 4) of subcutaneous diseases. Three-dimensionally (3D) printed plastics [1 mm in thickness, chlorinated polyethylene (CPE)] are embedded into the PDMS (15:1 mixing ratio, 1:300 pigment loaded) before curing, at a depth of around 2 mm. As illustrated in fig. S23, the locations of mechanical mapping are labeled on each sample. Young’s modulus of the plastics was characterized by nanoindentation (fig. S12). The elasticity difference between CPE (~33 MPa) and PDMS (~303 kPa) corresponds to that between normal subcutaneous tissue and fibrosis, characterized by shear wave ultrasound elastography (22, 27). We use type 1 ANIRP to map samples 1 to 3. For Sample 3, the middle locations (6, 7, 10, and 11, in fig. S23C) are labeled by a blue frame, where a high-resolution mapping by type 2 ANIRP is conducted to better detect the deeply embedded structures (~2.5 mm in depth and 0.7 cm in length) as shown in fig. S23D. Locations m1 and m2 are measured separately by a single-ERM ANIRP. We also use type 2 ANIRP to map sample 4. The mapping results are visualized by interpolation plots, as shown in Fig. 4B and fig. S24. We unify the color bar from 130 to 190 Hz for type 1 ANIRP results and from 200 to 300 Hz for type 2 ANIRP results.

Fig. 4. Spatial mapping of skin biomechanics using ANIRP.

(A) Schematic illustration of PDMS models where selected regions (highlighted in brown) are embedded with hard chlorinated polyethylene (CPE) plastic (~33 MPa, at depth ~2 mm) to simulate skin abnormality. In sample 3, several regions (highlighted in pale pink) embed the plastics at a deeper depth (~2.5 mm). (B) Measured mapping results of samples 2, 3, and 4 using ANIRP. (C) Schematic illustration of an artificial skin with plastics embedded at various depths, ranging from 0.13 to 3.13 mm. The locations of ANIRP measurements are labeled with black dots. (D) Measured response frequency using ANIRP as a function of embedding depth of the plastics in the corresponding regions of artificial skin shown in (C). (E and F) Analysis of ANIRP in its spatial differentiation. The experiments collect the response frequency (f) from ANIRP as an actuator is moving across the edge of an embedded plastic (~2 mm in depth), as illustrated in (E). The corresponding response frequency is plotted in (F). Scale bars, 2 cm in (B) and 0.5 cm in (C) and (E).

Figure S25 presents the strip plots and classifications of the spatial mapping results. The mapping results are classified into two to three clusters that are labeled in different colors. Each color represents a condition of the artificial skin: with embedding objects, without embedding objects, and with embedding objects at a deeper level (sample 3, middle only). In samples 1 and 2, type 1 successfully distinguishes positions 1 to 4 (fig. S23, A and B). However, during the mapping of sample 3, type 1 is not able to distinguish the positions in the middle (nos. 6 and 11, fig. S23C), due to the existence of some deeply embedded small structures (0.7 × 0.7 cm² squares, 0.1 mm in thickness, and ~2.5 mm in depth). These two points are labeled by arrows in the strip plot as unsuccessful classifications. This indicates that the spatial resolution of type 1 is mainly limited by the dimension of the ERM actuator and its spacing, which is roughly 2 cm. After switching to type 2, we realize a higher spatial density (~1 cm) of mechanical mapping, as illustrated in sample 3, middle, and sample 4. The results of sample 3, middle, are visualized and classified into 3 clusters in the strip plot, standing for (i) PDMS only, (ii) PDMS with deeper plastics embedded, and (iii) PDMS with plastics embedded (from low to high frequency). Figures S26 to S30 present the detailed frequency domain signals (0 to 300 Hz) and 5-s spectrograms corresponding to spatial mapping from samples 1 to 4. We use arrows to label the harmonics of 60-Hz noise and unify the color bar of the spectrograms.

