Biomimetic computer-to-brain communication enhancing naturalistic touch sensations via peripheral nerve stimulation

The study addresses the challenge that current neurotechnologies for restoring somatosensation via peripheral or central nervous system stimulation often evoke unnatural, paresthetic sensations due to synchronized, tonic stimulation patterns that do not follow natural tactile coding. The authors hypothesize that biomimetic stimulation, derived from realistic models of mechanoreceptor population dynamics, can produce neural activity that better resembles natural touch, leading to more intuitive sensations and improved functional outcomes. They aim to design biomimetic stimulation strategies using an in-silico model (FootSim), validate their transmission and similarity to natural touch in animals (cats) by recording at DRG and spinal cord, and evaluate sensation quality and functional benefits in implanted lower-limb amputees using a closed-loop neuro-robotic leg.

Prior work shows that implantable BCIs and neurostimulation of peripheral nerves, spinal cord, and somatosensory cortex can restore some lost sensations, improving prosthesis use and acceptance. However, common neuromodulation uses predefined constant frequencies, synchronously activating neurons and likely contributing to paresthesia. Natural touch involves probabilistic, asynchronous activation across afferent classes with early convergence in cuneate nucleus and S1. Biomimetic sensory feedback approaches in upper-limb prosthetics have yielded more intuitive sensations and better object interactions than non-biomimetic patterns, but were not tested in lower-limb amputees and lacked mechanistic insight along the somatosensory neuroaxis. The literature suggests frequency strongly shapes percept quality, and aggregate multi-afferent coding may be necessary to reproduce naturalistic percepts.

Multilevel framework: (1) In-silico modeling (FootSim), (2) animal validation in decerebrated cats, and (3) human clinical testing with implanted amputees and real-time neuro-robotic device. In-silico (FootSim): FootSim models foot sole cutaneous afferents (FAI/FAII/SAI/SAII) using microneurography-fitted firing models and a mechanical skin stress model. The foot sole can be populated with specific afferent types or full realistic distributions. Authors applied 2 s ramp-and-hold stimuli (0.15 s on-phase, 0.3 s off-phase) plus low-amplitude environmental noise (up to 0.5% of max) over the whole plantar surface. For each scenario (FAI, FAII, SAI, SAII, and FULL population), FootSim provided spiking responses across afferents, aggregated into PSTHs. Smoothed PSTHs modulated stimulation frequency; amplitude and pulse width remained constant per pattern. Five biomimetic patterns were defined (FAI-, FAII-, SAI-, SAII-like, and FULL biomimetic), encoding temporal dynamics while maintaining constant charge parameters (frequency as quality determinant; charge as intensity determinant). Animal experiments: Two adult decerebrated cats. A cuff electrode was placed on the tibial nerve for stimulation; amplitude tuned just above threshold (60 µA) defined by dorsal spinal afferent volleys (> meanBaseline + 2.5*stdBaseline). Recordings: 32-channel linear probe in L6 spinal cord gray matter, and 32-channel Utah array in L6 DRG. Conditions included five biomimetic stimulation patterns, a tonic 50 Hz stimulation, and natural touch (cotton swab rubbing; 15 repetitions). Biomimetic patterns repeated 90 times each. Signals were preprocessed (comb notch at 50 Hz and harmonics; high-pass 30 Hz; artifact removal). Spiking extracted by band-pass 800–5000 Hz and threshold crossing; PSTHs computed with bin sizes guided by Freedman–Diaconis rule (10–53 ms). LFPs (30–300 Hz) and current source density (CSD) analyses were performed; similarity quantified via cross-correlation, Pearson correlation, and KL divergence versus natural touch. Human clinical study: Three unilateral transfemoral amputees (ClinicalTrials.gov NCT03350061) were implanted with four TIME-4H intraneural electrodes (14 active sites each; 56 AS total per participant) in the tibial nerve. Mapping characterized perceptual thresholds and max charge per AS at 50 Hz, ensuring safety (<120 nC). Sensation locations (projected fields) spanned frontal, central, lateral metatarsus, and heel. Open-loop naturalness: Participants received 2 s current pulse trains with increasing (0.5 s), static (1 s), and decreasing (0.5 s) phases using encoding strategies: (i) Linear amplitude neuromodulation, (ii) Sinusoidal pulse-width modulation, (iii) Poisson frequency variant (mean 50 Hz), and (iv) Biomimetic patterns (SAI-, SAII-, FAI-, FAII-like, and FULL). Randomized trials were repeated three times per condition per AS; participants rated naturalness (VAS 0–5) blinded to encoding. Intensity was monitored to avoid bias. Closed-loop neuro-robotic device: Real-time system integrated a sensorized insole (heel, lateral/medial, frontal sensors), a microprocessor-based prosthetic knee-foot system (Ossur RHEO KNEE XC), a microcontroller implementing encoding (including precomputed BIOM frequency trains with linear amplitude modulation), and a multichannel stimulator (STIMEP) delivering intraneural stimulation through TIMEs. Feedback conditions: NF (no feedback), LIN (linear amplitude mapping; 50 Hz tonic), DISC (time-discrete short trains at gait-phase transitions, 50 Hz), BIOM (biomimetic frequency modulation plus linear amplitude modulation; three channels chosen per participant by area-specific naturalness). Functional tasks with two participants (S1, S2): (a) Stairs Task (ST): 30 s sessions, 10 repetitions per condition; outcome: laps/session and post-session self-reported confidence (VAS 0–10). (b) Cognitive Dual Task (CDT): walking 5 m baseline and while spelling backward a five-letter word; 10 repetitions per condition; outcomes: walking speed (m/s) and mental accuracy (% correct letters). Statistical analyses used normality tests, ANOVA or Kruskal–Wallis with post-hoc corrections; effect sizes reported.

