Introduction

Motion signals of human body carry various kinds of information on the physiological processes, which are closely relevant to human body conditions such as movement state, respiration, pulse, cardiovascular state, gesture language and symptom1,2,3,4,5,6,7. In this respect, reliable tools for convenient signal acquisition and efficient data processing are significant and highly desired. Nowadays, advanced technologies based on flexible electronics and intelligent medical devices allow thorough human body motion monitoring by offering numerous portable, flexible, and reliable methods8,9. In this case, continuous daily healthcare for both patients and normal people can be implemented by real-time signal interpreting and online transmission, which can greatly improve their safety and life quality10,11. In particular, novel skin-interface devices such as E-skins, textile sensors, and hydrogel sensors have attracted widespread attention due to their potential for applications in human body motion monitoring12, disease treatment13, drug delivery14, blood glucose monitoring15, and blood pressure analysis16. For instance, in the past years, various attempts have been made to design signal collectors in skin-interface devices for long-term body motion monitoring during natural daily activities, which would combine lightweight, flexibility, high permeability and imperceptibility17,18,19,20,21,22,23,24,25,26,27,28,29. The recent emergence of wearable motion sensors based on liquid metals30, hydrogels31, graphene aerogels32, functional fibers33, and piezoelectric materials34,35 have enabled flexible devices to be applied in body motion monitoring and short-term healthcare. However, strict requirement of processing environmental, relatively large weight along with low comfort level and inevitable partial rigid design still restricts their daily use.

Meanwhile, security is a precondition for large-scale application of skin-contact wearable electronics. Moisture permeability and gas permeability of the device dominantly decide its long-term wearing comfort level, as well as the biocompatibility25,30. Obstructing local sweat evaporation and air convection for a long time may cause skin inflammation. In some special cases such as high humidity, short circuit caused by sweat and water vapor accumulation leads to electronic component failure and, consequently, to the damage of the sensor as a whole24. Besides, the release of silver ions should be avoided in wearable products (i.e., garments, accessories, and portable electrical devices) because of the possible problems associated with anaphylaxis or biological toxicity36. Although there are a lot of advanced packaging technologies for flexible devices, the issue of optimizing the competition between moisture/gas permeabilities and encapsulation still remains unsolved.

Another challenge in relation to wearable sensors is the signal classification capability. In general, sensors subjected to stretching or bending are able to accurately collect specific signals within a wide range of human body motions (e.g., finger, elbow, or knee bending)23,30. Those signals, due to the significant difference in their amplitude and frequency, can be effectively classified via traditional approaches such as discrete Fourier transform (DFT) method. However, accurate classification for similar-movement related responses or slight action has not been achieved yet. In recent years, artificial intelligence (AI) (typically, machine learning) is being increasingly applied to various fields, including optical recognition37, object classification32,38, and tactus technology in robotics39,40,41. By virtue of high-dimensional feature extraction capability, machine learning is competent to classify the signals that appear to be irregular. Therefore, combining machine learning with advanced wearable sensors is of great significance for highly efficient and accurate human motion tracking as well as for the development of next-generation AI daily healthcare systems.

In this study, we propose an ultralight, flexible, and biocompatible all-fiber motion sensor (AFMS) for wearable electronics such as human motion monitoring with high recognition rate. The functional material of the sensor was presented by radial anisotropic porous silver fibers (RAPSF) fabricated via a modified blow-spinning technique with environmental control strategy. In the proposed method, the phase separation was achieved during jet flow process of precursor solution via the adjustment of compressed air temperature and flow rate. Unilateral pre-nucleation of the collected precursor fiber (AgNO3) film was provided by adjusting the temperature of the collector. The RAPSF with a single fiber conductivity of ~7.96 × 103 S m−1 and an adjustable film resistance from 80 to 213 Ω were obtained after 3 h of subsequent UV-radiation., The as-prepared RAPSFs were afterwards stacked through a layer-by-layer assembly strategy into a flexible, multilayer AFMS for human motion tracking. The sensor exhibited an ultralight weight (68.7 mg cm−3), elevated moisture permeability (59.7 g m−2 h−1), high gas permeability (low gas resistance of 9.8 × 10−3 Pa at the flow rate of 31.6 L min−1), good biocompatibility, and excellent motion signal identification performance in elbow, knee, and finger movements. Moreover, further processing of a sufficient amount of data using machine learning model demonstrated the possibility to use the AFMS as an AI sensor for identification of different throat motion with the accuracy above 85%, demonstrating the great potential of AFMS as an ultralight, flexible, and imperceptible device for wearable electronics and next-generation intelligent medical technology.

