Prosthetics & Rehabilitation Robotics

1. Scope and field map

This note covers the full chain from limb-loss replacement to neurological recovery: upper- and lower-limb prosthetics, sensory implants (cochlear, retinal), brain-machine interfaces (BMIs) for motor restoration, exoskeletons for spinal-cord injury (SCI) and stroke rehabilitation, end-effector robotic therapy, and the surrounding clinical, regulatory, and reimbursement scaffolding. It is the clinical neighbour of exoskeletons, surgical-robotics, teleoperation-haptics, and the existing sibling prosthetics (which focuses narrowly on myoelectric hands and osseointegration).

Three intertwined disciplines converge here:

• Biomechanics and gait science — Perry’s gait phases, joint loading, musculoskeletal modelling (biomechanics).
• Neural interfacing — surface EMG through implanted electrodes through intracortical arrays, paired with stimulation for sensory return (bioinstrumentation).
• Mechatronic device design — actuators, compliant sockets, microprocessor control, embedded inference (motors-electric, impedance-control, realtime-embedded).

Roughly 60 million people worldwide live with a major limb amputation or congenital limb difference (WHO/ISPO 2022 estimate). Of these only an estimated 10–15% have access to any prosthesis, and far fewer to a powered one — the global disparity drives ICRC norms and parallel low-resource design tracks (ReMotion, Jaipur Foot) alongside high-end Western platforms.

The clinical workflow that wraps every device decision is itself a system: prescription (physiatrist / surgeon), socket fabrication (prosthetist / orthotist), control fitting (clinical engineer), training (occupational / physical therapist), follow-up and revision. Engineering choices ripple across all five — a hand that adds 200 g may save 20% classification accuracy but reduce hours-of-wear by 30% if the user finds it too heavy. Outcome is always the integral of capability over time-in-use, not peak benchmark capability.

1.1 Historical arc

• Antiquity — wooden / iron toe and leg replacements (Egyptian Cairo toe, 950 BCE); Roman bronze Capua leg (300 BCE).
• 16th c. — Ambroise Paré’s articulated mechanical hand (1564); Götz von Berlichingen iron hand (1504).
• 1816 — James Potts’s “Anglesey leg” — articulated wood and catgut tendons; ancestor of modern transfemoral architecture.
• 1860s — American Civil War drives prosthetics industry; Hanger founded 1861 by amputee James Hanger.
• 1948 — Reinhold Reiter, Germany, first myoelectric hand prototype.
• 1960s — Soviet Central Institute of Prosthetics demonstrated multi-channel myoelectric control; Otto Bock Z6 commercial myoelectric hand 1968.
• 1969 — Brånemark coins “osseointegration.”
• 1984 — Van Phillips patents Flex-Foot (Cheetah lineage).
• 1997 — Ottobock C-Leg, first commercial microprocessor knee.
• 2004 — Kuiken TMR; 2006 BrainGate first cursor; 2007 i-Limb first multi-articulating hand.
• 2014 — DEKA / LUKE Arm FDA cleared; Raspopovic intraneural TIME stimulation.
• 2017–2023 — Speech BCIs, Neuralink first-in-human, PSYONIC L7259 reimbursement, OPRA HDE.

2. Upper-limb prosthetics

2.1 Three control paradigms

Paradigm	Mechanism	Pros	Cons
Body-powered (BP)	Cable from harness driven by scapular protraction; opens/closes hook or hand	Robust, low maintenance, tactile via cable tension, cheap	Cosmetically obvious harness, limited DoFs, fatigue
Myoelectric (single-site / DC)	Two antagonist sEMG sites drive open/close on one DoF; co-contraction switches grip	Cosmesis, less harness	One DoF at a time, heavier, expensive
Pattern-recognition (PR)	6–8 sEMG channels classified into postures via LDA/SVM/CNN	Multiple grips, intuitive, regression for proportional control	Drift, limb-position effect, training burden

Body-powered hooks (Hosmer Dorrance 5XA, TRS Grip 2/3) remain the most-used active terminal device worldwide because of cost (~$5–10k fitted), durability, and inherent proprioception through cable feedback (Biddiss & Chau 2007). The split between BP and myoelectric in US transradial users is roughly 40/60 by some surveys, with body-powered persisting strongly among heavy-duty occupations.

2.2 Targeted Muscle Reinnervation (TMR)

Pioneered by Todd Kuiken and Greg Dumanian at the Rehabilitation Institute of Chicago (RIC, now Shirley Ryan AbilityLab), TMR (Kuiken et al., RIC, 2004) reroutes the residual median, ulnar, radial, and musculocutaneous nerves onto segments of remaining muscle (pectoralis major segments for shoulder-disarticulation, biceps/triceps for transhumeral). After 3–6 months of reinnervation, surface EMG over each muscle segment expresses an independent volitional signal, effectively converting a proximal amputation into something closer to a transradial control problem. TMR also reduces neuroma-related pain in roughly 70% of cases (Dumanian 2019, JAMA Surgery) and is now standard of care at major US centres.

Regenerative Peripheral Nerve Interfaces (RPNI) — Cederna and Kemp (Michigan, 2014 onward) — wrap a transected nerve fascicle in a free muscle graft, producing a small reinnervated muscle target that amplifies the nerve signal for sEMG recording or implantable electrode capture. RPNIs are paired with surgical AMI procedures (see below) in modern bionic arms.

Agonist-Antagonist Myoneural Interfaces (AMI) — Hugh Herr’s group at MIT (Srinivasan et al. 2017, 2020) preserve a mechanical agonist-antagonist pair across the residual limb by surgically connecting opposing muscle bundles via a tendon; voluntary contraction in one stretches the other, giving the user a sense of joint state (proprioception) plus high-SNR EMG. Applied in both above-knee and below-elbow amputations; partnered with osseointegration in the “Ewing amputation” technique (Carty et al., 2018, Sci Transl Med).

2.3 sEMG signal acquisition

Surface EMG is the workhorse signal. Typical fitting uses 8–16 dry stainless or Ag/AgCl wet electrodes embedded in the socket inner wall over forearm extensors/flexors. Signal properties:

• Amplitude 10 µV – 5 mV; bandwidth 10–500 Hz usable.
• Front-end: instrumentation amplifier (TI INA333, INA826 or Analog Devices AD8232 / ADS1299) with CMRR > 100 dB at 60 Hz, programmable gain ~1000, then bandpass 20–450 Hz and 50/60 Hz notch.
• ADC 16-bit at 1–2 kSPS; 24-bit ΔΣ on multichannel (ADS1299 8-ch at 16 kSPS).
• Per-channel power < 5 mW for a wearable.

High-density sEMG (HD-sEMG) — 64 to 256+ electrode grids (OT Bioelettronica Quattrocento, Delsys Galileo) record spatially extended muscle activity, enabling motor-unit decomposition (Holobar & Zazula 2007 convolutive blind source separation). HD-sEMG drives the next generation of decoders by recovering individual motor-unit firing trains instead of bulk envelopes — Farina et al. (2017, Nat Biomed Eng) showed neural-drive decoders that outperform envelope-based PR.

Implanted Myoelectric Sensors (IMES) — Alfred Mann Foundation (Weir et al. 2009), commercialised through Sigenics. Each IMES is a 2 mm × 16 mm capsule placed surgically in target muscle, communicating inductively at 6.78 MHz with an external coil in the socket. iEMG bandwidth extends to 2 kHz, SNR 30–50 dB, no electrode-shift artefact. Used in DARPA Revolutionizing Prosthetics; investigational under FDA IDE.

2.4 Pattern recognition and decoding

Englehart-Hudgins time-domain (TD) feature set (2003) is still the clinical baseline:

• Mean Absolute Value (MAV): $\text{MAV}=\frac{1}{N}{\sum }_{i=1}^N|x_i|$
• Waveform Length (WL): $\text{WL}={\sum }_{i=2}^N|x_i-x_{i-1}|$
• Zero Crossings (ZC) with 10 µV deadband
• Slope Sign Changes (SSC)

Window 150–250 ms with 50–75% overlap; 300 ms total latency budget before user perceives lag. Linear Discriminant Analysis (LDA) over 8 channels × 4 features = 32-D feature vector classifying 5–7 grips achieves 85–95% accuracy in able-bodied subjects, 70–80% in transradial amputees.