Animal electrophysiology:

• Biomimetic stimulation PSTH dynamics were highly correlated with DRG and spinal responses, indicating faithful transmission of temporal patterns along the somatosensory neuroaxis.
• DRG LFP amplitude distributions: KL divergence versus natural touch was much smaller for biomimetic than tonic stimulation (biomimetic vs natural KL = 0.26; tonic vs natural KL = 2.06). Cross-subject natural touch comparison yielded KL = 0.66, suggesting generalizable natural dynamics.
• Spinal CSD similarity to natural touch: biomimetic vs natural correlation = 0.11 (p = 0.005); tonic vs natural correlation = −0.03 (p = 0.344); biomimetic vs tonic correlation = −0.13 (p = 0.001). Channel-wise analyses confirmed higher similarity for biomimetic vs natural and near-zero for tonic vs natural; shuffled controls approached zero.
• Spatial similarity along spinal dorsoventral axis: electrical stimulation produced higher inter-channel similarity than natural touch. Biomimetic FULL median correlation across channels was 0.60 (25th 0.45, 75th 0.67) vs tonic 50 Hz median 0.88 (25th 0.83, 75th 0.90). FULL biomimetic differed significantly from single-afferent biomimetics and from tonic and natural (multiple χ² tests; e.g., 50 Hz–Natural p < 0.001; Natural–FULL biom p = 0.035). Human open-loop naturalness (VAS 0–5):
• Across all tested active sites and three participants, biomimetic encodings elicited more natural sensations than linear encoding. Grouped means: S1: 3 ± 0.18 (BIOM) vs 1 ± 0.35 (LIN); S2: 2 ± 0.16 vs 0.5 ± 0.17; S3: 2 ± 0.36 vs 1 ± 0.18. Significance: S1 p < 0.001, f = 0.62; S2 p < 0.001, f = 0.89; S3 p = 0.0026, f = 1.01.
• Biomimetic outperformed sinusoidal and Poisson strategies on multiple patterns and participants (e.g., S1: p < 0.001 for SAI, FAI, FAII, FULL vs sinusoidal; p < 0.001 for SAI, SAII, FAI, FAII, FULL vs Poisson). No single biomimetic type dominated across all locations, likely reflecting differing recruited afferent mixes per channel/area. Closed-loop functional outcomes:
• Stairs Task (ST) speed (laps/session): BIOM improved over LIN, DISC, and NF. S1: BIOM 4.9 ± 0.1 vs LIN 4.5 ± 0.1 (p < 0.001), DISC 4.6 ± 0.1 (p = 0.004), NF 4.3 ± 0.1 (p < 0.001), effect size f = 2.14. S2: BIOM 4.3 ± 0.4 vs LIN 3.8 ± 0.1 (p = 0.047), DISC 3.6 ± 0.1 (p = 0.047), NF 3.5 ± 0.1 (p < 0.001), f = 1.5.
• Self-reported confidence (VAS 0–10) in ST: S1: BIOM 9.75 ± 0.26 vs LIN 8.75 ± 0.62 (p = 0.015), DISC 7.83 ± 0.39 (p < 0.001), NF 6.67 ± 0.49 (p < 0.001), f = 2.57. S2: BIOM 6 ± 0.3 vs LIN 5.37 ± 0.23 (p = 0.014), DISC 5.17 ± 0.25 (p < 0.001), NF 3.83 ± 0.25 (p < 0.001), f = 3.09.
• Cognitive Dual Task (CDT) mental accuracy (%): S1: BIOM 76 ± 16 vs LIN 58 ± 20 (p = 0.016), DISC 58 ± 11 (p = 0.028), NF 52 ± 17 (p = 0.004), f = 0.57. S2: BIOM 94 ± 9.6 vs LIN 72 ± 17 (p = 0.044), DISC 50 ± 37 (p < 0.001), NF 48 ± 14 (p < 0.001), f = 0.97. Walking speed in CDT was higher with feedback vs NF for S2 (p < 0.001) and in BIOM (p < 0.001) and LIN (p = 0.002) for S1. Overall, biomimetic stimulation produced neural dynamics closer to natural touch than tonic stimulation and yielded more natural percepts and better functional performance in implanted amputees.