Results

Preparation process of RAPSFs

The RAPSFs were produced using a modified blow-spinning system, including a liquid-gas two-channel six-needle module, a compressed air temperature regulator, a temperature-controlled fiber collector, and subsequent UV radiation (Fig. 1a). The final product had the form of a specific-sized RAPSF film with silver metallic luster (see Fig. 1b). The proper temperature of fiber collector (250 °C) was employed to accelerate the reduction reaction from Ag+ to metallic silver, as well as to speed up the crystal growth process. The precursor of RAPSF contained two solvents (dichloromethane and acetonitrile) with different boiling points and evaporation rates. The environmental control apparatus enabled, the compressed air to be heated to 60 °C in the course of blow spinning, which was higher than the boiling point of dichloromethane (Fig. 1c). Therefore, rapid evaporation of dichloromethane occurred during the fiber jet process and led to the emergence of surface voids on the final product (Fig. 1d, Supplementary Movie 1). In this work, different concentrations of dichloromethane (0, 2, 5, and 10 wt.%) were applied to reach various surface states of fibers, as well as to elucidate the effect of dichloromethane on their electrical properties (Supplementary Fig. 1). It is noteworthy that the precursor solution cannot be spun when the concentration of dichloromethane is over 12 wt.%.

Fig. 1: Preparation and micromorphology of RAPSF.
figure 1

a A modified blow-spinning system with environmental control accessories, including temperature controlled compressed air regulator and fiber collector. b Photograph of as-prepared RAPSF with silver metallic luster. c Schematic diagram of the blow-spinning and phase separation processes. d Photograph of jet flow at the needle tip during blow spinning, captured by a high-speed camera. e False-color SEM image of a single RAPSF (f) and its zoom-in image. g Cross-sectional SEM image of RAPSF, evidencing the porous structure within the polymer skeleton. h XRD spectra of as-prepared RAPSF and reference pattern (JCPDS file no. 04-0783, silver). i XPS spectrum differentiation imitating analysis of the Ag3d peak of RAPSF. j BET results for RAPSF and SF.

Full size image

Micromorphological characteristics

According to the scanning electron microscope (SEM) results, the as-prepared RAPSF film exhibited a network structure at the micro-scale (Supplementary Fig. 2), which was similar to that of silver fibers (SF)39. As shown in Fig. 1e, the single RAPSF fiber had a porous surface structure. The false-color SEM images revealed different particle structures in two sides of the fiber, including small granular islands on the top (away from the substrate (shown with green color), and a continuous film structure on the bottom (near to the substrate (highlighted with pink). The formation of the former is codetermined by the phase separation during blow spinning and the partial crystallization of silver particles under UV irradiation. The electrical properties of islands are unfixed and can be directly adjusted by changing the fiber shape, which further influences the silver particle distribution density through the surface. On the contrary, the formation of the continuous structure on the near-substrate side of the fiber is due to the effect of the high temperature substrate, being conducive to the early crystallization and the continuous silver grain growth. The tight intergranular connection allows such a continuous structure to remain electrically stable while changing the fiber shape. A zoom-in SEM image of the RAPSF in Fig. 1f depicts a secondary porous structure of the fiber skeleton (shown with a red arrow). Except for the skeleton surface, the cross-sectional SEM image of RAPSFs discloses a porous structure inside the fiber skeleton (Fig. 1g), indicating that the phase separation has occurred on both the surface and inside of precursor fibers. As mentioned above, silver particles on the RAPSF surface acted as the electrical functional components. According to the crystallographic analysis of the top and bottom sides of the RAPSF (Fig. 1h, i, and Supplementary Fig. 3), the reduction reaction of Ag+ took place on both these sides and metallic silver particles formed. Furthermore, silver particles allowed the polyvinyl pyrrolidone (PVP) skeleton to retain its structural stability (Supplementary Figs. 4, 5)42.

