First, the poly(L-lactic acid)-poly(trimethylene carbonate) (PLLA-PTMC) films were prepared via a solvent casting method. PLLA-PTMC (60:40 mass ratio, Jinan Daigang Biomaterial Co., Ltd., China) was dissolved in trichloromethane (AR, Beijing Tongguang Chemical Co., Ltd., China) at a mass-to-volume (w/v) ratio of 1:10. The mixture was stirred overnight to ensure complete dissolution. A 4 mL aliquot of the solution was uniformly cast onto a smooth glass substrate and cured at 4 °C for 12 h, yielding a flat PLLA-PTMC film (thickness ~120 μm) as the substrate of the device. Next, the oriented PCL fibers were fabricated via electrospinning. PCL (molecular weight ~80,000, Shanghai Yuanye Bio-Technology Co., Ltd., China) was dissolved in hexafluoroisopropanol (AR, Shanghai Aladdin Biochemical Technology Co., Ltd., China) at a ratio of 1:9 (w/v) and stirred overnight. A 2 mL aliquot of solution was loaded into a 10 mL syringe. The PLLA-PTMC film was fixed onto the rotating drum at a speed of 850 rpm. A high voltage of 15 kV was applied to the needle tip, while the drum was grounded at -3 kV. The PCL solution was extruded at a flow rate of 0.05 mL/min for 30 min, resulting in oriented PCL fibers (thickness ~30 μm) deposited onto the PLLA-PTMC substrate. Subsequently, Au electrodes were fabricated via magnetron sputtering. A custom-designed mask plate (6 channels, 250 μm line width, 1 mm spacing) was affixed to the PCL/PLLA-PTMC bilayer film. The masked film was subjected to magnetron sputtering to deposit 10 nm thick Ti (1.5 Å/s, 420 V × 0.35 A) and 100 nm thick Au (1.5 Å/s, 440 V × 0.1 A). Then, the planar electrodes were peeled from the glass substrate and wrapped around a Kirschner needle (diameter: 1.5 mm) to form a 3D tubular structure. The device was interconnected to a flexible printed circuit board (FPCB) using hot-pressed silver paste zebra connectors. Finally, both the zebra connectors and the FPCB were encapsulated using commercial low-toxicity silicone adhesive (KWIK-SIL, World Precision Instruments, USA), protecting them from corrosion induced by biofluids.

All surgical procedures were performed on 8-week-old female Sprague-Dawley (SD) rats (n = 6 animals at each time point, weight 200-250 g). All experimental procedures were approved by the Experimental Animal Ethics Committee of the Chinese Institute for Brain Research, Beijing (approval number: CIBR-IACUC-087). First, under isoflurane anesthesia, the right sciatic nerve was exposed under sterile conditions. Next, a 6 mm nerve defect was created, and the device was sutured to the proximal and distal nerve stumps using tension-free 8-0 sutures. Then, the FPCB was subcutaneously implanted in the thigh. Finally, muscle and skin were closed using 4-0 sutures.

Gait performance was evaluated in both healthy rats and rats with NeuroSyn at 2-week intervals. Six rats (weight 200-250 g) were involved at 2, 4, 6 and 8 weeks post-implantation, respectively, and six healthy rats (weight 200-250 g) without surgery were used as controls. Rats in all groups underwent gait training and recording in identical conditions. They were immobilized via the forelimbs using a custom-made vest, allowing hindlimb contact with the treadmill on which they walked freely. During walking, gait patterns were captured using a high-speed camera (100 fps) to record continuous joint position changes, and DeepLabCut was used to extract joint locations from all videotapes. For each gait cycle, 21 kinematic parameters were computed and detailed in Supplementary Table 1. For quantitative analysis of locomotion performance, principal component analysis (PCA) was applied to normalized kinematic waveforms. Specifically, PCA was performed on data from all rats together after correlation-based normalization [adjusting the mean of the data to 0 and the standard deviation (SD) to 1], which enabled the comparison of variables with disparate values and variances. Then, locomotion performance was quantified as the Euclidean distance between the rats with NeuroSyn and healthy rats in the PC space defined by the first three PCs. Finally, extraction of key gait variables was determined by correlation between each gait variable and locomotion performance.