To analyze the ability to detect subcutaneous abnormal tissues in different depths, we calibrate the response frequency of specially designed PDMS artificial skin (15:1 mixing ratio), with a piece of stair-shape plastics embedded. A type 2 ANIRP (ERM actuator powered by 3.0 V) was used. Figure 4 (C and D) illustrates the study of depth calibration. In this experiment, we use a specially designed ANIRP with a single ERM actuator (0.5 cm) to measure the response frequency of the artificial skin at various locations. The depth of the embedded structure ranges from 0.13 to 3.13 mm, with a step of 0.3 mm, as shown in the cross-sectional illustration. The results of response frequency are revealed in Fig. 3D. The response frequency reaches a saturated value when the embedded structure is close to the surface (<2 mm) of PDMS. A difference in response frequency exists for deeply embedded structures up to 3 mm, indicating the ability of subcutaneous diagnosis. Figure S31 presents the detailed frequency domain signals (200 to 300 Hz) and spectrograms corresponding to Fig. 4 (C and D). We use arrows to label the harmonics of 60-Hz noise and unify the color bar of the spectrograms. As illustrated in previous research, the ERM actuator produces both vertical and horizontal waves of vibration (51). The higher the input power, the larger the amplitude. Therefore, to achieve a deeper detection limit, we use a larger ERM actuator (type 3 ANIRP, 1.4 cm, as illustrated in fig. S32A) powered by 4.5 V DC. We first calibrate the intrinsic vibration frequency of the ERM actuator, as shown in fig. S32C. Then, we tested the design on a similar skin model (CPE in PDMS, fig. S33A), of which the detection limit exceeds 9 mm (fig. S33B). Corresponding spectra and spectrograms are illustrated in fig. S34. This covers the typical tumor location of most skin tumors (59–62). Given larger ERM actuators and stronger power input, deeper detection is expected to be realized. To better analyze the performance of ANIRP, we developed another series of artificial tissues using PDMS (10:1, ~514 kPa, comparable to subcutaneous skin lesion) embedded Ecoflex 00-30 (~43 kPa, comparable to normal skin), as shown in fig. S33C. With type 1 ANIRP (1.0-cm ERM actuator, powered by 2.5 V DC), the depth limit of detection decreases to ~2 mm (figs. S33D and S35). With type 3 ANIRP (1.4-cm ERM actuator, powered by 4.5 V DC), a depth limit of detection of 4 mm can be realized (figs. S33E and S36). Figure 4 (E and F) shows the study of the edge effect. When moving the position of the 0.5-cm ERM actuator against the border of a plastic-embedded sample (2 mm in depth and 15:1 mixing ratio), a decrease in response frequency is observed, as shown in Fig. 3F. This is explained by a smaller area of the embedded sample in contact with the actuator, and a relatively lower contribution of high frequency. A steep drop occurs when two-thirds of the actuator is removed from the harder part of the sample. High spatial sensitivity is demonstrated in this experiment as the response frequency changes sharply when there are small variations in the location of the ERM actuator (<1 mm). This feature of ANIRP compensates for the actuator density in spatial resolution, as detection of lesions can still be achieved when the lesioned area is not fully covered by the actuator. Figure S37 presents the detailed frequency domain signals (0 to 300 Hz) and spectrograms corresponding to Fig. 4 (E and F). We use arrows to label the harmonics of 60-Hz noise and unify the color bar of the spectrograms.

Since both the embedded depth and the overlapping area affect the response frequency, we use a type 2 ANIRP (2 × 2 ERM actuators, powered by 3.0 V DC) to measure the response frequencies on four locations around embedded object. We use 15:1 PDMS with plastic object (15 mm in diameter and 2 mm in thickness) embedded at various depths (−0.5, −1, −1.5, −2, and −2.5 mm) as the artificial skin models with lesion. Multiple different relative locations (11 for each skin model) between the ANIRP and the artificial skin lesion are chosen (fig. S38A). We obtain a dataset (55 in total) relating response frequencies (f₁, f₂, f₃, f₄) and coordinates of embedded object (x, y, z), which is randomly divided into the training set (45 sets) and the test set (10 sets). Leave-one-out cross-validation (LOOCV) was used in the training set for comprehensive evaluation of the models. The root mean square errors (RMSEs) of the test sets are generally smaller than the LOOCV training sets, indicating negligible overfitting. The RMSE is used to evaluate the performance of coordinates prediction. Through machine learning, accurate prediction (0.79 to 1.50 mm for x-y prediction RMSE and 0.14 to 0.39 mm for z prediction RMSE) of embedded object location can be achieved (fig. S38B and table S5).