The information about porosity, obtained from the BET tests, reveals an increase in the pore and surface areas of the RAPSF with a dichloromethane concentration of 5 wt.% relative to those of the SF (Fig. 1j). Particularly, the pore area in the RAPSF for the pore diameter lower than 10 μm is twice higher than that of the SF. As follows from the SEM images (Supplementary Fig. 6), the use of different concentrations of dichloromethane results in various porous structures of the final fibers (in terms of pore scale and distribution). The surface porous and secondary porous structures (Supplementary Fig. 7) codetermine the ultralight weight of as-prepared RAPSF with an ultra-low density of 68.7 mg cm−3 and an overall weight of 7.95 mg, which is of significant importance for portable and wearable electronics.

Electrical properties

The conductivity of a single RAPSF in flat state was mainly determined by the presence of silver particles across the fiber surface and was measured to be 7.96 × 103 S m−1. The average path impedance of the probe-RAPSF-probe system was confirmed repeatedly as about 2.1 kΩ, reflecting a stable resistive characteristic (Supplementary Figs. 8, 9). Once the fiber is bent to different radius of curvature, its conductivity shows regular changes (Fig. 2a). A similar situation was observed in the RAPSF film, where, however, the change rate of the conductivity was higher than that in a single fiber (Fig. 2b). Besides, the ability to adjust the resistance of the RAPSF film (15 mm×10 mm×50 μm) between 80 and 213 Ω by varying the bend radius was shown in Fig. 2c. Wherein, bending radius from 12 to 1 mm can be considered as the linear range for stably monitoring of the RAPSF (Fig. 2c inset) with a numerical relation as:

$$R = - 8.73 \ast r + 214.78$$
(1)

where R is the resistance of RAPSF, r is the bending radius, the R-square of the fitting result is about 0.991. The finite element analysis (FEA) was further employed to investigate the fiber state during bending. As shown in Fig. 2d, the stress concentration occurred in the midpiece of the single fiber during bending rather than the uniform strength on the fiber surface. Such a stress concentration led to a local density variation of silver particles on the top lateral of the fiber exposed to bending, leading to a change in conductive path (Fig. 2e, orange line). Assuming that bending has occurred in the x-z plane, the cross-section in the y-direction along the fiber axis can be considered as the neutral surface, in which the stress remains unchanged (Fig. 2e, red dashed line). In this case, the top lateral is stretched and the bottom lateral is compressed. Because of the continuous silver particle structure on the bottom lateral, the change in resistance of the RAPSF during bending can be attributed to the density variation and junction mismatch of silver particles on the top lateral. This can be further proved by local surface stretch bending simulation via FEA. Figure 2f, g show the FEA simulation results revealing the unlinked particles and stress concentration at the cracks even at a small degree of fiber stretching. To benefit from the inherent flexibility of the polymer fiber skeleton, the resistance of the RAPSF can be restored to its plane state level after the bending is released (Fig. 2h). The existence of the flexible polymer fiber skeleton ensures that the cracks will not form on the fiber surface when it is bent. Thus, during the durability testing, the RAPSF exhibited remarkable electrical and mechanical stabilities, being able to endure up to 10,000 cycles of bending under the bending radius of 8 mm.