At 4 and 8 weeks post-implantation, peripheral nerve signals were recorded in vivo using NeuroSyn (6 channels) during treadmill walking (n = 6 animals at 4 and 8 weeks post-implantation respectively, weight 200-250 g). Rats in all groups underwent nerve signal acquisition and analysis in identical conditions. The signals were acquired at 30 kHz, pre-amplified and digitized online using Open Ephys Acquisition Board (Open Ephys, Lisbon). Subsequently, these signals were bandpass-filtered (0.1 to 20 Hz, fourth-order Butterworth filter). Linear regression referencing (LRR) was then applied to further reduce ambient noise artifacts. All subsequent neural signal acquisition and processing procedures were consistently maintained under identical parameters and methodologies. To mitigate different noise levels at 4 and 8 weeks post-implantation, a median-based z-score method was utilized to normalize the signal amplitude into units of median absolute deviation (m. a. d.) to the median. Next, the spectrum and power were analyzed by aligning the time to swing onset. Finally, the SNR was calculated as the ratio of signal variance during the swing phase (σ_Sw) to that during the stance phase (σ_St), as given by Equation (1):

(1) $$ \mathrm{SNR}(\mathrm{~dB})=10 \times \log _{10} \frac{\sigma_{\mathrm{SW}}^{2}}{\sigma_{\mathrm{St}}^{2}}\\ $$

Swing index was calculated as the ratio of the difference of power to the sum of power during swing and stance state, as given by Equation (2):

(2) $$ \text { Swing Index }=\frac{\mathrm{P}_{\text {Swing }}-\mathrm{P}_{\text {Stance }}}{\mathrm{P}_{\text {Swing }}+\mathrm{P}_{\text {Stance }}}\\ $$

At 8 weeks post-implantation, M1 signals (array center at AP: 1.75 mm, ML: 2.25 mm) were recorded in vivo simultaneously with peripheral nerve signals during treadmill walking (n = 3 animals, weight 200-250 g). A microwire array (diameter: 0.05 mm, 8 channels) was used for M1 signal recording, while NeuroSyn was used for peripheral nerve signal recording. Both signals were sampled at 30 kHz, pre-amplified and digitized online using Open Ephys. This cohort of animals (n = 3) was used exclusively for comparing peripheral nerve signals recorded by NeuroSyn with M1 signals, while the cohort (n = 6) was used to assess peripheral nerve signal recovery and gait restoration at 4 and 8 weeks postimplantation, as well as their synchronization.

Recordings of descending and ascending signals were also conducted at 8 weeks post-implantation (n = 6 animals, weight 200-250 g); however, during this procedure, the rats remained under anesthesia rather than walking on the treadmill. Animals were anesthetized using isoflurane, and a sterile surgical procedure was performed to expose the affected side. Then, a commercial cuff electrode (6 channels) was implanted near the proximal end of the nerve as a control. For ascending signal recording from the foot, needle electrodes were inserted into the plantar surface of the right paw as stimulating electrodes, and the electrical pulse stimulation (±10 mA, 1 Hz, 0.1 ms pulse width) was applied. Signals were then simultaneously recorded from both NeuroSyn and the cuff electrode using Open Ephys. Next, for descending signal recording form the spinal cord, two needle electrodes were inserted at the L1 and S2 segments respectively as stimulating electrodes, and the electrical pulse stimulation (±2 mA, 1 Hz, 0.1 ms pulse width) was applied. Signals were again collected simultaneously from NeuroSyn and cuff electrodes using Open Ephys.

EMG signals were recorded at 4 and 8 weeks post-implantation. Under isoflurane anesthesia, the rats were operated to expose the sciatic nerves of both affected and healthy sides (n = 5 animals at 4 and 8 weeks post-implantation respectively, weight 200-250 g). Rats in all groups underwent EMG signal recording in identical conditions. Then, a commercial double-hook electrode was placed proximal to the sciatic nerve as the stimulating electrode, and needle electrodes were inserted into the gastrocnemius muscle as the recording electrodes. Next, sciatic nerve stimulation was performed using single pulses (0.1 mA, 0.5 mA, 1 mA, 3 mA, 5 mA, 10 mA), and five consecutive EMG signals were recorded by an electrophysiology instrument (6240EC, Chengdu Instrument Factory, China). Finally, the peak-to-peak amplitudes of the EMG signals were analyzed. When NeuroSyn was used as a stimulating electrode in a separate cohort of animals (n = 6 animals at 8 weeks post-implantation, weight 200-250 g), two adjacent electrode channels were selected for stimulation, with the positive and negative poles connected accordingly. Next, sciatic nerve stimulation was performed using square-wave pulses (±5 mA, 1 Hz, 0.1 ms), and five consecutive EMG signals were recorded.