ANIRP for monitoring skin status

We further demonstrate the capability of skin status monitoring in Fig. 5. Figure 5A shows a series of experimental photos with the lower arm experiencing a relax-contract-relax cycle by lifting weights. The continuous monitoring of skin mechanics is presented in Fig. 5B, as an increase in response frequency is observed during muscle contraction. The signal during contract is not as stable as the relaxing phase, due to the tremor. Likewise, we use a PDMS skin model to demonstrate the skin status with different curvatures (Fig. 5C), and the resulting changes in response frequency (Fig. 5D). Figures S39 and S40 present the detailed frequency domain signals and spectrograms corresponding to Fig. 5. We also confirm that ANIRP is able to capture ultrashort muscle changes, including tremors (<0.1 s). This is because some of the subcutaneous muscles undergo fast contract-relax cycles during the tremor. We use type 1 ANIRP (powered by 3.0 V DC) to detect single tremor events on the right forearm of a volunteer. By slicing the signal into 1-s slices at a step of 50 ms, we are able to clearly distinguish and timestamp the tremor event, which is estimated to be <0.1 s. The corresponding spectrograms and spectra are illustrated in fig. S41.

Fig. 5. On-body demonstration of ANIRP.

(A and B) Continuous measurements (B) of ANIRP on a forearm (front side) during a relax-contract-relax cycle of lifting a heavy weight (A). (C and D) Continuous measurements (D) of ANIRP on an artificial skin during its outward-inward-outward bending (C). The transition timestamps of status are labeled by red dash lines. (E) Measured response frequencies and estimated modulus (using type 1 ANIRP, at 2.5 V) of normal skin from various regions of the body, including the bicep, forearm, face, and forehead (n = 5). (F) Measured response frequency and estimated modulus of porcine skin a various burning levels defined by burning time (n = 3).

Mitigating motion artifacts is essential to the performance of wearable sensors. Here, we conduct a comparison between the situation with and without motion artifacts during the measurement of real human skin. Motion artifacts are created by violently swinging the arm being tested. The tests are completed on the right forearm of the subject, by an ANIRP with a 1-cm ERM actuator operated at 2.5 V. Figure S42 (B and C) shows the 5-s raw signal before digital filters of two statuses: (i) staying still (B) and (ii) swinging (C). A 0.1-s slice of the raw signal is followed, to show the vibration at high frequency (>20 Hz). Figure S42 (D and F) are the low-frequency components after being processed by a 20-Hz low-pass filter (second Butterworth). This part of the signal is mainly contributed by artifacts instead of the vibration from the actuator. In fig. S42 where the arm is staying still, we can also observe a tiny fluctuation, due to the existence of pulses. Namely, we obtain the high-frequency components by applying a 20-Hz high-pass filter (second Butterworth), as illustrated in fig. S42 (E and G). The signal in fig. S42G (with motion artifacts) seems less stable in magnitude than that in fig. S42E (without motion artifacts), but still has the same frequency. Figure S43 shows the corresponding spectra and spectrograms of these experiments. As shown in fig. S43A, when motion artifacts exist, the response frequency (~122 Hz) is relatively weaker while the low-frequency noise (<5 Hz) is relatively higher. However, the SNR can be easily enhanced by applying digital filters. Apart from postprocessing, we may also enhance the SNR by increasing the light intensity of the LED-PD sensor, as discussed before.