Fig. 2: Electrical performance and FEA of RAPSF.
figure 2

I/V testing results of a single RAPSF and b RAPSF film with different bending radius on the PET substrates. c Resistance of the RAPSF film at different bending radius and the film recovery. The inset shows linear range of resistance for RAPSF film from 12 to 1 mm bending radius. d FEA stress simulation of single fiber for stress analysis in bending state. e Schematic diagram of conductivity of RAPSF with different radius of curvature; the red dashed lines indicate the neutral surface and the orange lines indicate the electron transport paths. f FEA stress simulation of single RAPSF surface, revealing the junction mismatch between silver particles and g corresponding stress along the horizontal central axis. h SEM image of RAPSF film exposed to bending. i Cyclic bending testing of RAPSF film with a bend radius of 8 mm.

Full size image

To assess the functional characteristics of the produced layers in practical use, the as-prepared RAPSFs were layer-by-layer assembled into an AFMS as illustrated in Fig. 3a. First, the polyacrylonitrile (PAN) substrate was fabricated by blow spinning with a roller collector. The PAN fiber film was peeled off from the collector after one hour of drying under an infrared light. Second, a silver nanowires (AgNWs)/PVP/alcohol dispersion was locally applied onto the PAN substrate via dip coating, forming AgNWs electrodes (Supplementary Fig. 10, 11). After that, the as-prepared RAPSF was hot-pressed (60 °C, 750 Pa) on the PAN substrate and the obtained top-layer PAN film was locally spray-coated with polyvinyl butyral (PVB) alcohol dispersion as the adhesive. Benefiting from the low fiber-space ratio and the low intrinsic bulk density of the fiber mesh, the AFMS exhibited an ultra-low density of 68.7 mg cm−3, allowing it to be placed on the surface of setaria viridis (Fig. 3b, Supplementary Movie 2).

Fig. 3: Fabrication and characteristics of the AFMS.
figure 3

a Schematic of the fabrication process of the AFMS. b Photographs of the AFMS on setaria viridis, revealing its ultralight weight. c SEM images of the PAN fiber film substrate and d AgNWs/PAN structure. e Moisture permeability (red) and gas permeability (blue) for AFMS, nickel foam, PI film, PET film, and PDMS film. f ICP-MS results on silver ion release detection during 24 h of magnetic stirring. g Implanted biocompatibility tests of the AFMSs implanted into the subcutis of mice (C57BL/6) for 1, 5, 10, and 15 days, and the control group (non-implantation).

Full size image

Biocompatibility and stability

Breathability is one of the most important characteristics of skin-contact electronics because of their direct impacts on safety and comfort level. In this study, all the blow-spun RAPSF, the PAN film, and the dip-coated AgNWs were found to have the low volume fractions, which were attributed to their similar network structures (Fig. 3c, d, and Supplementary Fig. 12). The network structures for different layers codetermined the high gas permeability (the low gas resistance of 9.8 × 10−3 Pa at the flow rate of 31.6 L min−1) and the elevated moisture permeability of 59.7 g m−2 h−1 (1432.8 g m−2 in a day) for the sensor (Fig. 3e, Supplementary Fig. 13), which was comparable to those of a porous nickel foam and obviously better than those of other flexible substrates such as polyimide (PI), polyethylene terephthalate (PET), and polydimethylsiloxane (PDMS) (Supplementary Fig. 14). The leakage of metal ions is also a serious biosecurity issue for wearable electronics. In order to measure the ion release level (mainly for Ag-ions), the AFMS was dipped in normal saline and exposed to magnetic stirring at a rotational speed of 600 rpm. After 24 h of stirring, inductively coupled plasma mass spectrometry (ICP-MS) results of the final normal saline revealed a low Ag-ion concentration of 0.1359 mg L−1, indicating the good bonding between the RAPSF and the PAN fiber film substrates (Fig. 3f, and Supplementary Fig. 15)25. This means that the AFMS is able to retain its structural stability even when subjected to moving or sweating. The accelerated oxidation experiment was afterwards carried out to measure the stability and inoxidizability of the AFMS as well as the all-fiber sensor in the air at 80 °C (Supplementary Figs. 16, 17). The resistance change rate (ΔR/R) was found to increase by 0.9% only after 96 h of testing, which was extremely low for nanoscale metallic materials and indicated remarkable stability and inoxidizability of the AFMS.