For classification of neural data into swing/stance states, we used a neural network model. The architecture consisted of a 1D convolutional neural network (1D CNN) and a long short-term memory (LSTM) layer. Moreover, the training dropout technique was applied to the LSTM and Dense layers. Since swing states were always fewer than stance states, we randomly selected an equal number of stance state samples to balance the training dataset. Mean Squared Error (MSE) was used for the loss function.

At 4 and 8 weeks post-implantation, rats were euthanized by intramuscular injection of an overdose of xylazine/ketamine (n = 3 at 4 and 8 weeks post-implantation respectively, weight 200-250 g). Rats in all groups underwent immunofluorescence staining in identical conditions. Sciatic nerve segments located at 2 mm from the proximal end were dissected, immediately immersed in 4% paraformaldehyde (10:1 fixative-to-tissue volume ratio), and fixed for 24 h at 4 °C. Then, samples were dehydrated in graded sucrose solutions, embedded in optimal cutting temperature compound, and snap-frozen in liquid nitrogen. After that, transverse nerve sections were stained with the following primary and secondary antibodies: Mouse anti-βⅢ Tubulin antibody (Abcam, USA), Rabbit anti-S100 antibody (Abcam, USA), Goat anti-rabbit IgG (H+L) cross-adsorbed secondary antibody (Alexa Fluor^TM 647, Abcam, USA), and Goat anti-mouse IgG, IgM (H+L) secondary antibody (Alexa Fluor^TM 488, Abcam, USA). Finally, sections were visualized using an orthogonal confocal microscope (LSM780, Zeiss, Germany).

At 4 and 8 weeks post-implantation, rats were anesthetized by isoflurane and peripheral blood was collected from the orbital vein using a vacuum blood collection tube (n = 3 animals at 4 and 8 weeks post-implantation respectively, weight 200-250 g). Rats in all groups underwent blood chemistry analysis in identical conditions. After 2 h of clotting at room temperature, whole blood samples underwent centrifugation (3,000 rpm, 15 min) to isolate serum. Biochemical and electrolyte tests were performed by Servicebio Biotechnology Co., Ltd.

Data are all expressed as mean ± SD. Differences between groups were analyzed using one‑way analysis of variance (ANOVA) (^*P < 0.05, ^**P < 0.01, ^***P < 0.001), followed by Tukey’s post hoc test using SPSS software (version 27.0).

The schematic illustration and fabrication process of NeuroSyn are shown in Figs. 1a, b, respectively. Photographs of NeuroSyn in the planar format and conduit format connected to a custom FPCB are shown in Figs. 1c, d. We use biodegradable PLLA-PTMC as the device substrate (~120 μm), providing mechanical support with an elastic modulus of 2.2 MPa, close to that of peripheral nerves (0.58-24.4 MPa)^[38-40]. After immersion in PBS (pH 7.4, 37 °C) over a time frame of 6 weeks, the elastic modulus of PLLA-PTMC increases to 359.1 MPa (Week 6), and the breaking elongation decreases from 638.7% (Week 0) to 393.7% (Week 6), showing a gradual transition toward a stiffer and more brittle state [Supplementary Figure 1A]. However, compared with the widely used non-biodegradable material polyimide (PI), which exhibits elastic modulus and breaking elongation of 2,398.7 MPa and 8.2% respectively [Supplementary Figure 1B], PLLA-PTMC retains substantially higher flexibility to conform to the shape of peripheral nerves and prevent deformation caused by movement, thereby supporting high-fidelity signal recording. Furthermore, aligned PCL fibers (thickness ~30 μm) are fabricated on the PLLA-PTMC substrate via electrospinning [Figure 1B], providing topographic guidance cues that effectively promote directional axon regeneration^[36,41]. To ensure reliable signal acquisition, thin-film Au (100 nm) with Ti (10 nm) adhesion layer is deposited on the PCL layer by magnetron sputtering through shadow masks, yielding patterned 6-channel fiber electrodes that feature a large effective sensing area, low impedance and a desirable SNR. The electrodes are arranged in an L-shaped configuration [Figure 1C], with recording ends aligned to the PCL fibers and neurite growth, facilitating stable electrode-tissue interfaces. The opposite ends of the electrodes serve as the electrical interface, where zebra connectors are thermally laminated to link the fiber electrodes to the FPCB. The device exhibits excellent mechanical compliance [Figure 1C], and when rolled, it forms a 3D cylindrical architecture that closely mimics native nerve structure [Figure 1D]. SEM confirms the microstructure of the Au fiber electrode on aligned PCL fibers [Figure 1E and F], while EDS verifies uniform Au distribution across the fibrous matrix [Supplementary Figure 2]. To assess guidance effects, dorsal root ganglion (DRG) neurons are co-cultured with NeuroSyn for 7 days. Specifically, SEM and immunostaining images show axons extending unidirectionally along the PCL fibers [Figure 1G and H, Supplementary Figure 3].