As a noninvasive wearable, ANIRP is generally safe for wearing without leading to skin irritations or inflammation. However, the production of heat from LEDs or the ERM actuators may lead to safety concerns during the measurements, especially during extended monitoring of skin. Here, we validate the compatibility by measuring the temperature over time (fig. S44). A 2-min running of a 1-cm actuator powered by 2.5 to 3.5 V DC leads to an increase in temperature but not exceeding 45°C (fig. S44A), while a 0.5-cm actuator powered by 2.5 V DC does not exceed 45°C. When operated at 3.0 and 3.5 V, the 0.5-cm ERM actuator has a safe temperature within 20 s, which meets the requirement of minimum actuation time (fig. S44B). The IR LED also generates heat, as shown in fig. S44 (C to F). We avoid overheating of the LED by: (i) operating at a lower bias (1.2 to 1.3 V), (ii) using multiple LEDs (four LEDs, <1.3 V), and (iii) using shining mode instead of continuous lighting. Multiple LEDs greatly improve the signal quality at lower bias, as shown in fig. S45. Figure S46 shows equivalent circuit diagrams of the ANIRPs. The driver module performs moderately high-voltage actuation (between 2.0 and 3.0 V) through controlled sequences of pulse-width modulation. As the resistance changes due to potentiometer control, the driver provides proportional levels of actuation to the patch, allowing for precise and predictable voltage-to-vibration levels. The circuit design features the ATmega328PU microcontroller with actuators, and corresponding LEDs in parallel as well as load resistors, connected directly to pulse-width modulation on the controller. Individual actuators are driven and isolated for precise control over the sensing region. This architecture makes scaling the design of the ANIRP efficient and reasonable. The driver and patch interface via electrical connections between corresponding components, including pin-to-actuator, pin-to-LED, and lead-to-PD, ensuring precise signal transmission. We also design a Bluetooth low energy module based on ESP32C3 for the automation of data acquisition and wireless application (fig. S46C and table S6). The PD in the LED-PD sensor is coupled with a trans-impedance amplifier as analog signal amplification. The ESP32 unit receives the amplified signal while controlling the LED. To maintain a high SNR, electrical traces are spread out, reducing interference and optimizing performance. Figure 5E shows the measurement of different regions (bicep, forearm, face, and forehead) of normal skin by ANIRP. We use a type 1 ANIRP at 2.5 V in the measurements. The response frequency ranges from 80 to 120 Hz, which falls in the linear range of calibration. The response frequency varies due to the difference between skin properties and subcutaneous components (fat tissues, muscles, or bone). The response frequencies of healthy skin provide a standard reference for skin or subcutaneous disease diagnosis. Our results correspond to previously published research in trend, while difference exists owing to different calibration methods (1). Figure S47 shows the corresponding spectra and spectrograms of the measurements. Figure 5F shows the capability of ANIRP to measure skin burning damage. We use a piece of porcine skin in this experiment and burn a selected area (5 × 5 cm²) with a torch for a controlled time to create various levels of skin burning. Each stage is measured three times, as we observe a significant increase in the response frequency with the increase in burning level. The corresponding spectrum and spectrograms are provided in fig. S48 (B and C).