In order to evaluate the biosecurity of the sensor, in vivo biocompatibility tests were performed on 5 mice (C57BL/6) (Supplementary Fig. 18). According to Fig. 3g, neither osteoporosis nor edema was observed in the epidermal cells on the first day of experiments. Besides, there was also no obvious thickening of the epidermis and the dermis in the collagen fibers was sufficiently thick. However, the subcutaneous tissue was infiltrated with inflammatory cells, indicating that there was an inflammatory response caused by foreign body invasion. After 5 days, the overall structure of the skin exhibited only a slight abnormality and the inflammatory cells in the hypodermis were significantly reduced. After 10 days, the overall structure of the skin was basically restored, the collagen fibers in the dermis were loosely arranged, and no inflammatory cells were observed in the subcutaneous tissue. In this case, the overall structure of the skin was similar to that of the control group (non-implantation) and entered the inflammation-negative range. Except for restorable local inflammation caused by the rejection reaction, no irreversible tissue necrosis occurred within 15 days after in vivo biocompatibility tests, indicating a good biocompatibility of such an AFMS.

Motion monitoring and AI signal classification

The lightweight, flexibility, mechanical stability, and biocompatibility of the AFMS joint guarantee the feasibility of its practical application. As a proof of concept, we built a motion monitoring system based on the AFMS and an electrical experimental platform. This system is able to track the sensor resistance changes in real time, reflecting the actual motion situation, including the movement of fingers, elbow, knee, and throat (Fig. 4a), within a short response time (45 ms) and short recovery time (62 ms) (Fig. 4b). The AFMS was first applied to monitor the motions of fingers, elbow, and knee, and the results showed the regular changes in the electrical signals caused by sensor bending (Fig. 4c, Supplementary Figs. 19, 20, and Supplementary Movie 3). The signal amplitude was determined by the degree of sensor bending, and the signal frequency of signal corresponded to the motion frequency. In this case, different kinds of electrical signals could be easily classified into corresponding motions via DFT analysis (Supplementary Figs. 21–23). In addition, a five-channel motion monitoring system was built to recognize simple gestures. In such a system, five AFMSs were respectively pasted on five fingers. As shown in Fig. 4d, gestures, referred to as “good”, “yes”, and “OK”, were clearly identified depending on the resistance-bend relationship of the sensor. These data show great potential of the designed AFMS for real-time tracking of human body motion.

Fig. 4: AFMS for motion monitoring and AI wearable electronics.
figure 4

a Schematic diagram of wearable application scenarios of the AFMS. b Response time and recovery time of the AFMS. c Current-time signals collected by the AFMS attached on the finger surface during bending. d Gestures recognition through the resistance change rate by the AFMSs attached on the five finger surfaces. e Five types of throat motion responses including speaking (“Goodbye” and “Hello”), swallowing- and coughing-related signals. f Model training and identification process, wherein the pre-collected data are divided into a training set and a testing set with respect to a ratio of 8:2 for 5-fold cross validation. g Confusion matrix of identification for throat motion signals. h Classification results of a 58 s continuous signal containing four types of throat motions.