Deep learning decoders — Atzori et al. (Ninapro 2014, Sci Data 2016) released a 27-subject sEMG database that catalysed CNN/LSTM/transformer decoders. Côté-Allard et al. (2019, IEEE TNSRE) showed transfer-learnable CNNs cutting calibration time. Modern hands run a quantised CNN (50k params) on a Cortex-M7 (STM32H7) under TensorFlow Lite Micro at 30 Hz inference rate, fitting in < 32 kB RAM.

Simultaneous proportional control — regression schemes (Hahne 2014, IEEE TNSRE) map EMG features continuously to joint velocity rather than discrete posture choices, enabling simultaneous wrist + grip motion. Used in modern Coapt and Infinite Biomedical clinical systems.

2.5 Clinical PR controllers

Product	Vendor	Approach	Notes
CompleteControl Gen2	Coapt LLC (Chicago)	LDA + adaptation, calibrate by mimicking grips	First FDA-cleared PR controller (2013); pairs with i-Limb, bebionic, Hero Arm
Sense	Infinite Biomedical Technologies (IBT)	LDA with onboard adaptation	Direct competitor; preferred by some VA centres
AxonSoft	Ottobock	DC + grip switching for bebionic / Michelangelo	Not PR per se but mature DC tooling
Symphonie Aktiv	Romedis	DC + simple two-DoF switching	Conservative European fitting

2.6 Multi-articulating hands

Device	Vendor	Active DoFs	Mass	Grip force	Price (USD)	Notes
i-Limb Quantum	Touch Bionics / Össur	6	432–536 g	100 N	$30–70k	36 grip patterns; gesture (i-Mo) switching
bebionic Gen3	Ottobock	6	495–598 g	140 N	$30–50k	Industrial robustness
Michelangelo	Ottobock	2 (thumb + coupled 4-finger)	420 g	70 N	$40–60k	Anatomical thumb opposition
Ability Hand	PSYONIC (Akhtar 2023)	6	490 g	50 N pinch / 145 N power	~$25k	Compliant fingertips, force feedback, Medicare L7259
Hero Arm	Open Bionics (UK)	6	280–340 g	8 kg lift	~$10k	3D-printed sockets; “covers” (Iron Man, Star Wars, Frozen) under Disney licence; paediatric leader
Vincent Evolution 4	Vincent Systems (Germany)	6	280 g	60 N	$25–40k	Smallest multi-articulating
COVVI Nexus	COVVI (UK)	6	538 g	110 N	~$30k	Bluetooth fitting app, magnetic coupler
Esper Hand	Esper Bionics (Ukraine/US)	6	380 g	80 N	$25–35k	Cloud-connected adaptive learning
NeoMano	BrainCo (China/US)	2 (grasp assist)	80 g (glove)	8 kg	$5–8k	Lightweight orthotic glove, not full prosthesis
LUKE Arm	Mobius Bionics (DEKA / DARPA RP2009)	10 powered (shoulder–wrist)	4.8 kg full	grip 90 N	$100–200k	Foot-IMU and switch control; FDA cleared 2014
Modular Prosthetic Limb (MPL)	JHU/APL (research)	26 DoF (17 active)	4.7 kg	70 N	research	Used in Collinger 2013 cortical control
Atom Touch	Atom Limbs	14 DoFs	~3.5 kg	grip 90 N	est $20k (target)	Early-2026 IDE pilot; consumer ambition

2.7 Shoulder-disarticulation and transhumeral

The proximal amputee faces a controllability cliff: only chest/shoulder EMG remains for direct control. Solutions:

• TMR + PR (Kuiken/Hargrove) — converts shoulder-level into transradial-like signal space.
• DEKA/LUKE Arm — foot-IMU shoes provide additional control DoF; user pitches/rolls feet to drive shoulder and elbow modes.
• MPL + cortical BCI — research-grade, Collinger 2013 Lancet demonstrated 7-DoF reach + grasp from M1 microelectrode arrays in a tetraplegic.

3. Lower-limb prosthetics

3.1 Gait phases

Perry’s 1992 division remains canonical: stance (initial contact, loading response, mid-stance, terminal stance, pre-swing) and swing (initial swing, mid-swing, terminal swing). Stance is ~60% of cycle; swing ~40%. Joint torques peak at heel strike and toe-off; ground reaction force peaks ~1.2× body weight at heel strike on level ground, 2–3× during running.

3.2 Passive feet

Device	Vendor	Type	Push-off energy	Mass	Notes
SACH foot	many (Otto Bock, Kingsley, Hangar)	Solid Ankle Cushion Heel	0% return	400 g	K1–K2 ambulators; cheapest
Single-axis foot	Ottobock 1A30	Hinged ankle	~5% return	450 g	Smoother heel-strike than SACH
ESAR foot	Össur Vari-Flex, Ottobock 1C30 Trias	Energy-storing carbon	25–35% return	350–600 g	K3 community walker
Flex-Foot Cheetah	Össur (Van Phillips 1984 patent)	Sprinting C-blade	up to 90% return	270 g	Paralympic standard; not for walking
Pro-Flex	Össur	Three-blade carbon	30% return; 27° dorsiflexion	685 g	Improved push-off vs Vari-Flex
Renegade	Freedom Innovations	Carbon ESAR	30%	500 g	Active K3–K4

ISO 10328 governs lower-limb structural strength testing — three load levels (P3, P4, P5) at body masses 60, 80, 100 kg with cyclic 2 million-cycle fatigue plus ultimate static load. The standard is mandatory for CE and required by most insurance plans before reimbursement.

3.3 Microprocessor knees (MPK)

Sensor cycle 20–50 ms; finite-state machine over knee angle θ_k, angular velocity ω_k, axial load F_z (4-element strain-gauge bridge in pylon), shank IMU.

Device	Vendor	Mass	Battery	Notable feature
C-Leg 4	Ottobock (1997, ref. standard)	1235 g	45 h	Hydraulic damping, first MPK
Genium	Ottobock	1450 g	5 days	Stumble recovery; stair ascent in alternating steps
Genium X3	Ottobock (DoD / Walter Reed)	1480 g	5 days	IP68 saltwater; runs
Rheo Knee XC	Össur	1610 g	36 h	Magnetorheological fluid damper
Power Knee	Össur	2700 g	1 day	Powered swing initiation, only fully active commercial knee
Plié 3	Freedom Innovations	1170 g	3 days	Fastest sensor cycle (10 ms)
Linx	Endolite/Blatchford	1450 g (knee+ankle)	3 days	Integrated knee + ankle control
Kenevo	Ottobock	1500 g	2 days	K2-ambulator MPK; safer-fall logic

Damping control example (C-Leg): stance flexion damping starts at 4 N·m·s/rad, drops linearly to 1 N·m·s/rad through mid-stance, rises in extension before heel strike to prevent hyperextension. Swing damping is tuned to user cadence. Hydraulic valve actuated by stepper at 100 Hz.

Reimbursement caveat: in the US, Medicare’s K-level system (K0 = no ambulation, K4 = athletic) gates MPK access — only K3 or higher qualify. The VA covers MPKs more liberally; in single-payer systems coverage is uneven (UK NHS only began wider C-Leg coverage in 2016).

3.4 Powered ankles and knees

Active devices add motors that perform net positive work, not just dissipate it.

• BiOM / Ottobock Empower — Hugh Herr’s MIT spin-out (2007), acquired by Ottobock 2017. Series-elastic actuator drives ankle dorsiflexion/plantarflexion; net positive work ~10 J per stride, matching biological ankle push-off. Battery 3000 steps; mass 2.1 kg.
• Össur Power Knee — first commercial powered knee (2006, 2018 redesign); motor + harmonic-drive at knee axis; assists stand-to-sit and stair ascent.
• Össur Proprio Foot — semi-active microprocessor ankle adjusting dorsiflexion angle for slope; not net positive work but adaptive.
• Open-source ankle — Open Source Leg (OSL, Rouse et al., U Michigan 2019) — full schematic + firmware available for research; used in 10+ academic labs.

