Loading
Old version — revision 3
This is a fixed snapshot of Neuroprosthetics, saved by Import. It is not edited and it is not updated; the article may have changed since.
Edit summary: Re-import of content/neuroprosthetics.md — the file changed after the initial import. Not an edit.
Read the current versionWhat this revision changedAll 3 revisions
The engineering field that builds devices interfacing directly with the nervous system to restore lost sensory, motor, cognitive, or autonomic function.
Neuroprosthetics is the engineering discipline concerned with devices that exchange signals directly with the nervous system in order to replace a function the nervous system has lost. It spans devices in routine clinical use, such as the Cochlear implant, the spinal cord stimulator and Deep brain stimulation, as well as devices that exist only in single-patient research studies, such as cortical arrays that restore a sense of touch. The problem that unites them is the interface rather than the electronics: how to place a durable transducer against excitable tissue that reacts to its presence, and how to encode information in a form neurons will accept.
Four families are usually distinguished, though real devices increasingly cross the boundaries.
Sensory prostheses convert an external stimulus into neural activity. The cochlear implant and the Retinal implants and visual prostheses are the archetypes, and vestibular implants for bilateral vestibular loss have reached small human trials. These devices have to solve an encoding problem, because electrodes cannot reproduce the natural code of the sense organ.
Motor prostheses read intent and act on it. Cortical systems for people with paralysis fall under Brain–computer interface, and peripheral systems that read muscle activity fall under Myoelectric prosthetics. Functional electrical stimulation drives the user's own paralysed muscles or spinal circuitry instead of an external actuator.
Autonomic and organ-directed devices are the largest category by number and receive the least attention. Sacral neuromodulation for bladder dysfunction, vagus nerve stimulation for epilepsy and for rehabilitation after stroke, phrenic pacing for ventilator dependence, and spinal cord stimulation for chronic pain together account for far more implants than every cortical device combined.
Cognitive prostheses attempt to restore a computation rather than a signal path. The only serious example is the hippocampal Memory prosthesis, which remains at the proof-of-concept stage.
Every device is a loop with four parts, and each part fails in its own way.
The transducer couples electrical or optical energy to tissue. The options range from penetrating microelectrode arrays such as the Utah array, through the surface arrays used in Electrocorticography interfaces, to nerve cuffs, intrafascicular electrodes, and the endovascular approach of the Stentrode. Closer coupling yields more information per channel and more tissue damage.
The front end amplifies microvolt signals in the presence of stimulation artefacts thousands of times larger, within a power budget that keeps the surrounding tissue from heating by more than a fraction of a degree.
The algorithm converts signals into commands, or sensor data into stimulation patterns. On the read side this is Neural decoding, which has improved faster than the hardware. On the write side it is an encoding problem with no comparable body of theory.
The actuator is a robotic limb, a synthesized voice, a stimulator on the user's own spinal cord, or a pattern of current in a sensory nucleus. Systems that both read and write are called bidirectional, and the field's hardest problems are concentrated in them.
The clinically settled devices restore hearing, suppress the symptoms of movement disorders, manage pain, and control bladder and bowel function. They are unglamorous, well reimbursed in wealthy countries, and are improving incrementally.
The research frontier is bidirectional restoration after paralysis and amputation, and three results define it. Intracortical microstimulation of human somatosensory cortex produces tactile sensations that a participant localizes to specific fingers of a paralysed hand, and adding that feedback roughly halved the time needed to complete an object-transfer task with a robotic arm.12 Implanted electrodes in residual peripheral nerves evoke graded touch in amputees, and combined with skeletal attachment through Osseointegration they have supported prostheses used at home for years.3 Epidural spinal stimulation, targeted to the dorsal roots that recruit specific leg muscle groups, restores stepping in people with severe spinal cord injury.4 A cortical implant wired to such a stimulator gave one participant volitional, thought-driven walking, and produced neurological recovery that persisted with the system switched off.
Each of these involves single-digit numbers of participants. A device cleared in the United States in late 2024 delivers non-invasive transcutaneous spinal stimulation to improve hand and arm strength after cervical injury, and it is the first of this generation to reach the market.
Reading intent out of the nervous system is comparatively tractable, because a decoder can be trained on labelled attempts and improved offline without touching the tissue. Writing information in is harder, for structural rather than temporary reasons.
An electrode excites every axon within its field indiscriminately, including axons of passage that belong to unrelated circuits. Natural sensory input arrives as spatiotemporally patterned activity across specific cell types. Electrical stimulation instead activates a sphere of tissue synchronously, a pattern the brain has never encountered and has no code for. Percepts evoked this way can be described but feel unnatural: pressure, tingling or a vague spot of light, rather than texture, temperature or shape. Repeated stimulation causes adaptation, so amplitude has to be modulated to keep a percept stable. Every stimulation pulse also saturates the recording amplifiers, which is why bidirectional devices need artefact-rejection schemes that discard part of the signal they were built to collect.