Full size image

On the other hand, with the continuous improvement of medicine science, coughing detection is increasingly becoming of extreme importance for monitoring infectious pneumonia, especially caused by SARS-CoV-2 variants known as Covid-1943. Small amplitude, similar frequency, complex phase relationships, indistinct features, and signals interference of throat motion co-determined the complexity of coughing detection. As a proof-of-concept, we used the AFMS to collect four types of throat motion signals, including two speaking signals (“Goodbye” and “Hello”), swallowing, and coughing (Fig. 4e). In this case, amplitudes and frequencies of the two speaking signals were too close to be classified via DFT analysis. In order to overcome these issues, machine learning method was employed. For comparison, three common machine learning models, including XGBoost (XGB), k-nearest neighbor (KNN), and support vector machine (SVM), were introduced to handle the classification for throat motions44. Particularly, a 5-fold cross validation was pre-performed before starting machine learning (Fig. 4f, Supplementary Fig. 24). The classification results revealed a good grade comprising few outliers, concentrated distribution, and high mean value, as well as high classification speed of XGB (Supplementary Figs. 25, 26, Supplementary Table 1). Particularly, the XGB achieved the highest accuracy and stayed steady when the frames length is set higher than 180 points (Supplementary Fig. 27). The confusion matrix of the identification task exhibited a high accuracy of the correction rate of this system (Fig. 4g). Finally, the trained XGB was used to classify a 58 s continuous mixed signal. Based on the results, this model allows one to effectively classify four types of the above throat motions and noise signals (Fig. 4h). The good recognition accuracy and classification abilities of the AFMS make it possible for spotting the early onset of the viral throat illnesses and for tracking the related symptoms.

Discussion

In this work, the RAPSFs were fabricated by a modified blow-spinning technology with environmental control strategy. A single RAPSF fiber exhibited a porous and radial anisotropic structure due to the phase separation and temperature-controlled grain growth process. The heterogeneous texture of silver particles on the RAPSF surface enabled the resistance of the film to be adjusted between 80 and 213 Ω depending on the deflected shape. Benefiting from the network structure and flexible polymer skeleton, the RAPSF film revealed the high mechanical stability and the ability to endure over 10,000 bending cycles. As a proof-of-concept for body motion and throat monitoring, an AFMS based on RAPSF, PAN, and AgNWs was manufactured via a layer-by-layer assembly strategy for body motion monitoring. Good performance of the AFMS as the sensor was demonstrated via classification of signals collected during the finger and elbow bending, the knee motion, and different gestures. Moreover, the analysis of a sufficient amount of data collected using the AFMS revealed a high identification accuracy of the latter (more than 85%) as an AI throat motion sensor for tracking different throat motions. The ultralight weight, unique flexibility, good gas/moisture permeability, comfortable wearing, high biocompatibility, and signal classification superiority of the designed AFMS are conducive to the potential application of RAPSF-based all-fiber sensors in flexible electronics and wearable AI healthcare.

Methods

Materials

Silver nitrate (AgNO3, 99.8% purity) was purchased from Beijing Chemical Works. Polyvinyl pyrrolidones (PVP) with Mw of 360,000 and 1,300,000) were supplied by Alfa Aesar. Acetonitrile (99.0% purity), polyvinyl butyral (PVB, Mw = 170,000), polyacrylonitrile (PAN, Mw = 150,000), and dimethyl sulfoxide (DMSO, 99.8%) were purchased from Shanghai Aladdin Bio-Chem Technology Co., LTD. Cupric chloride dihydrate (CuCl2·H2O, 98%), ethylene glycol (EG, 99%), and dichloromethane (CH2Cl2, 99.5%) were provided by Macklin Biochemical Co., Ltd. Fluorocarbon surfactant (FS-3100) was purchased from DuPont Research & Development and Management Co., Ltd. The copper tape electrode was commercially available. All reagents were used without further purification.