The powered-ankle clinical claim is reduced metabolic cost (Herr & Grabowski 2012, PRSB: ~14% lower than passive ESAR for transtibial K3 walkers) plus more natural gait kinematics; longevity, weight, and noise have slowed adoption.

3.5 Osseointegration

Brånemark Sr.’s 1969 discovery of titanium osseointegration (originally dental) extended to limbs by Rickard Brånemark (Integrum, Gothenburg). Three commercialised systems:

System	Origin	Implant geometry	Surgical stages	Status
OPRA	Integrum (Sweden)	Threaded Ti-6Al-4V fixture, screw-in abutment	Two-stage (6 months apart)	FDA HDE 2020 (transfemoral)
ILP / Endo-Exo	ESKA / Orthodynamics (Germany)	Press-fit porous Ti, intramedullary	Single stage	CE; widely used in DE
POP (Percutaneous Osseointegrated Prosthesis)	DJO / Utah	Press-fit Ti with hydroxyapatite coating	Single stage	FDA IDE; Utah veterans cohort

Advantages over socket: direct skeletal load transfer, eliminates socket sweat/wounds, restores osseoperception (limb sense via bone vibrations), increased prosthesis-use time from 60% to >95% of waking hours, dramatic Q-TFA quality-of-life gains (Brånemark et al. 2014, Bone Joint J., 51-patient series).

Risks: deep infection 2–6%, superficial stoma irritation 30–50% lifetime, fixture loosening < 5% with proper rehab. Safety abutment engineered to fracture at ~80 N·m bending to prevent femoral fracture in falls. ISO 14801 governs implant fatigue testing.

3.6 Socket design

Even with osseointegration on the rise, > 95% of users still rely on a socket. Design lineage:

• Patellar-Tendon-Bearing (PTB) transtibial socket (Radcliffe & Foort 1961, UC Berkeley) — load-bearing on patellar tendon and posterior soft tissue.
• Total Surface Bearing (TSB) — distributes load over residual limb evenly; needs precise fit.
• Quadrilateral / Ischial Containment transfemoral — Sabolich-Bertschi geometry, contains ischium for proximal stability.
• MAS (Marlo Anatomical Socket) — Ortiz 2002 — high-mobility transfemoral; narrower medio-laterally.

Suspension:

• Suction (pin-lock with silicone liner Alpha SmartTemp, Iceross Comfort).
• Elevated vacuum (Harmony, Ottobock; LimbLogic, OPL) — subatmospheric pressure 5–20 kPa via mechanical or electronic pump; reduces residual-limb volume fluctuation, improves proprioception.
• Self-suspending bracketless designs (Pin Lock 6, Locking Lanyard).

Liners: silicone gel (Alpha SmartTemp, Iceross Comfort Air), polyurethane (Ottobock Skinguard), thermoplastic elastomer (Aegis Plus). Liner durometer 15–35 Shore A.

CAD/CAM fabrication: scan residual limb with 3D scanner (Standard Cyborg, Structure Sensor, Vorum CANFIT, Hangar Insignia in-clinic platform); modify in CAD; CNC-carve foam positive or directly 3D-print check socket. Hangar Insignia (rolled out 2018) is the largest US in-clinic CAD/CAM workflow. Definitive sockets are typically carbon-fibre layup over the corrected positive.

3.7 ICRC and low-resource prosthetics

The International Committee of the Red Cross (ICRC) and ISPO publish norms for polypropylene-thermoformed sockets and modular feet for use in low- and middle-income countries. Jaipur Foot (BMVSS, India, 1968 onward) and ReMotion knee (D-Rev, 2014) target sub-$100 component cost. Trade-offs: durability + serviceability prioritised over kinematic refinement.

3.8 Worked example — gait state machine for transfemoral MPK

Inputs: knee-flexion θ_k (encoder), knee angular velocity ω_k, axial pylon load F_z (strain-gauge bridge), shank IMU (3-axis accel + gyro).

Sensor cycle: 50 Hz, 20 ms loop period; worst-case latency to damping change 30 ms.

State transitions (simplified Ottobock-style):

From	To	Trigger
Mid-stance	Pre-swing	F_z < 0.3·BW AND θ_k > 5° AND ω_k > 0
Pre-swing	Swing	F_z < 0.05·BW
Swing	Terminal-swing	ω_k changes sign
Terminal-swing	Initial contact	F_z > 0.3·BW

Damping schedule: stance flexion damping 4 N·m·s/rad at heel-strike, decays linearly to 1 N·m·s/rad through mid-stance, extension damping rises to 5 N·m·s/rad near terminal swing to prevent hyperextension. Swing damping is tuned offline to user cadence (cadence-following PID).

Stumble recovery (Genium): if ω_k drops below threshold during swing (toe catches), valve clamps to high flexion damping immediately; gyro cross-check on shank IMU confirms perturbation before triggering, reducing false-positive locking.

3.9 Worked example — osseointegration mechanical loading

Implant: OPRA fixture, Ti-6Al-4V, 100 mm length, 12 mm OD intramedullary section, screw-threaded for primary stability. Abutment screws into fixture, protrudes 30 mm through skin.

Load case: 80 kg transfemoral user, stance-phase peak ground reaction force ~1.2 × BW = 942 N axial. Bending moment at bone-implant interface on uneven heel strike with 50 mm moment arm: M ≈ 47 N·m.

Stress check: σ_bend = M·c/I; hollow Ti rod (12 mm OD, 6 mm ID) gives I = π(12⁴ − 6⁴)/64 ≈ 942 mm⁴, c = 6 mm. σ ≈ (47 000 N·mm · 6 mm) / 942 mm⁴ ≈ 300 MPa — well below Ti-6Al-4V yield (~830 MPa). The safety abutment is the planned fuse: fractures at ~80 N·m to protect the bone-implant interface in falls. Bone remodelling (Wolff’s law) consolidates over 6–12 months per the Integrum two-stage rehab protocol; load progression follows weekly milestone schedule.

4. Sensing for prosthetic control

4.1 Peripheral nerve interfaces (PNI)

A spectrum from least to most invasive:

Interface	Origin	Geometry	Channels	Selectivity	Chronic stability
Cuff (epineurial)	Naples, GA Tech	Wraps nerve trunk	2–16	Low (fascicle group)	High; years
Flat Interface Nerve Electrode (FINE)	Tyler, CWRU	Flattens nerve, multi-contact	8–16	Medium	High; > 5 yr clinical
Longitudinal Intrafascicular (LIFE)	Yoshida, Lawrence	Wire through fascicle, parallel	4	High	Months
Transverse Intrafascicular Multichannel (TIME)	Micera (EPFL/SSSA)	Penetrating ribbon, 14 sites	14–56	Very high	1–3 months human
Utah Slanted Electrode Array (USEA)	Normann, U Utah	100-microelectrode array, graded depth	100	Very high	1–6 months
Regenerative sieve	Akin, Edell	Nerve regrows through perforated chip	varies	Single-fibre potential	Limited

The TIME electrodes from Silvestro Micera’s group (Lausanne/SSSA) enabled the LifeHand 2 (Raspopovic et al. 2014, Sci Transl Med) and SensArs trials — multi-fascicle intraneural stimulation delivering graded, position-localised touch percepts.