Two approaches attack the specificity limit. Biomimetic stimulation shapes pulse trains to imitate the dynamics of natural afferent firing instead of delivering constant-amplitude trains, which makes evoked touch more natural. Optogenetics achieves genuine cell-type specificity in animals but requires gene delivery and implanted light sources, and no human neuroprosthesis uses it.
Why this asymmetry mattersA motor prosthesis that reads well and writes badly is still useful, because vision substitutes for missing touch. A sensory prosthesis has nothing to compensate for a poor write channel. That is why hearing and vision restoration have plateaued at low information rates while motor decoding has kept improving.
Chronic implants fail in predictable ways. Penetrating electrodes provoke a foreign-body response: microglia activate, astrocytes encapsulate the shank, and neurons retreat from the recording tip, so signal yield declines over months to years. Insulation absorbs water and delaminates, thin metal traces corrode, and hermetic feedthroughs leak. Percutaneous connectors, still standard in academic systems, are an infection route and tether the user to a laboratory.
Three further limits recur beyond the hardware. Decoders drift as the recorded population changes and need recalibration. Almost all headline performance is measured with a technician present in a controlled setting, and it degrades outside that setting. Finally, the number of independent control or perceptual dimensions a person can use saturates well below the electrode count, for the same reason in every modality: current spread makes nearby channels non-independent.
Implanted neurotechnology raises questions that ordinary medical devices do not. Recorded neural activity supports inferences the user did not intend to disclose, the concern formalized under Mental privacy and addressed in legislation under Neurorights. Devices that modulate mood or motivation raise questions about who authors an action. A device implanted for therapy in a person who cannot easily have it removed also creates an asymmetric dependence on the manufacturer's continued existence. That exposure became concrete when retinal implant support was withdrawn from several hundred patients.
Access follows wealth rather than need. Cochlear implantation rates differ by more than an order of magnitude between high- and low-income countries, and every newer device costs more. The distributional pattern examined in Access and inequality is already the current state of the most successful neuroprosthesis ever built, not a hypothetical.
The point where restoration shades into Human enhancement is contested rather than settled. Devices such as those catalogued in Sensory augmentation use the same interfaces to add capacities rather than restore them. Speculative uses such as Brain-to-brain interfaces communication borrow the vocabulary of clinical neuroprosthetics while resting on far weaker demonstrations.
Two trends look likely to shape the next decade. Materials and packaging are improving faster than algorithms. Thin-film flexible arrays, carbon-fibre electrodes and fully implanted wireless systems address the specific failure modes that have limited chronic recording, and commercial programmes at Neuralink, Synchron and others are funding that engineering at a scale academic laboratories never could. Miniaturized wireless nodes of the kind described under Neural dust and ultrasonic implants would change the placement calculus if they can be made to work at cortical depth. Further from the clinic, bio-interfacing artificial neurons aim to make the stimulating electronics themselves fire like nerve cells. Printed devices of this kind have stimulated Purkinje neurons in slices of mouse cerebellum, but none has been reported implanted in a living brain.5
The competing trajectory is biological. The regenerative approaches covered under Limb regeneration, along with cell replacement therapies, would remove the need for an interface entirely in some indications, and external assistive devices such as the Powered exoskeletons avoid implantation altogether. Which path prevails depends on an unresolved empirical question: whether the write-channel problem yields to better encoding and better electrodes, or whether artificial stimulation is permanently limited to coarse, unnatural percepts that the brain can learn to use but never to interpret as its own.
paperFlesher, S. N. et al. "Intracortical microstimulation of human somatosensory cortex." Science Translational Medicine, 2016.↩One participant with a spinal cord injury; the percepts were localized to individual fingers, and the outcome is what he reported feeling rather than a measured functional gain.
paperFlesher, S. N. et al. "A brain-computer interface that evokes tactile sensations improves robotic arm control." Science, 2021. ↩
paperOrtiz-Catalan, M. et al. "Self-contained neuromusculoskeletal arm prostheses." New England Journal of Medicine, 2020.↩A brief report on a small series of arm amputees using the system in daily life, with no control group and no randomization.
paperWagner, F. B. et al. "Targeted neurotechnology restores walking in humans with spinal cord injury." Nature, 2018. ↩
paperHadke, S. S. et al. "Printed MoS2 memristive nanosheet networks for spiking neurons with multi-order complexity." Nature Nanotechnology, 2026.↩Stimulation of Purkinje neurons in mouse cerebellar slices; no living animal or implanted device.