Optimization of PCL concentration (5%-23%) reveals a concentration-dependent effect on fiber architecture and neuronal guidance: higher concentrations (> 9%) increase fiber diameter but progressively disrupt orientation, while lower concentrations (≤ 9%) yield discontinuous fibers with bead formation [Supplementary Figure 4]. The 9% formulation achieves an optimally aligned structure [Figure 1F and I], supporting directional axonal growth of DRG neurons, whereas disruption of this alignment results in disorganized growth [Supplementary Figure 5]. Therefore, a PCL concentration of 9% is selected for device fabrication.

Electrochemical impedance spectroscopy of Au electrodes (electrode area: 0.07 cm²) in PBS (pH 7.4) shows an impedance of 290.5 ± 48.6 Ω at 1 kHz [Figure 1J]. Notably, the impedance remains stable after 68 days in PBS (pH 7.4, 37 °C) [Figure 1K], indicating long-term recording stability under physiological conditions. After the cyclic bending test 10,000 cycles (bending radius: 5 mm, frequency: 1 Hz) to mimic the repetitive movement under physiological conditions, the average impedance increases by 20% at 1 kHz [Supplementary Figure 6A], which remains within the acceptable range for neural signal recording. In addition, SEM images reveal no evidence of microcracks on the Au films after either the rolling process to form the conduit or the cyclic bending test [Supplementary Figure 6B], confirming the mechanical robustness of the electrode. Furthermore, we evaluate the feasibility of NeuroSyn for neural stimulation to promote peripheral nerve repair and/or monitor recovery status. Specifically, CV reveals a cathodic charge storage capacity (CSC) of 24.1 ± 2.79 mC/cm² for NeuroSyn [Supplementary Figure 7A], exceeding that of previously reported Au electrodes^[42,43]. This improvement in CSC is attributed to the enlarged effective surface area generated by the Au nanofiber structure, which could allow neuronal activation at lower current amplitudes, thereby reducing the risk of neural damage^[44]. Additionally, the electrodes show only a modest CSC decrease after soaking in PBS (pH 7.4, 37 °C) for 6 days, likely due to Cl^--induced passivation^[45], with CSC maintained at 15.1 ± 3.54 mC/cm² [Supplementary Figure 7B]. Under daily monophasic stimulation for 6 days (0.5 V, 20 Hz, pulse width 1 ms, 1 h daily), both impedance and CSC remain stable [Supplementary Figure 7C and D], confirming NeuroSyn’s feasibility for early electrical stimulation therapy following PNI.