The findings support the central hypothesis that time-varying, model-informed biomimetic peripheral nerve stimulation can encode artificial tactile information with spatiotemporal neural dynamics resembling natural touch. In animals, biomimetic patterns propagated from the stimulated nerve to DRG and spinal cord, matching natural touch more closely than tonic stimulation in PSTH, LFP distributions, and spinal CSD topology, and reducing synchronized overactivation that may underlie paresthesia. In humans, biomimetic encodings consistently increased perceived naturalness across implant channels versus conventional linear, sinusoidal, or Poisson strategies, though the optimal biomimetic pattern varied by foot region and likely by the composition of recruited afferents. Implemented in a real-time neuro-robotic leg, biomimetic feedback improved mobility (stair-walking speed), confidence, and cognitive dual-task performance without compromising gait speed, indicating more efficient sensorimotor integration and reduced mental workload. These results substantiate biomimetic encoding as a key design principle for next-generation neuroprostheses and neuromodulation devices, aligning artificial stimulation with physiological coding to enhance perception and function.

This work presents a neuroscience-driven, multilevel framework that integrates realistic in-silico mechanoreceptor modeling (FootSim), animal neurophysiology, and human clinical testing to realize biomimetic peripheral nerve stimulation. The approach enables faithful transmission of naturalistic temporal patterns along the somatosensory neuroaxis, evokes more natural perceptions than standard encodings, and improves functional outcomes in lower-limb amputees using a neuro-robotic prosthesis. The study positions biomimetic stimulation as a foundational feature for future neuroprostheses and broader neuromodulation applications. Future research should: (i) refine models to incorporate richer mechanics (e.g., shear, sliding) and real-time computation; (ii) investigate higher-level processing (gracilis/cuneate, thalamus, cortex) and embodiment; (iii) personalize encoding to local afferent distributions using adaptive/ML methods; and (iv) validate in larger, randomized, controlled clinical trials with comprehensive gait and neurophysiological assessments.

Animal natural touch was delivered manually rather than with a robotic, precisely controlled mechanical stimulator, limiting standardization. FootSim did not explicitly model shear forces or lateral sliding, likely reducing accuracy for SAII (skin stretch) responses. Biomimetic patterns used in closed-loop were computed offline; a fully real-time, possibly implantable system is still needed. Clinical assessments involved small sample sizes; Participant 3 did not perform functional tasks. Naturalness was rated via a single-item VAS; broader questionnaires could capture richer perceptual dimensions. Future trials should include quantitative gait analyses, fatigue and neurophysiological measures, and randomized, double-blind designs to establish generalizability and clinical efficacy.