Characterization

The transmitted infrared (IRT) images were captured by a thermal imager (DT-980, CEM, China). The X-ray powder diffraction data were collected using an XRD diffractometer (D/max 2500, Rigaku, Japan) in Cu Kα radiation (λ = 1.54178 Å). The micromorphological analysis was performed by means of a field emission scanning electron microscope (FE-SEM, LEO-1530, Zeiss, Germany) with a EDS attachment module. An X-ray photoelectron spectrometer (ThermoFischer, ESCALAB Xi +, America) equipped with an Al Kα radiation source (hν = 1486.6 eV) was employed to acquire the surface element information. Thermogravimetric analysis (TGA) and differential thermal analysis (DTA) were performed using a thermogravimetric analyzer (STA 449 F3, Jupiter, Germany). The electrical properties of as-prepared samples were examined on an IV/CV probe installation (PRECISION SYSTEMS INDUSTRIAL LIMITED, HK, China), including an EPS4 probe station, a PCA1000 LCR instrument, and PPTS-CV IV/CV test software. The IV/CV probe test system was linked to an electrochemical workstation (Metrohm Multi Autlab M204, Switzerland) for accurate data recording. The hydrophobicity of as-prepared samples was measured via an OCA15 Pro video-based optical contact-angle measuring system (OCA15Pro, Dataphysics, Germany). The mechanical properties were assessed using a universal testing machine, combined with a PC-controlled electrochemical workstation (CHI 630E, CH Instrument, China) in a two-electrode system. The gas permeability was measured with an automatic filter material tester (CERTITEST 8130 A, TSI, USA). The ion concentration in water was evaluated using an inductive coupled plasma emission spectrometer (Agilent 5110 ICP-OES, USA).

Fabrication of RAPSF

In a typical process, 0.25 g of PVP were first dispersed in 2.10 g of acetonitrile, and stirred for 5 h. After that, 1.50 g of AgNO3 and 50 μL of FS-3100 were added into the solution and stirred for more than an hour until completely dissolved. In order to achieve the phase separation, different contents of dichloromethane (0, 2, 5, and 10 wt.%) were added into the solution and thoroughly stirred for an hour to obtain a brown blow-spinning precursor solution. To perform blow-spinning, the precursor solution was jetted out to a temperature-controlled metal base (250 °C) with an injection speed of 1.5 mL h−1. The compressed air was heated to 80 °C by a temperature control accessory. In this process, the airflow velocity was set to a value of 3.0 m3 h−1. The photo-reduction of Ag+ ions was achieved after 3 h of UV irradiation from simultaneously two UV lamps (250 W, Philips Lighting).

Fabrication of PAN fiber substrates

First, 0.14 g of PAN (Mw = 150,000) were dispersed in 100.00 g of DMSO and stirred for 2 h at 70 °C. To perform blow spinning, the solution was injected into an integrated blow-spinning apparatus with a liquid-gas two-channel six-needle module and a rotated fiber collector. During blow spinning, the airflow velocity was set to 4.5 m3 h−1, and the injection speed of the solution was kept at a value of 2.0 mL h−1.

Synthesis of silver nanowires

AgNWs were produced according to a previously described method45. In brief, 0.8 g of PVP (Mw = 360,000) were first dissolved in 100 mL of EG and stirred for 2 h. After that, 1.0 g of AgNO3 was added into the solution and stirred for an hour until thoroughly dissolved. Then, 1.6 mL of as-prepared CuCl2·2H2O/EG solution (3.3 mM) were rapidly injected into the mixture and mildly stirred for an hour. The mixture solution was afterwards transferred in a preheated silicone oil bath and exposed to heating at 130 °C for 3 h. To achieve the product separation, the solution was alternately cleaned with acetone and ethanol for three times during 10 min of centrifugation at a rotational speed of 3000 rpm. The separated AgNWs (with radius of 25 nm and lengths of 200 μm) were dispersed in a PVP/ethanol solution (0.5 wt.%), forming a 100 mg/g AgNWs/ethanol mixture solution for further use.

Committee that granted approval

Animal studies are approved by Animal Management Committee of Chongqing Science and Technology Commission.

Electrical tests

The electrical tests of the RAPSFs were performed using a probe station and an electrochemical workstation (Supplementary Fig. 8). Prior to testing, the single RAPSF and RAPSF films were unilaterally fixed on a flexible PET substrate.

Signal classification and machine learning mode

In this work, three machine learning models were used for signals classification, including XGB, KNN and SVM (for further details and principles, please refer to Supplementary Information).