We usually use ANIRP in a sequential mapping mode, as shown in Fig. 2A, when ERM actuators are only powered individually. Real-time signal recording is demonstrated in movies S3 to S6. To analyze the cross-interference of ERM actuators when they are powered simultaneously, we designed the cross-talk analysis experiment as shown in fig. S49A. We use an artificial skin model of PDMS (15:1) with CPE partially embedded as a lesioned region in contrast to normal PDMS as the normal region. Two 1-cm ERM actuators are deployed on the normal (ERM-1) and lesioned (ERM-2) regions and powered at 2.5 V in this experiment. An LED-PD sensor is placed at an equal distance from each of the ERM actuators. Three recordings are carried out when: (i) only ERM-1 is on, (ii) only ERM-2 is on, and (iii) both ERM-1 and ERM-2 are on. In experiment (i), an average response frequency is detected as 151.1 Hz, while experiment (ii) results in 187.1 Hz. In experiment (iii), two response frequencies are observed (except for the harmonics of 60 Hz) simultaneously, as shown in fig. S50. Among the six recordings, the average of the lower response frequency is 150.9 Hz, while the average of the higher response frequency is 185.2 Hz. As illustrated in fig. S49B, the lower and higher response frequencies (corresponding to normal and lesioned regions) without interference [in black, experiments (i) and (ii)] and with interference [in red, experiment (iii)] are compared. T tests are carried out and result in being not significant for both normal and lesioned groups (n = 6 and α = 0.05 for both groups; P = 0.9291 for the normal group and P = 0.2484 for the lesioned group). It is proved that the vibration from different sources is independent and produces a linearly combined mixed vibration signal that can be easily analyzed through FFT. This feature lays a foundation for another working mode of ANIRP apart from sequential mapping, which allows for a faster analysis of large areas. Figure S51 shows the simultaneous detection mode of ANIRP when multiple ERM actuators can be powered at the same time. When skin lesions exist, the response frequency of the lesioned region can be detected in the mixed signal, which usually differs from the normal response frequency. The actuation region can therefore be reduced by half at each stage to identify the precise location of the abnormal area. We also demonstrate the sequential mapping of a 4 × 4 array of locations on a PDMS sample with CPE embedded (2 mm in depth), as illustrated in fig. S52.

The development of the ANIRP marks a notable advancement in dermatological diagnostics, providing a noninvasive, easily operated, and real-time method for measuring the biomechanical properties of skin. By integrating multiple actuators with NIRS sensors, the ANIRP offers a comprehensive tool for rapid and efficient diagnosis and monitoring of skin conditions, facilitating a better patient experience and improving the quality of life. Compared with other EMM sensors, ANIRP measures skin mechanics under an unconventional mechanism based on frequency damping. This results in a much simpler yet robust sensor design, making the ANIRP more accessible and practical for widespread clinical use. Using ERM actuators instead of piezoelectric transducers or LRA actuators, the fabrication of ANIRP is easier under a low-cost fabrication protocol. This not only reduces production costs but also ensures a high level of reliability and consistency in the device’s performance. The ANIRP’s ability to map skin elasticity and other mechanical properties enhances our understanding of skin disorders, aiding in early detection and continuous health monitoring. This enables health care providers to intervene promptly and effectively at an early stage of disease progression, benefiting patients through improved diagnostic accuracy and personalized treatment plans. Future enhancements of ANIRP will focus on increasing diagnostic accuracy and refining device sensitivity, aiming to better meet the challenges of clinical diagnostics. Integrating ERM actuators with higher mechanical output and vibration amplitude enables ANIRP to stimulate deeper subcutaneous tissue, thus substantially enhancing the detection limit of ANIRP. Using ERM actuators with higher rotation frequency makes it feasible to obtain better continuous monitoring of subcutaneous muscle tension, as the temporal resolution is determined by the response and intrinsic frequency. In general, this innovative approach combines mechanical and optical diagnostics, showcasing an advanced method in the management and treatment of skin diseases.

MATERIALS AND METHODS

All procedures were conducted in accordance with the National Institutes of Health (NIH) Guide and with the approval of the Office of Human Research Ethics at the University of North Carolina at Chapel Hill, under the protocol no. 22-0163. Grubb’s test was used to exclude the outliers in data analysis, and t test was used to analyze the difference between groups. The sample size of experiments was three to nine, with experiments repeated technically or biologically (stated in detail in Discussions).

The fabrication of PDMS artificial skins

Normal artificial skin

We used commercially available PDMS to prepare artificial skin models for its similarity to skin in viscoelasticity. The main component and the curing agent of silicone elastomer (Sylgard 184, DOW) are mixed under a ratio from 40:1 to 5:1 (w/w). Silicone-compatible pigments (Silc Pig, Smooth-on) are added into the mixture under a ratio of 1:300 (w/w), to better simulate the UV-visible (UV-vis) absorption of skin and tissues. The mixture was then thoroughly mixed and degassed under vacuum for ~10 min and allowed to stay in an oven of 60°C for 18 hours until fully cured.