4.2 Sensory feedback modalities

Modality	Mechanism	Latency	Resolution	Invasiveness
Vibrotactile (Tan group, etc.)	ERM/LRA motor on intact skin	20 ms	Coarse (1–4 chan)	Non-invasive
Electrotactile	Transcutaneous current pulses	5 ms	Medium (8 chan)	Skin only
Mechanotactile	Servo push-pin on skin	50 ms	Localised	Non-invasive
Skin stretch	Tangential motor on skin	30 ms	Direction cues	Non-invasive
Targeted Sensory Reinnervation (TSR)	Hand nerves rerouted to skin patch	n/a (always-on)	Touch on chest patch felt as “hand touch”	Surgical
Intraneural (FINE, TIME)	Direct fascicle stimulation	1 ms	Fine; percept-mapped	Implant
DRG stimulation	Dorsal root ganglion electrodes	5 ms	Dermatome-mapped	Implant
Cortical (ICMS in S1)	Intracortical microstimulation in somatosensory cortex	1 ms	High; finger-region mapped	Intracortical

Flesher et al. (2016, Sci Transl Med) demonstrated naturalistic touch via Utah-array ICMS in S1; Ganzer et al. (2020) closed the loop with motor cortex BMI. Vibrotactile remains the most-deployed non-invasive option in commercial hands (PSYONIC Ability Hand ships with closed-loop vibrotactile to the residual limb).

4.3 Worked example — sEMG classification pipeline

Setup: 8-channel sEMG (Delsys Trigno wireless, or Coapt embedded), 1 kHz sample rate, 200 ms window, 100 ms increment.

Per-channel features (Hudgins TD set):

• MAV: $\text{MAV}=\frac{1}{N}{\sum }_{i=1}^N|x_i|$
• WL: $\text{WL}={\sum }_{i=2}^N|x_i-x_{i-1}|$
• ZC (zero crossings, 10 µV deadband)
• SSC (slope sign changes)

8 channels × 4 features = 32-dimensional feature vector per window. LDA classifier trained on 5 postures (power, pinch, tripod, lateral, hook) + rest.

Typical performance:

Metric	Able-bodied	Transradial amputee
Classification accuracy	85–95%	70–80%
Time-to-grip	250 ms	350 ms
False-trigger rate	3–5%/min	8–15%/min

Clinical-grade decoders post-process with a majority-vote smoother over 3 windows to reduce false transitions and impose a minimum dwell time (~200 ms) before accepting a new posture.

4.4 Limb-position effect and adaptation

EMG feature distributions shift with arm posture due to gravity-induced co-activation and electrode shift relative to underlying muscle. Fougner et al. (2011, IEEE TNSRE) quantified ~15–25% PR accuracy drop and showed adding an IMU as classifier input restores most of the loss. Modern systems train across postures and adapt online (Coapt CCE incremental adaptation, Hahne 2017).

4.5 Embedded inference

Modern multi-articulating hands run real-time classification on a Cortex-M7 SoC (STM32H7 at 480 MHz with FPU, or NXP i.MX RT) for low-power local inference; higher-end designs (LUKE Arm) use Cortex-A53 with embedded Linux. LDA + sliding-window vote fits in < 10 kB RAM with sub-1 ms inference; quantised CNNs (Atzori 2016 Ninapro CNN at ~50k params) run at 30 Hz on the same hardware under TensorFlow Lite Micro. Power budget: < 200 mW continuous for the controller, with the actuators dominating overall consumption. See realtime-embedded and microcontrollers.

5. Brain-machine interfaces

5.1 Signal modality and bandwidth

Modality	Spatial resolution	Bandwidth (Hz)	Channel count	Invasiveness
EEG (scalp)	cm	0.5–50	8–256	Non-invasive
MEG	cm	0.5–100	100–300	Non-invasive (shielded room)
fNIRS	cm	0.01–0.1	8–64	Non-invasive
ECoG (subdural)	mm	0.5–200	64–256	Craniotomy
µECoG / HD-ECoG	sub-mm	1–500	256–1024	Craniotomy
Intracortical microelectrode (Utah, Neuralink, Paradromics)	< 100 µm; single unit	DC–10 kHz	96–8192	Cortical penetration
Endovascular (Stentrode)	mm	1–200	16	Catheter-deployed

5.2 BrainGate consortium (Hochberg, Donoghue, Henderson)

Open clinical study of the Blackrock 96-channel Utah array in M1 of tetraplegic participants since 2004. Milestones:

• 2006 (Hochberg, Nature 442) — first human cursor control by neural ensemble.
• 2012 (Hochberg, Nature 485) — reach + grasp of DLR-LWR robotic arm to drink coffee.
• 2013 (Collinger, The Lancet) — 7-DoF MPL control by C5-C6 tetraplegic (Jan Scheuermann).
• 2017–2021 — speech decoding via 128-channel ECoG (Edward Chang, UCSF) and Utah-array intracortical (Krishna Shenoy / Frank Willett, Stanford).
• 2023 (Willett et al., Nature) — Utah-array speech BCI at 62 words/min from motor cortex.

5.3 Commercial / pivotal BCIs (status mid-2026)

Device	Vendor	Approach	Status mid-2026
Utah Array	Blackrock Neurotech	96-channel Si microelectrode	Workhorse research platform; > 20 years
N1 Link	Neuralink	1024 electrodes on 64 flexible threads; robot-inserted	PRIME study, 3+ participants; Convoy follow-on
Stentrode	Synchron	Endovascular stent with 16 contacts in superior sagittal sinus	COMMAND trial; less-invasive niche
Connexus	Paradromics	421-electrode high-channel array	First-in-human 2024; channels expanded 2025
Layer 7 Cortical Interface	Precision Neuroscience	Thin-film µECoG, 1024 sites	FDA 510(k) cleared single-use 2025
CorTec Brain Interchange	CorTec (Germany)	Bidirectional ECoG	CE trials
BCIs in tetraplegic / ALS users	Various	Cursor, speech, robotic-arm	Multiple active IDEs

5.4 Decoding methods

• Kalman filters with neural state → cursor velocity (Wu 2002; Kim 2008).
• Refit Kalman (Gilja et al. 2012) — periodic retraining absorbing user adaptation.
• Recurrent neural networks (Sussillo, Pandarinath et al. 2016) for nonlinear decoding.
• Latent dynamic models (LFADS — Pandarinath 2018, Nat Methods) — sequential VAEs for trial-to-trial neural dynamics.
• Transformer-based decoders (2023–2025) — for speech and high-dim hand kinematics.

5.5 Engineering challenges

• Mechanical micromotion of penetrating electrodes (~10–30 µm/day brain pulsation) drives gliosis and signal decay over 6–12 months — the Neuralink retraction event in 2024 underscored chronic stability as the open problem.
• Wireless power and telemetry: N1 Link transmits BLE-class data through bone via inductive link.
• Surgical robotics: Neuralink’s R1 surgical robot threads 64 polymer probes at < 5 µm precision; competitors use stereotactic placement.

5.6 Worked example — Kalman cursor decoder

Two-state random-walk model on cursor velocity ${\mathbf{v}}_t\in {\mathbb{R}}^2$:

• State equation: ${\mathbf{v}}_t=A{\mathbf{v}}_{t-1}+{\mathbf{w}}_t$, A = 0.94·I, w ~ N(0, Q).
• Observation: ${\mathbf{y}}_t=C{\mathbf{v}}_t+{\mathbf{n}}_t$, where ${\mathbf{y}}_t$ is the 96-channel binned spike-count vector (20 ms bins), C is fit by ridge regression on a calibration block.

Refit-Kalman (Gilja 2012) periodically re-estimates C using assumed-intent labels from the calibration cursor task. Latent dynamic models (LFADS, Pandarinath 2018) replace the linear state equation with an RNN encoder/decoder.

5.7 BCIs for non-motor restoration

• Communication — Stanford / UCSF speech BCIs (Willett 2023; Metzger 2023) decode phonemes or articulator motion from M1 and ventral sensorimotor cortex. 62–90 wpm with 25–28% word-error rate.
• Mood / closed-loop psychiatry — research, NeuroPace and academic groups for treatment-resistant depression.
• Memory — DARPA RAM, Mayberg / Suthana groups; hippocampal closed-loop stimulation.

These are adjacent fields; in the prosthetics context they overlap on shared hardware (Utah array, Neuralink threads) and signal-processing infrastructure.

