Hair-Thin Nerve Implants Decode Phantom Leg Movements in Above-Knee Amputees, Opening Path to Neurally Controlled Prosthetics

Overview

A team led by Chalmers University of Technology in Sweden has demonstrated, for the first time, the decoding of intended leg movements directly from peripheral nerves in people with above-knee amputations. The proof-of-concept study, published in Nature Communications, used hair-thin electrodes implanted in the sciatic nerve and a bio-inspired artificial intelligence decoder to classify phantom movements of the knee, ankle, and toes with accuracy significantly above chance. The results suggest a pathway toward prosthetic legs that respond to the wearer’s neural intentions rather than relying on pre-programmed gait cycles.

What We Know

The study enrolled two individuals with transfemoral (above-knee) amputations who were active users of passive prosthetic devices. Surgeons implanted four transversal intrafascicular multichannel electrodes (TIMEs) into the tibial branch of each participant’s sciatic nerve, providing 56 active recording sites sampled at 30 kHz. Each electrode is roughly the width of a human hair and flexible enough to conform to the nerve’s internal structure.

Participants were asked to attempt six phantom limb movements: knee flexion and extension, ankle dorsiflexion and plantarflexion, and toe flexion and extension. In the first participant, 91 percent of implanted electrode channels showed responsiveness to at least one phantom movement, with firing rates that differed significantly across joint types. The second participant showed a lower but still meaningful 37.5 percent responsiveness rate.

To decode these signals, the researchers built a spiking neural network (SNN) that processes time-based electrical spikes rather than continuous numerical values, mimicking more closely how biological neurons communicate. The SNN outperformed both linear support vector machines and multilayer perceptrons in movement classification, achieving 58 percent accuracy across six movement classes in the first participant and 68 percent across four classes in the second when using an optimized double-encoding scheme. While these numbers may appear modest in absolute terms, they represent statistically significant improvements over conventional decoders and were achieved with data from a single nerve branch.

Combining intraneural signals with residual muscular activity from the same electrodes boosted performance further, reaching 64 percent for the first participant and 73 percent for the second. The researchers also found that adding more electrodes improved accuracy in a near-linear fashion, with a Pearson correlation of 0.93 in the first participant, suggesting that scaling up the number of implants could yield substantially better results.

A particularly significant finding is that the motor and sensory maps recorded from the nerve showed minimal overlap, with as little as 7 percent spatial intersection between motor-responsive and sensory-responsive channels. This early segregation within the sciatic nerve means the same implant could, in principle, provide both natural neural control and sensory feedback simultaneously without the two functions interfering with each other.

“When you tell your body to move, signals travel through nerves to muscles, even if the limb is gone. The challenge is extracting that information and understanding the neural code behind it,” said Giacomo Valle, assistant professor at Chalmers and the study’s senior author. The team’s approach of using spiking neural networks to interpret the nerve’s native signaling language, rather than forcing continuous-value abstractions onto inherently spike-based data, appears to be a key factor in the decoder’s superior performance.

What We Don’t Know

The study has several acknowledged limitations. With only two participants and a 90-day implantation period, long-term stability and generalizability remain unestablished. All decoding was performed offline; whether the system can achieve real-time performance sufficient for walking remains untested. The researchers also noted that foreign body response was observed at the implantation site upon electrode removal, raising questions about chronic implant viability.

The accuracy figures, while statistically superior to baselines, would need to improve substantially before clinical deployment. Whether scaling from four to more electrodes, or implanting in additional nerve branches, can close the performance gap is an open question. The study also cannot objectively verify the timing of phantom movements, since there is no proprioceptive feedback to confirm when participants actually initiated each attempt.

Analysis

Most prosthetic legs in clinical use today rely on pre-programmed movement patterns triggered by mechanical sensors or surface electromyography. Direct neural control has been demonstrated for upper-limb prosthetics, including by previous work from Chalmers’ own Biomechatronics and Neurorehabilitation Laboratory, but lower-limb applications have lagged behind. The challenge is partly anatomical: above-knee amputations sever the leg’s motor nerves far from their original targets, making signal extraction difficult.

This study’s contribution is showing that meaningful motor information persists in the proximal sciatic nerve even after above-knee amputation, and that a bio-inspired decoder can extract it. The bidirectional capability of the implant, enabling both motor readout and sensory stimulation through the same device, could simplify the engineering challenge of building a fully neurally integrated prosthetic leg.

The research was funded by the European Research Council under the FeelAgain grant, and the team has made both de-identified participant data and analysis code publicly available. The next step, according to the researchers, is integration with an actual prosthetic leg for real-world testing.