Artificial skin with embedded objects

We used fused deposition modeling 3D printing (Ultimaker S3) to customize the hard plastic objects. The thickness of 3D-printed CPE is controlled at 1 mm (or other desired thickness, as specified). Some tiny spacers are printed together at the edges of the CPE objects to control the depth of the embedded objects. Then, the spacers of the objects are attached to the bottom of the petri dish by scotch tape.

We used commercially available PDMS to prepare artificial skin models for its similarity to skin in viscoelasticity. The main component and the curing agent of silicone elastomer (Sylgard 184, DOW) are mixed under a certain ratio [15:1 or 10:1 (w/w), as specified]. Silicone-compatible pigments (Silc Pig, Smooth-on) are added into the mixture under a ratio of 1:300 (w/w), to better simulate the UV-vis absorption of skin and tissues. The mixture was then thoroughly mixed.

We poured the mixture into the prepared petri dish with the plastic objects and degassed them under vacuum for ~10 min. Then, they were allowed to stay in an oven of 60°C for 18 hours until fully cured. We carefully broke the petri dish and extracted the artificial skin model with objects embedded.

The fabrication of ANIRP prototypes

A commercially available Pyralux Kapton soft PCB material made of PI sandwiched between copper was double-sided sprayed on and coated with masking paint (Krylon). The paint mask was then partially removed using an IR laser ablation system to expose unwanted copper, which was removed by etching in ferric chloride solution (MG Chemicals 415) for 15 min. The top-side pattern for ERM actuators and the back-side pattern for LED-PD sensors were ensured to be well aligned during the laser ablation and connected by through-holes. The soft PCB was then rinsed with water and acetone to remove the remaining etchant and paint mask. Surface mount electrical components (LEDs, PDs, and the FPC connector) were soldered using solder paste and hot air guns. The structure of the soft PCB was tailored by UV laser ablation to introduce hollow structures for conformal deployment. The ERM actuators (1.0 or 0.5 cm) are attached to the designated locations on the top side of the soft PCB and soldered on. The full list of electronic components is listed in tables S2 and S6.

The measurement of response frequency using ANIRP

An ANIRP was deployed onto an artificial skin model or real skin with an FPC cable connected. After a basic cleaning of the skin area, the ANIRP was mounted to the skin by covering a piece of medical dressing (Tegaderm, 3M) for uniform and conformal device placement. The FPC cable was connected to the driver module. The IR LEDs were provided with a DC bias of 1.2 to 1.4 V. The actuators were powered in a certain order that can be preprogrammed (notes S2 and S3). The ground (GND) and output channels of the PDs were connected to the analog-digital converter (ADC) of Intan RHD, where the data were amplified and recorded.

Hair effect analysis

A type 1 ANIRP was deployed on the volunteer’s right forearm with a piece of medical dressing (Tegaderm, 3M). The location of deployment was around the surface of the brachioradialis muscle, which is the border of hairy and hairless skin. The distance between the two locations for measurement was ~1 cm to minimize the systematic difference in skin elasticity and subcutaneous tissues. The IR LEDs were provided with a DC bias of 1.4 V. The actuator was powered by 2.5 V DC. The GND and output channels of the PD were connected to the ADC of the Intan RHD, where the data were amplified and recorded. Each location was repeatedly measured five times.

Porcine skin burning analysis

We first labeled the burning area on the porcine skin (~5 mm in thickness) as a 5 cm × 5 cm square. We used moderate flame to evenly burn the surface of the selected area for a controlled time of 0 to 5 min. A type 1 ANIRP was deployed onto the burned area with a piece of medical dressing (Tegaderm, 3M). The IR LEDs were provided with a DC bias of 1.4 V. The actuator was powered by 2.5 V DC. The GND and output channels of the PD were connected to the ADC of the Intan RHD, where the data were amplified and recorded. Each burning level was repeatedly measured three times.