6. Bionic eyes

Device	Vendor	Approach	Status
Argus II	Second Sight (US)	60-electrode epiretinal	FDA 2013; company discontinued 2019; orphaned ~350 users (Ross 2022, Spectrum)
Alpha-AMS	Retina Implant (Germany)	1500-pixel subretinal photodiode	CE 2013; company closed 2019
PRIMA	Pixium Vision (FR) / Science Corp acquisition	Subretinal photovoltaic, ~378 pixels, NIR-illuminated by goggles	Pivotal PRIMAvera EU; FDA pre-IDE
IRIS-V	Pixium (defunct line)	Epiretinal	Halted
Argus III concept	Vivani / academic spin-offs	—	—
Cortical visual prosthesis	UTAH/Moran Eye/CORTIVIS, Bionic Vision Technologies (AU)	V1 microelectrode arrays	Research; Fernández 2021 Sci Transl Med
Orion	Cortigent (Vivani Medical)	V1 cortical stimulation	FDA Breakthrough; trials

The Second Sight collapse left ~350 Argus II users without OEM support and is the cautionary tale for any implanted system without long-term vendor commitment. Cortical V1 approaches (Pelayo, Normann, Bionic Vision Tech) offer access to patients without functional retina but face spatial-resolution and percept-stability limits.

7. Cochlear implants

The most successful neural prosthesis by user count: > 1 million worldwide as of 2025.

Vendor	Flagship	Electrodes	Coding strategies	Notes
Cochlear (AU)	Nucleus 8	22 intracochlear	ACE (Advanced Combination Encoder); SmartSound iQ2	~60% global share
Advanced Bionics (US, Sonova)	HiRes Ultra 3D	16	HiRes Optima; ClearVoice	MRI-conditional
MED-EL (AT)	SYNCHRONY 2	12 (long electrode 31.5 mm)	FS4, FineHearing	Longest electrode, best low-freq
Oticon Medical	Neuro Zti	20	Modiolar	Bone-anchored option

Coding strategies:

• CIS (Continuous Interleaved Sampling) — Wilson et al. 1991 NIH — interleaved biphasic pulses, no inter-electrode interaction.
• ACE (Advanced Combination Encoder) — Cochlear, picks N-of-M strongest frequency bands per cycle.
• Fine Structure (FSP, FS4, MED-EL) — preserves temporal fine structure on apical electrodes for music perception.
• HiRes Optima — Advanced Bionics, hybrid envelope + fine timing.

Bilateral and bimodal: dual CI (bilateral) or CI + contralateral hearing aid (bimodal) restores sound localisation cues — interaural time and level differences. ABI (Auditory Brainstem Implant) for users with absent auditory nerve, e.g. NF2 patients.

Surgical and audiological workflow: mastoidectomy through facial recess; electrode threaded into scala tympani via cochleostomy or round window. Activation 2–4 weeks post-op; programming (mapping) refines T-levels (threshold) and C-levels (comfort / maximum) per electrode over a year of follow-up. Outcome variance is large — children implanted under 18 months reach age-appropriate speech in mainstream school in > 70% of cases (Geers, NIH CDaCI 2016 cohort); late-deafened adult outcomes vary with cochlear nerve health.

Hybrid electric-acoustic stimulation (EAS): for users with residual low-frequency hearing, a shorter electrode preserves apical hair cells; acoustic amplification covers low frequencies while electrical stimulation covers high. MED-EL’s “Flex” electrode family targets this. Music perception remains weaker than speech because temporal fine structure encoding is imperfect under CIS / ACE.

Cost and access: $40 000 – 100 000 for device + surgery + initial therapy in the US, typically covered by major insurers and Medicare. Speech processor (the external worn unit) is replaced every 5–7 years. Worldwide access remains inequitable — over 90% of children with profound bilateral hearing loss in LMICs are not implanted.

8. Exoskeletons for rehabilitation and assistance

Detailed coverage in exoskeletons; clinical use cases:

Device	Vendor	Application	Notable feature
EksoNR / Indego	Ekso Bionics (US)	SCI, stroke, MS rehab	FDA cleared inpatient; modular Indego
ReWalk	ReWalk Robotics (IL/US)	T7–L5 SCI personal use	First personal-use FDA 2014
Atalante	Wandercraft (FR)	Self-balancing, hands-free SCI	No crutches; uses dynamic walking
HAL	Cyberdyne (JP)	Neuro-rehab via bio-electric signal	Reads volitional EMG/EEG intent
Rex	Rex Bionics (NZ)	Self-balancing, low-mobility users	Standing rehab
Phoenix	suitX / Ottobock	Lightweight SCI mobility	Modular
Keeogo	B-Temia (CA)	Stroke / weakness	Knee-only
MyoSuit	MyoSwiss (CH)	Soft assistive	Cable-driven, day use

Stroke vs SCI indications:

• Stroke rehab leverages exoskeletons for repetitive task-specific training (Krakauer 2006 paradigm: high-intensity, task-specific, early intervention). Lokomat-style overground gait training combined with body-weight support.
• SCI (complete) cannot recover volitional motor below lesion; exoskeletons substitute for gait. ASIA grade governs eligibility.
• Phase III RCT evidence is mixed: exoskeleton training is non-inferior to conventional therapy for chronic stroke (Mehrholz Cochrane 2017) and improves bone density / cardiovascular markers in SCI but mobility transfer to home use is limited.

9. Rehabilitation robotics

9.1 End-effector vs exoskeleton designs

Class	Examples	Pros	Cons
End-effector planar	MIT-Manus / InMotion ARM (Krebs, Hogan 1998); Bi-Manu-Track	Single contact point; simple kinematics	Doesn’t constrain individual joints
End-effector 3D	InMotion WRIST, HapticMaster	More workspace	Larger, heavier
Exoskeleton upper-limb	Armeo Spring (Hocoma, passive); ArmeoPower; Harmony (Harmonic Bionics)	Joint-by-joint training	Alignment matters; safety joints
Exoskeleton lower-limb	Lokomat (Hocoma, treadmill); Andago, FLOAT body-weight support	High dose	Equipment cost; non-trivial setup
Hand-specific	Amadeo (Tyromotion), HandTutor, Gloreha glove	Distal training	Single joint focus

The MIT-Manus (Krebs & Hogan 1998, IEEE TNSRE) is the canonical impedance-controlled rehabilitation robot — back-drivable SEA-style design, 6-DoF planar with assist-as-needed control. Commercialised as InMotion ARM / Bionik Laboratories. Install cost $80–150k per clinic.

9.2 Functional Electrical Stimulation (FES)

Stimulation of paralysed peripheral nerve to elicit muscle contraction; restores grasp (Freehand System — historic; NEC Bioness H200), gait (Bioness L300, WalkAide). FES can be combined with exoskeletons in hybrid exoskeleton-FES systems (e.g. MUNDUS, Cybathlon FES bike).

Implantable FES: Ripple Neuromed Networked Neuroprosthesis (Kilgore, Peckham, Cleveland) — implanted electrodes restore grasp in C5/C6 tetraplegia; commercial path uncertain post-Freehand.

9.3 Therapy paradigms

Krakauer’s principles for motor recovery: task-specific, high-intensity, early, with feedback. Massed practice (1000+ reps/session) is feasible with robotic delivery in a way human therapists cannot sustain. Assist-as-needed (Reinkensmeyer 2007) gradually withdraws robot torque so user does more work; error augmentation may accelerate learning in some patients.