The degradation of NeuroSyn is evaluated under accelerated conditions (PBS, pH 7.4, 65 °C). The device consists of ~0.51 wt% Au, with the remaining components composed of biodegradable polymers. Specifically, degradation proceeds through polymer hydrolysis and fragmentation (day 23), followed by the breakup of the Au thin films (day 68), leaving ~0.02 wt% residual Au fragments by day 148 [Figure 1L]. The degradation rate of the device under physiological conditions can be estimated by the Arrhenius equation^[46,47]:

(3) $$ \mathrm{k}=\mathrm{k}_{0} \cdot \exp \left(-\mathrm{E}_{\mathrm{A}} / \mathrm{RT}\right) $$

where k₀ is the pre-exponential factor, E_A is the activation energy for hydrolytic degradation (83.7 kJ/mol for PLLA^[48], 15.6 kJ/mol for PCL^[49]), and R is the universal gas constant [8.314 J/(mol·K)]. Consequently, the physiological degradation rate (37 °C) is 14.7 (for PLLA) or 1.65 (for PCL) times slower than accelerated degradation (65 °C). This indicates that the device maintains structural integrity throughout the critical nerve regeneration window while ensuring complete resorption within approximately 2.7 years post-implantation, which is suitable for long-term signal monitoring without permanent foreign body presence. In addition, reducing Au thickness to 30 nm facilitates device fragmentation into smaller pieces, while maintaining a reasonable impedance at 1 kHz (4,636.2 ± 1,532.3 Ω) [Supplementary Figure 8], comparable to previously reported Au electrodes^[50,51]. Moreover, studies on submicron Au (50-250 nm), micro-Au (> 20 μm), and bulk Au implants consistently show that Au ions are slowly released via macrophage-mediated surface oxidation, taken up by local macrophages, mast cells, and fibroblasts, and stored in lysosomes for months^[52-54]. Furthermore, the biological effects of Au ions are confined to cells adjacent to the Au particles, with no detectable spread to distant organs, such as the lung and liver. Au is observed in the kidney as faint staining only in animals implanted with multiple Au grids (total mass ~5.5 mg), further supporting the local safety profile^[54]. Notably, the total Au mass in our device prior to degradation is approximately 81.1 μg, which is far below the threshold associated with systemic accumulation.

We apply NeuroSyn to the sciatic nerve with a gap defect (6 mm) in SD rats to evaluate its ability to monitor nerve recovery and decode motor intentions over 8 weeks, which matches the time frame for nerve regrowth. Peripheral nerve regeneration follows a well-defined biological timeline. Within 24-48 h after injury, axonal degeneration and segmental disintegration of the distal axon begin. One week after injury, macrophages efficiently clear myelin and axonal debris. Under the guidance of Schwann cells, damaged axons form growth cones and begin to regenerate along the Büngner band, while the supportive capacity of Schwann cells begins to decline after approximately two months, which is a key factor limiting successful reinnervation of target muscles^[55]. Therefore, monitoring neural signals during this window enables assessment of regeneration status and provides timely stimulation or rehabilitation guidance. Once functional reinnervation is fully established (3-6 months)^[15], the device can degrade naturally in vivo, eliminating the need for a second surgery and reducing associated infection risks. The device is implanted at the injury site, bridging the proximal and distal ends of the damaged nerve [Figure 2A]. Immunofluorescence staining of the regenerated nerve tissue (transverse section, 2 mm from the proximal end) shows a larger myelinated area (S100, red) surrounding nerve axons (β-tubulin, green) at 8 weeks than that at 4 weeks post-implantation, suggesting the progressive maturation of myelination and the repair of injured nerve tissues [Figure 2B, Supplementary Figure 9A]. The statistically average axon density increases from 1,670.4 /mm² at 4 weeks post-implantation to 3,464.5 /mm² at 8 weeks post-implantation (P < 0.05), indicating the active nerve regeneration [Supplementary Figure 9B]. Moreover, the longitudinal sections of the regenerated nerve at 8 weeks post-implantation show wellaligned axons (β-tubulin, green) traversing the conduit, indicating that the regenerated nerve achieves a continuous, oriented architecture [Supplementary Figure 10].

We further assess nerve conduction recovery via EMG of the gastrocnemius muscles on the injured side at 4 and 8 weeks post-implantation by applying stimulation currents of 0.1-10 mA at the proximal end of the injured nerves [Figure 2C, Supplementary Figure 11], with the contralateral healthy side as a control. At 4 weeks post-implantation, limited axonal growth and poor myelination result in incomplete reinnervation and undetectable EMG signals. At 8 weeks post-implantation, although EMG responses remain below healthy levels at greater stimulation currents, enhanced axonal regrowth and myelination restore neuromuscular connectivity, yielding increased EMG amplitudes and decreased EMG latencies across all stimulation intensities compared with those at 4 weeks post-implantation. These results suggest that we can potentially use NeuroSyn to deliver electrical stimulation to assess recovery by interrogating EMG signals in target muscles, which will be discussed in the following section.