Motion artifact analysis

A type 1 ANIRP was deployed on the volunteer’s right forearm with a piece of medical dressing (Tegaderm, 3M). The IR LEDs were provided with a DC bias of 1.4 V. The actuator was powered by 2.5 V DC. The GND and output channels of the PD were connected to the ADC of the Intan RHD, where the data were amplified and recorded. The vibration response signal was first recorded when the arm was staying still, then recorded when the arm was periodically swinging without changing the muscle status. The ANIRP was not removed until both signals (staying still and swinging) were recorded.

The optical simulation by MCX lab

The geometric optics simulation was carried out by MCX Lab in Python (pmcx). First, we defined a three-layer skin model, including epidermis (0.1 mm in thickness), dermis (1 mm in thickness), and subcutaneous muscle (15.9 mm in thickness). The corresponding optical parameters of 900-nm NIR light (scattering coefficient, absorption coefficient, anisotropy factor, and refractive index) of each layer are provided in table S1 (63). The data provided here were converted into mm⁻¹. For balancing resolution and computation workload, the voxel length is defined as 0.1 mm. We carried out the simulation in a space of 50 mm × 50 mm × 25 mm (x ∙ y ∙ z), while the three-layer model was centered in the space (10 to 40 mm, 10 to 40 mm, and 3 to 17 mm). We then defined the light source as zenith Gaussian (zGaussian), which has a Gaussian distribution of light intensity by the zenith angle, to simulate the angular light intensity distribution of an LED. The x and y coordinates of the light source are both 25 mm. We defined two separate detectors (detection radius, 1 mm) on both sides of the light source symmetrically. The simulation started from 0, with a duration of 5 ns.

Flux graph

We used 10⁷ photons in this simulation, with the light source placed 0.5 mm away from the surface of the epidermis. The flux graphs (Fig. 3B) were sliced along the x direction starting from the center of the model to the edge, every 3 mm. The color bar was unified with contrast adjusted.

Sensor-skin distance simulation

We used 10⁶ photons in this simulation, with the light source placed at a varying distance from the surface of the epidermis (0 to 2 mm, every 0.1 mm). The two detectors are placed at the same z coordinate with the light source, to keep the same distance from the skin, with x and y coordinates to be (24 mm, 24 mm) and (26 mm, 26 mm). The received photons are counted by different sensor-skin distances, as shown in fig. S16E. The simulation was repeated three times with different seed values.

LED-PD separation simulation

We used 10⁶ photons in this simulation, with the light source placed 0.5 mm away from the surface of the epidermis. The two detectors are placed at the same z coordinate as the light source, to keep the same distance from the skin. However, they had varying x and y coordinates, (25 mm − distance, 25 mm − distance) and (25 mm + distance, 25 mm + distance), to simulate the separation of LED and PD. The distance value ranges from 1 to 5 mm every 0.1 mm. The received photons are counted by different LED-PD separations, as shown in fig. S16F. The simulation was repeated three times with different seed values.

Light intensity simulation

We used varying photon numbers in this simulation, with the light source placed 0.5 mm away from the surface of the epidermis. The two detectors are placed at the same z coordinate with the light source, to keep the same distance from the skin, with x and y coordinates to be (24 mm, 24 mm) and (26 mm, 26 mm). The photon number varied from 10⁴ to 10⁵ every 10⁴, 10⁵ to 10⁶ every 10⁵, and 10⁶ to 5 × 10⁶ every 10⁶. The received photons are counted by different sensor-skin distances, as shown in fig. S16G. The simulation was repeated three times with different seed values.

Supplementary Materials

The PDF file includes:

Notes S1 to S3

Tables S1 to S6

Figs. S1 to S52

Legends for movies S1 to S6

Other Supplementary Material for this manuscript includes the following:

Movies S1 to S6