9.4 Outcome measures

Measure	Domain	Range	Notes
Fugl-Meyer (FMA-UE / FMA-LE)	Stroke motor recovery	0–66 UE, 0–34 LE	Gold-standard upper-limb
ARAT (Action Research Arm Test)	UE function	0–57	19 items, time-limited
10-Metre Walk Test (10MWT)	Gait speed	m/s	Predicts community ambulation > 0.8 m/s
6-Minute Walk Test (6MWT)	Endurance	metres	Surgery / SCI tracking
Berg Balance Scale	Balance	0–56	Fall-risk threshold ≤ 45
Modified Ashworth Scale (MAS)	Spasticity	0–4	Subjective, low ICC
FIM / WeeFIM	Independence	18–126	Discharge planning
TUG (Timed Up and Go)	Functional mobility	seconds	> 13.5 s = fall risk
ABILHAND / DASH	Self-report UE	varied	Patient-centred
Box and Block Test	Manual dexterity	blocks/min	Hand training

10. Pediatric prosthetics

Children present unique constraints: rapid growth (socket revision every 12–18 months), psychological identity formation, smaller cosmetic envelopes, lower force tolerance. Lessons learned:

• Open Bionics Hero Arm (UK) — leverages 3D printing for low-cost replacement; partnership with Disney enables Marvel/Star Wars/Frozen covers that change the social conversation around limb-difference.
• Limbitless Solutions (UCF) — open-source 3D-printed arms gifted to children via grant funding.
• e-NABLE — global volunteer network printing low-cost (sub-$50) body-powered hands for paediatric and low-resource users; > 8000 devices distributed.
• Growth-tracking sockets — modular distal extensions; some research into adaptive sockets (Anantharaman, Univ. of Salford 2022) using bladders.

Weight is the primary engineering constraint — a pediatric hand must be < 250 g to be tolerable on a 6-year-old.

11. Materials and fabrication

Component	Common materials	Standards
Socket (definitive)	Carbon-fibre/epoxy layup over check positive; PETG/PP for check	ISO 22523
Socket (developing-world)	Polypropylene thermoformed	ICRC norms
Pylon	Aluminium 6061-T6 (light, K1–K2); Ti-6Al-4V (active, K3+)	ISO 10328
Liner	Silicone gel (Iceross, Alpha SmartTemp); PU; TPE	ISO 22675
Hand frame	Aluminium + glass-filled nylon (Ability); Al + carbon (i-Limb)	—
Finger drive	Brushed DC + worm gear; SEAs in research (RIC Hand 6)	—
Battery	Li-ion 7.4 V 800–2000 mAh; LiFePO4 in high-rate hands	IEC 62133
Osseo fixture	Ti-6Al-4V or Ti-CP grade 4 with HA / SLA surface	ISO 14801, ASTM F1108
3D-print prototypes	Mecuris, Open Bionics: PA12 SLS, PETG FDM, MJF	—

composites-taxonomy covers carbon-fibre layup detail; fatigue-analysis for cyclic-load failure prediction in pylons and osseo abutments.

12. Regulatory and reimbursement

12.1 US FDA

• 510(k) clearance — most prosthetic hands, MPKs, exoskeletons (predicates exist).
• De Novo — novel devices without predicate but moderate-low risk (Indego, Atalante).
• PMA (Premarket Approval) — Class III implants: cochlear implants, retinal prostheses, BCIs.
• IDE (Investigational Device Exemption) — pre-market investigational use (Neuralink, Paradromics, OPRA pre-HDE).
• HDE (Humanitarian Device Exemption) — small populations (OPRA 2020).
• Breakthrough Device Designation — Neuralink, PRIMA, Connexus, Layer 7.

12.2 EU MDR

CE marking under MDR (2017/745, effective 2021): most prosthetics Class IIa; powered exoskeletons IIb; implantable BCIs Class III. Notified Body review for IIb+. Many small-batch / pediatric devices struggle with MDR’s notified-body cost, partially driving the 2024 transitional extensions.

12.3 Key standards

• ISO 10328 — Lower-limb structural strength.
• ISO 22523 — Prosthetics and orthotics: requirements and test methods.
• ISO 22675 — Ankle-foot assembly testing.
• ISO 14801 — Dental implant fatigue (used for osseo abutments).
• ISO 13485 — Quality management for medical devices.
• ISO 14971 — Risk management.
• IEC 60601-1 — Electrical safety for medical devices (powered prostheses, MPKs, exoskeletons).
• IEC 62366 — Usability engineering.
• IEC 80601-2-78 — Particular requirements for medical robots for rehabilitation, assessment, compensation, alleviation.
• ANSI/RESNA WC-19 — Wheelchair tie-down (not directly prosthetic).

12.4 Reimbursement (US-centric)

Prosthetic devices billed under HCPCS L-codes (L5000–L8499). Key codes:

• L5856 — addition for microprocessor knee with swing + stance.
• L5857 — addition for microprocessor knee, stance-only.
• L5859 — powered knee addition (Power Knee).
• L5973 — Empower-class powered ankle addition.
• L5980-L5987 — additions for energy-storing feet (Vari-Flex, Pro-Flex).
• L6925/L6935 — myoelectric hand.
• L7259 — multi-articulating hand (PSYONIC, Hero Arm, Vincent — newly added 2022, watershed).
• L7400 series — additions / repairs.

Medicare K-level determination gates access to MPK and advanced feet. The K-classification (K0 = none, K4 = athletic) is set by the prosthetist with physician concurrence; K3+ unlocks ESAR feet, K3+ MPK. Documentation rigor varies hugely across MACs (Medicare Administrative Contractors).

VA (US Department of Veterans Affairs) coverage is broader — covers Genium X3, Empower, multi-articulating hands without strict K-level gating; drives the high-end clinical market.

Single-payer systems (UK NHS, Australia DVA, Canada provincial schemes) vary — UK only began wider C-Leg coverage in 2016. Australia NDIS opened high-tier funding for working-age users.

Out-of-pocket reality: even with insurance, US users typically face $1000–3000/yrincopays,gloves,chargerreplacement,socketrevisions.Catastrophicwithoutinsurance:abasictransfemoralMPKsetupcanrun$80–150k all-in including socket, foot, fitting, training.

13. Companies and ecosystem

Big four limb prosthetics (full-line, vertically integrated):

• Ottobock (Duderstadt, Germany; founded 1919) — broadest portfolio, dominates lower-limb MPK and shoulder-disarticulation; acquired BiOM (Empower), and Otto Bock HealthCare.
• Össur (Reykjavík, Iceland; founded 1971) — strong in feet (Flex-Foot, Cheetah, Pro-Flex), Rheo Knee, Power Knee, owns Touch Bionics (i-Limb).
• Hanger Inc (US; founded 1861 — Civil War origin) — largest US clinical provider, 800+ patient-care clinics; Insignia CAD/CAM platform.
• Endolite / Blatchford (UK) — Linx integrated knee/ankle, Echelon foot family.

Multi-articulating hand specialists: PSYONIC, Open Bionics, Vincent Systems, COVVI, Esper Bionics, BrainCo, Mobius Bionics (DEKA tech), Atom Limbs (early-stage), Phantom Neuro (peripheral nerve add-on, not hand directly).

Components / sockets: Steeper (UK, Espire elbow), Coyote Composites (carbon hardware), College Park (Helix3D, TruStep feet), Freedom Innovations (Plié, Renegade), Fillauer (componentry, Motion Control Utah Arm), Liberating Technologies.

Control electronics: Coapt (PR LDA), Infinite Biomedical Technologies (Sense), Ottobock Myobock electrodes, Liberating Technologies LTI sensors, Sigenics (IMES).

Exoskeletons: Ekso Bionics, ReWalk Robotics, Cyberdyne, Wandercraft, Rex Bionics, suitX (Ottobock subsidiary), B-Temia, German Bionic, Sarcos Robotics (industrial), Hyundai Robotics (industrial), Roam Robotics (skiing/recreation, broader).

Rehabilitation robotics: Hocoma (Lokomat, Armeo — owned by DIH Medical), Bionik Laboratories (InMotion), Tyromotion (Amadeo, Diego), Harmonic Bionics (Harmony SHR), Fourier Intelligence (CN, gait trainer).

Neuroprosthetics / BCI: Blackrock Neurotech, Neuralink, Synchron, Paradromics, Precision Neuroscience, CorTec, Onward (epidural spinal-cord stimulation for SCI gait), NeuroPace (epilepsy, not motor), Ripple Neuro.

Cochlear and retinal: Cochlear, Advanced Bionics (Sonova), MED-EL, Oticon Medical; Pixium → Science Corp acquisition path (PRIMA), Bionic Vision Technologies (AU), Cortigent (Vivani spin-out).