To assess NeuroSyn’s capacity for neural recovery monitoring and motor intention recognition, we record neural signals and gait simultaneously during the recovery of injured nerves [Supplementary Figure 12]. Gait kinematics are captured with a high-speed camera (100 fps) during treadmill walking [Figure 2D]. Using DeepLabCut, we extract key joint positions, including crest, hip, knee, ankle, metatarsophalangeal (MTP), and toe. Representative stick diagrams illustrate sequential hindlimb movements during a single swing phase in rats implanted with NeuroSyn at 4 and 8 weeks post-implantation, with each stick extracted from individual captured frames [Figure 2E]. Compared with healthy rats [Supplementary Figure 13A], rats at 4 weeks post-implantation exhibit markedly altered gait patterns, including prolonged swing duration. In contrast, at 8 weeks post-implantation their gait more closely resembles that of healthy controls. To quantify gait changes during nerve repair (2-8 weeks post-implantation), we apply PCA^[56] to 21 standardized locomotor metrics, encompassing temporal gait features, limb trajectories, joint angles, and limb coordination [Supplementary Table 1, Supplementary Figure 13B]. We utilize the gait performance index, which is defined as the Euclidean distance^[57] between device-implanted rats and healthy rats in the principal component space, where smaller values indicate a closer level of recovery. The index [Figure 2F] peaks at 4 weeks post-implantation and subsequently declines, consistent with catwalk recordings in previous reports^[58,59] and reflecting early denervation due to PNI followed by progressive functional recovery through reinnervation. At 8 weeks post-implantation, the overall gait performance exhibits a notable enhancement compared with that at 4 and 6 weeks (P < 0.001). In addition, among the 21 gait parameters analyzed, the changes in swing duration exhibit a stronger correlation with the progress of gait recovery, as revealed by the Pearson correlation analysis with the gait performance index [Supplementary Figure 13C].

During treadmill walking at 4 and 8 weeks post-implantation, neural signals are recorded via NeuroSyn simultaneously. The signals undergo bandpass filtering (0.1-20 Hz, fourth-order Butterworth) and are further refined through denoising by LRR^[60]. To facilitate meaningful comparisons across different time points, which may exhibit varying noise levels, the neural signals’ amplitudes are normalized by a robust median-based z-score method. This method standardizes raw values into units of median absolute deviation (m.a.d.) to the median, as shown in Figure 2G and H. Representative neural signals with corresponding gait maps reveal stronger neural activity at 8 weeks post-implantation compared with that at 4 weeks, with all six channels exhibiting a consistent trend [Figure 2G and H]. Furthermore, SNR analysis [Figure 2I] confirms this improvement, with SNR values increasing from 4 to 8 weeks (mean SNR: 4 weeks, 9.48 dB; 8 weeks, 19.5 dB; P < 0.01). These results indicate that NeuroSyn’s signal quality improves in parallel with the morphological and functional recovery of the nerve. Since the electrode impedance remains stable over time [Figure 1K], the increasing SNR is likely due to improved neural activity resulting from enhanced regenerated axons and myelination rather than changes in electrode properties.

We further investigate the synchronization between neural signals recorded by NeuroSyn and gait [Figure 2J-L]. We define the start of swing as time zero and analyze the signal power within 2 s before and after swing onset [Figure 2J and K]. Strong neural signals are observed only around 0 s at both 4 and 8 weeks post-implantation, and the signal power is higher at 8 weeks than that at 4 weeks. To quantify the synchronization between neural signal power and swing events, we calculate the swing index (see Experimental section for details) to assess changes in neural signals across different gait cycle stages. A positive swing index reflects increased power during the swing phase, indicating concurrent signal generation with swing. As shown in Figure 2L, the swing index at 8 weeks post-implantation is significantly higher than that at 4 weeks (mean swing index: 4 weeks, 0.235; 8 weeks, 0.434; P < 0.001). These findings show that neural signals are strongly correlated with the neural recovery process, with signal intensity, SNR, and signal-swing synchronization improving over the time frame of nerve recovery, which are consistent with the histological and electrophysiological results. Therefore, NeuroSyn can enable real-time tracking of recovery status after PNI through multiple parameters, including signal intensity, SNR, and gait synchronization, which is critical for ensuring timely diagnosis and treatment. By contrast, traditional diagnostic methods often rely on complex and bulky imaging equipment and are applied intermittently^[61,62], which precludes continuous monitoring and early diagnosis.