Open-source / academic platforms: Open Bionics Brunel/Hero designs, HANDi Hand (UNB), Open Source Leg (U Michigan), MIT Open Bionics platform, Imperial College OpenBionics (Liarokapis).

14. Cost ranges (USD, 2025–2026)

Device class	Typical price (device only)	Lifecycle cost factors
Body-powered hook + harness (fitted)	$5 000 – 10 000	Cable + hook replacement annually
Single-DoF myoelectric hand	$15 000 – 25 000	Glove, battery, electrode service
Multi-articulating hand	$30 000 – 80 000	Gloves $1500–3000/pair, finger service
LUKE / DEKA arm (shoulder-disarticulation full)	$100 000 – 200 000	High service interval
Microprocessor knee (C-Leg / Rheo)	$50 000 – 80 000 fitted	24-month service
Genium X3	$100 000 – 130 000	Military / VA / private
Powered knee (Power Knee)	$80 000 – 120 000	Battery, motor service
Powered ankle (Empower)	$50 000 – 70 000	Battery, push-off lifetime
Osseointegration (OPRA full procedure)	$100 000 – 150 000	Two surgeries + rehab
Cochlear implant (device + surgery + therapy)	$40 000 – 100 000	Speech processor upgrade every 5–7 yr
Retinal prosthesis (Argus II historic)	$150 000	Out of market
BCI (research, not commercial)	n/a	Per-trial costing
MIT-Manus install (clinical)	$80 000 – 150 000	Annual service contract
Lokomat install	$300 000 – 500 000	High utilization required to amortise
Exoskeleton (Ekso / ReWalk personal)	$80 000 – 150 000	Battery, software updates

15. Edge cases and failure modes

• Sweat / impedance drift — Skin impedance > 50 kΩ collapses CMRR. Dry electrode design (steel pickle-fork), adaptive thresholds, daily re-cal.
• Limb-position effect — Fougner 2011, fold IMU into PR classifier.
• Electrode shift — Volume conduction changes with socket donning/doffing; recalibrate.
• Battery anxiety — End-of-day cliff in hands; multi-day reserve essential for community use.
• Abandonment — 30–50% of upper-limb users abandon within 2 years (Biddiss & Chau 2007); top reasons: weight, training burden, lack of sensory feedback. The simpler the device, the higher the retention — hooks rule by ergonomic economy.
• Phantom limb pain — 50–80% chronic prevalence; mirror-box therapy (Ramachandran), TMR, AMI offer partial relief.
• Osseointegration deep infection — 2–6% lifetime; superficial irritation 30–50%; safety abutment engineered to fracture before bone.
• MPK fall in stance — battery dead → most MPKs default to high-damping safe mode; some (Genium) maintain stance support several minutes after power loss.
• Exoskeleton fall — soft landing is unsolved; Ekso emergency lower function; Wandercraft Atalante adds dynamic balancing reducing crutch dependency.
• BCI signal decay — gliosis around penetrating electrodes; Neuralink 2024 thread retraction; ECoG more stable but lower channel count.
• Implant explant — most cochlear and BCI users will face revision surgery in lifetime.
• Cybersecurity / wireless — BLE-pairing hands and cloud BCIs expand attack surfaces; FDA 2023 guidance on premarket cybersecurity now required.
• Pediatric weight ceiling — < 250 g for school-age hand; cosmetically expressive devices change abandonment dynamics.
• Insurance approval delays — multi-articulating hand denials common; appeal letters and clinical justification routine.

16. Future directions

• Bidirectional BCI for full bionic limb — closed-loop S1 stimulation + M1 decoding for motor + sensory restoration; demonstrated at research scale (Flesher 2021 Science; Ganzer 2020), commercial path open.
• Soft robotic hands — Pneumatic-net actuators (Polygerinos, Walsh) for assist-as-needed glove orthotics; lightweight, compliant, low force.
• Modular plug-and-play prosthetics — Phantom Neuro and others working on percutaneous Bluetooth modules — easily upgradeable electronics on permanent surgical foundation.
• AI gait adaptation — RL-trained intent recognition replacing finite-state machines (Lenzi 2022); slope and terrain adaptation without explicit modes.
• Hyperlocal manufacturing — 3D printing distributed via Open Bionics, Limbitless, Mecuris, e-NABLE; service network rather than central factory.
• Neural implants beyond Utah array — Neuralink polymer threads, Paradromics microwire, Precision Layer 7 thin-film, Sygnos / Ceribell ECoG; channel count climbing to thousands.
• Combined exoskeleton-FES hybrids — exo for structural support, FES for muscle activation, reducing battery and weight.
• OnwardSpinal-cord neuromodulation — Onward Medical’s ARC-EX (epidural stimulator) restored arm function in C5-C7 SCI (Lorach 2023 Nature with Courtine).
• Targeted Sensory Reinnervation + AMI — Ewing amputation (Carty 2018) + multi-DoF sensory return; tested clinically by Herr / Carty / Srinivasan groups.
• Speech BCIs maturing — 62 wpm (Willett 2023), > 90 wpm with ECoG (Metzger 2023 Nature), now translating into clinical pilots (UCSF NeuralKey).
• Pediatric growth-adaptive sockets — bladder-actuated sockets that re-shape with growth, reducing revision frequency.
• OPC-class durability for BCIs — chronic stability is the open scientific problem; if solved unlocks commercial volume for closed-loop motor-restoration BCIs.

17. Software and tooling

Tool	Origin	Use
Coapt CompleteControl Gen2	Coapt LLC	Clinical PR fitting (LDA + adaptation)
Infinite Biomedical Sense	IBT	Clinical PR fitting alternative
Ottobock AxonSoft	Ottobock	bebionic / Michelangelo configuration
OpenSim	Stanford (Delp)	Musculoskeletal simulation, gait
AnyBody	AnyBody Tech	Inverse dynamics, joint loading
Ninapro database	UniGE / Atzori	130k+ EMG trials, 50 subjects
HANDi Hand	UNB / Bahaei	Open-source research hand
Open Source Leg	U Michigan (Rouse)	Research transfemoral platform
OpenBionics	Imperial / Liarokapis	Open-source hand designs
Hangar Insignia	Hangar Inc	In-clinic CAD/CAM socket platform
Standard Cyborg	Standard Cyborg	3D scanning + cloud CAD for socket
Vorum CANFIT	Vorum (CA)	CAD/CAM prosthetics
Mecuris Solution Platform	Mecuris (DE)	3D-printed prosthetic configurator
MoBL-ARMS	Stanford	Upper-limb musculoskeletal model
BCI2000	Schalk et al.	General-purpose BCI research platform
OpenViBE / OpenBCI	various	EEG research and prototyping
MNE-Python	Larson et al.	M/EEG analysis
EMGLAB	open	sEMG decomposition
Neural Signal Processing Toolbox	Blackrock	Utah-array data analysis

See scientific-computing for general SciPy/NumPy stack and signal-processing for the underlying DSP theory.

18. Case studies

18.1 PSYONIC Ability Hand (Akhtar et al., J-BHI 2023)

University of Illinois spin-out led by Aadeel Akhtar. Six independently actuated digits with compliant rubber fingertips housing barometric pressure sensors; first multi-articulating hand to ship with closed-loop vibrotactile force feedback to the residual limb. Mass 490 g, 50 N pinch, 145 N power grip, 30 ms feedback latency. Qualified for Medicare reimbursement (HCPCS L7259) in 2022 — a watershed for US clinical access; before this code, multi-articulating hands were reimbursed only under bundled codes that effectively suppressed coverage. Documented deployment exceeded 350 users by 2024; clinical results show measurable improvement in grasp-quality metrics versus body-powered controls. Notable engineering choice: chose elastomer-compliant fingers over high-precision rigid mechanisms, accepting lower positional accuracy in exchange for impact robustness and intrinsic grasp adaptability.