Furthermore, we employ machine learning to decode the recorded neural signals and predict locomotor states. Neural signals acquired by NeuroSyn at 8 weeks post-implantation are input into a neural network model to classify swing and stance phases [Figure 3A]. The model consists of two main components: a 1D CNN for extracting local temporal and spectral features, and an LSTM for capturing long-range temporal dependencies. The processed features from both networks are concatenated and passed through a DENSE layer for combination and transformation. The model output is converted into a probability distribution by the SOFTMAX function for classification prediction. To enhance the robustness and prevent overfitting, we apply dropout techniques during training of the LSTM and Dense layers.

Representative outputs are presented in Figure 3B, showing both the actual gait phase recorded by the camera and the predicted gait phase. The model demonstrates a highly accurate state classification, achieving 97% accuracy for the swing phase and 93% for the stance phase, with an overall accuracy of 95% [Figure 3C], which is higher than other simpler machine learning models (best accuracy: 94.5%) [Supplementary Figure 14]. The model’s classification accuracy is further validated by the receiver operating characteristic (ROC) curve shown in Figure 3D, which yields an area under the curve (AUC) value of 0.991. These results confirm the excellent performance of the predictive model. Thus, peripheral neural signals recorded by NeuroSyn at 8 weeks post-implantation enable highly accurate prediction of movement patterns.

Based on the ability of peripheral nerve signals to predict movement patterns, we next compared their synchronization with gait against signals recorded from M1 region, which is a common approach for controlling exoskeleton robots. We demonstrate that peripheral signals exhibit more reliable synchronization with gait than brain signals. Concurrently with signal recording by NeuroSyn, we implant an 8-channel electrode array in the M1 region of the rat brain. At 8 weeks post-implantation, we record both M1 and sciatic nerve (SN) signals from rats walking on a treadmill, with representative traces presented in Figure 3E and F. The analysis of signal synchronization with gait reveals that neural signals are detected in both the M1 region and SN at the onset of swing (t = 0 s). Notably, the signals recorded by NeuroSyn in SN exhibit significantly higher power compared with those detected in the M1 region [Figure 3G]. Furthermore, the swing index reflects that SN signals recorded by NeuroSyn are significantly higher than M1 signals (mean swing index: M1, 0.0639; SN, 0.275; P < 0.01) [Figure 3H]. These findings demonstrate that SN signals are more strongly synchronized with swing events than M1 signals, indicating that NeuroSyn’s recordings more accurately reflect motor intentions. Integration with machine learning algorithms could enable reliable prediction of intended movements for exoskeleton control using the NeuroSyn platform.

The closed-loop system in exoskeleton robots requires the ability not only to accurately interpret signals generated by nerves and muscles, but also to provide electrical stimulation to enhance motor function^[63]. To validate NeuroSyn’s capacity for closed-loop control, we monitor ascending and descending induced signals in the nervous system, and use NeuroSyn as the stimulation electrode to activate target muscles.

At 8 weeks post-implantation, we first record evoked signals from the sciatic nerve using NeuroSyn during epidural spinal stimulation (±2 mA, 1 Hz, pulse width 0.1 ms) [Figure 4A]. A commercial six-channel cuff electrode is implanted simultaneously as a control. To eliminate stimulation artifacts, we apply currents of opposite phase, as artifacts invert polarity whereas endogenous neural signals do not. Quantitative analysis shows comparable waveforms [Figure 4B] and amplitude distributions [Figure 4C] (P = 0.689) between NeuroSyn and the cuff electrode. Next, we record ascending signals from the sciatic nerve using NeuroSyn during toe stimulation at 8 weeks post-implantation (±10 mA, 1 Hz, pulse width 0.1 ms) [Figure 4D]. Both NeuroSyn and the cuff electrode demonstrate comparable waveforms [Figure 4E] and amplitude distributions [Figure 4F] (P = 0.633). These results demonstrate that NeuroSyn reliably captures descending and ascending evoked signals at 8 weeks post-implantation, achieving the reception of commands from both the central nervous system and the peripheral effector.

To evaluate NeuroSyn’s electrical stimulation capability, we first apply NeuroSyn to the intact sciatic nerve, with three channels as the positive electrode and the remaining three as the negative electrode. EMG is recorded from the gastrocnemius muscle, with a commercial hook electrode serving as a control. NeuroSyn elicits EMG amplitudes comparable to those of the commercial electrode [Supplementary Figure 15A and B]. We then investigate the stimulation of injured sciatic nerves using NeuroSyn, by implanting it at 6-mm sciatic nerve defects [Figure 4G]. At 8 weeks post-implantation, stimulation via NeuroSyn evokes EMG signals successfully [Figure 4H and I] with average amplitudes of 0.1 mV. Although EMG amplitudes decrease after PNI, the measurable responses confirm functional reinnervation and stable stimulation throughout regeneration.

To evaluate whether the proposed system is compatible with real-time closed-loop applications, we conduct a simulated latency analysis to estimate the time delay between neural signal acquisition and the delivery of corresponding stimulation. This delay includes acquiring sciatic nerve signals, decoding the signals, recognizing the gait state, transferring the signals to TTL output and sending stimulation to the nerve [Supplementary Figure 16A]. To emulate real-time processing conditions, we use peripheral nerve signals recorded by NeuroSyn with different channels during locomotion. The analysis reveals an average latency of approximately 106 ms using 6-channel NeuroSyn, which is lower than that using single-channel NeuroSyn [Supplementary Figure 16B]. This latency falls within the time scale considered compatible with real-time locomotor assistance and exoskeleton control (~100-200 ms)^[64-66]. This demonstrates that NeuroSyn and the decoding system are technically capable of supporting real-time closed-loop neuromodulation. Collectively, the results suggest that NeuroSyn can potentially serve as a bidirectional interface to monitor motor function recovery in real time and deliver appropriate electrical stimulation during rehabilitation.

NeuroSyn is implanted into rat sciatic nerve defects and monitored over 20 weeks. The substrate gradually fragments, leaving only minor Au and PLLA-PTMC residues by week 20 [Figure 5A], supporting its structural stability during nerve regeneration and its ability to degrade into small fragments after device function is completed. Histological analysis of major organs (heart, kidney, liver, lung, spleen) at 8 weeks post-implantation reveals normal architecture comparable to healthy controls [Figure 5B], indicating no systemic toxicity. Additionally, hematological assessments, including complete blood counts and serum biochemistry at 4 and 8 weeks post-implantation, remain within physiological ranges [Figure 5C and D]. In addition to these systemic safety data, the average axon density at 8 weeks post-implantation (3,464.5 /mm²) is greater than that at 4 weeks post-implantation (1,670.4 /mm²), reaching approximately 50% of that in healthy rats (6,865.9 /mm²) [Supplementary Figure 9B]. This significant increase over time (P < 0.05) demonstrates active and progressive axonal regeneration within the conduit, supporting the favorable biocompatibility of our device. Furthermore, the mechanical compatibility of the device and the well-established material selection can both support the local biocompatibility. The elastic modulus of PLLA-PTMC during degradation (2.2-359.1 MPa) is comparable to that of rat sciatic nerves, which mitigates foreign body reactions caused by the mechanical mismatch between the implant and the nerve^[67-69]. Furthermore, the constituent materials, including PLLA-PTMC, PCL, and Au, are all demonstrated to have good compatibility. For example, PLLA-PTMC has shown only moderate tissue response over 90 days in a rat femoral nerve^[70], and a comprehensive FDA-commissioned safety profile further indicates that PCL is widely used in soft tissue applications with local responses comparable to or lower than those of non-PCL materials^[71]. Au is also a widely used electrode material in neural interfaces, with extensive literature confirming its histological safety and minimal foreign body reaction^[72-74]. Together, these findings demonstrate NeuroSyn’s favorable biodegradability and biocompatibility in physiological environments, which can eliminate potential infection risks and tissue damage associated with device removal.