18.2 Ottobock Genium X3 (DoD / Walter Reed program)

Top-end military / waterproof variant of the Genium MPK developed under contract with the US Department of Defense for use at Walter Reed Army Medical Center. IP68 (submersion to 3 m saltwater for 1 h), sand- and corrosion-resistant. 5-day battery, alternating-step stair ascent, ramp descent with controlled inertial damping, stumble recovery via gyroscope cross-check on knee angular velocity. Cost ~$120 000; service interval 24 months. The X3 set the durability bar that downstream Genium revisions and competitors target. Drives VA / DoD procurement and indirectly underwrites lower-tier civilian Genium variants through shared engineering investment.

18.3 Neuralink PRIME study first-in-human (2024)

Noland Arbaugh, C4–C5 tetraplegic, received the N1 Link implant in January 2024. 1024 electrodes across 64 polymer threads, robot-inserted into hand-knob region of motor cortex; transmits wirelessly to external puck. Within weeks Arbaugh demonstrated cursor control (chess, Civilization) at usable bandwidth — broke World Record for cursor-bits-per-second among BCI users. Subsequent thread retraction reduced active channels by ~85% before software compensation (improved spike-detection on remaining channels, dynamic reference re-fitting) restored performance — a critical real-world lesson in long-term intracortical mechanical stability. Second participant enrolled August 2024, third in late 2024. The PRIME study (FDA IDE) demonstrates that intracortical BCIs are now in clinical use; motor-output to a robotic limb (versus cursor) remains research-stage but converging.

18.4 Lorach et al. 2023 — gait restoration in chronic SCI

Onward Medical ARC-EX epidural spinal stimulator paired with a brain-computer interface from Courtine / Bloch (EPFL/Lausanne). Patient with chronic C5/C6 SCI regained volitional gait by decoding leg intent from ECoG over motor cortex and translating it into patterned epidural stimulation below the lesion. Demonstrated in Nature 2023; not a prosthesis in the limb-replacement sense, but a defining example of bypass-style neural prosthetics that may eventually combine with exoskeletons or muscle-driven gait without external mechanism.

18.5 PRIMA subretinal photovoltaic (PRIMAvera EU pivotal)

Pixium Vision / Stanford (Palanker) subretinal photovoltaic array — 378 pixels (each 100 µm) — illuminated by near-infrared via patient-worn goggles. The chip is passive: NIR-driven photodiodes injection-pulse currents into bipolar cells. Read-out is through residual retinal circuitry. The pivotal trial enrolled atrophic AMD patients and reported letter-acuity gains in 2024. Pixium financial restructuring led to acquisition by Science Corporation (Max Hodak’s company) in 2024. Demonstrates that the post-Argus II era is consolidating around photovoltaic and cortical strategies rather than epiretinal microelectrode arrays.

19. Citations (selected)

• Brånemark, P.-I. (1969). Osseointegrated implants in the treatment of the edentulous jaw. — foundational osseointegration.
• Brånemark, R., Berlin, Ö., Hagberg, K., et al. (2014). A novel osseointegrated percutaneous prosthetic system for the treatment of patients with transfemoral amputation. Bone Joint J. — OPRA outcomes.
• Englehart, K., & Hudgins, B. (2003). A robust, real-time control scheme for multifunction myoelectric control. IEEE TBME 50(7).
• Kuiken, T. A., Dumanian, G. A., Lipschutz, R. D., et al. (2004). The use of targeted muscle reinnervation for improved myoelectric prosthesis control. Prosthet. Orthot. Int.
• Cipriani, C., Controzzi, M., & Carrozza, M. C. (2008). The SmartHand transradial prosthesis. J. NeuroEng. Rehabil.
• Hochberg, L. R., Serruya, M. D., Friehs, G. M., et al. (2006). Neuronal ensemble control of prosthetic devices by a human with tetraplegia. Nature 442 — BrainGate.
• Hochberg, L. R., Bacher, D., Jarosiewicz, B., et al. (2012). Reach and grasp by people with tetraplegia using a neurally controlled robotic arm. Nature 485.
• Collinger, J. L., Wodlinger, B., Downey, J. E., et al. (2013). High-performance neuroprosthetic control by an individual with tetraplegia. The Lancet — MPL cortical control.
• Tan, D. W., Schiefer, M. A., Keith, M. W., et al. (2014). A neural interface provides long-term stable natural touch perception. Sci. Transl. Med.
• Raspopovic, S., Capogrosso, M., Petrini, F. M., et al. (2014). Restoring natural sensory feedback in real-time bidirectional hand prostheses. Sci. Transl. Med.
• Hargrove, L. J., Young, A. J., Simon, A. M., et al. (2018). Intuitive control of a powered prosthetic leg during ambulation: a randomized clinical trial. JAMA.
• Akhtar, A., Aghasadeghi, N., Hargrove, L., & Bretl, T. (2023). Compliant, high-bandwidth multi-articulating prosthetic hand with force feedback. IEEE J. Biomed. Health Inform.
• Biddiss, E., & Chau, T. (2007). Upper limb prosthesis use and abandonment. Prosthet. Orthot. Int.
• Resnik, L., Klinger, S. L., & Etter, K. (2014). The DEKA Arm: features, functionality, and evolution. Prosthet. Orthot. Int.
• Atzori, M., et al. (2016). Ninapro database for sEMG. Sci. Data 3.
• Fougner, A., Scheme, E., Chan, A. D. C., et al. (2011). Resolving the limb-position effect in myoelectric pattern recognition. IEEE TNSRE.
• Willett, F. R., Kunz, E. M., Fan, C., et al. (2023). A high-performance speech neuroprosthesis. Nature 620.
• Metzger, S. L., Littlejohn, K. T., Silva, A. B., et al. (2023). A high-performance neuroprosthesis for speech decoding and avatar control. Nature 620.
• Lorach, H., Galvez, A., Spagnolo, V., et al. (2023). Walking naturally after spinal cord injury using a brain-spine interface. Nature 618.
• Carty, M. J., Herr, H. M., Srinivasan, S. S., et al. (2018). The Ewing amputation: the first human implementation of the agonist-antagonist myoneural interface. Sci. Transl. Med.
• Krebs, H. I., Hogan, N., et al. (1998). Robot-aided neurorehabilitation. IEEE TNSRE — MIT-Manus.
• Krakauer, J. W. (2006). Motor learning: its relevance to stroke recovery and neurorehabilitation. Curr. Opin. Neurol.
• Mehrholz, J., et al. (2017). Electromechanical-assisted training for walking after stroke. Cochrane Database Syst. Rev.
• Wilson, B. S., Finley, C. C., Lawson, D. T., et al. (1991). Better speech recognition with cochlear implants. Nature 352 — CIS strategy.
• Perry, J. (1992). Gait Analysis: Normal and Pathological Function. SLACK Inc.
• Farina, D., Vujaklija, I., Sartori, M., et al. (2017). Man/machine interface based on the discharge timings of spinal motor neurons after targeted muscle reinnervation. Nat Biomed Eng.

20. Glossary

• AMI — Agonist-Antagonist Myoneural Interface (Herr, MIT).
• ARAT — Action Research Arm Test.
• BMI / BCI — Brain-Machine / Brain-Computer Interface.
• BP — Body-Powered prosthesis.
• CI — Cochlear Implant.
• CIS — Continuous Interleaved Sampling.
• DC — Direct Control (one antagonist pair per DoF).
• ECoG — Electrocorticography (subdural cortical recording).
• ESAR — Energy-Storing And Returning foot.
• FES — Functional Electrical Stimulation.
• FIM — Functional Independence Measure.
• FINE — Flat Interface Nerve Electrode (Tyler, CWRU).
• FMA — Fugl-Meyer Assessment.
• HCPCS — Healthcare Common Procedure Coding System (US).
• HDE — Humanitarian Device Exemption.
• HD-sEMG — High-Density surface EMG.
• ICRC — International Committee of the Red Cross.
• IDE — Investigational Device Exemption.
• IMES — Implanted Myoelectric Sensors.
• K-level — Medicare functional ambulation classification K0-K4.
• LFADS — Latent Factor Analysis via Dynamical Systems.
• LIFE / TIME — Longitudinal / Transverse Intrafascicular electrodes.
• MAS — Modified Ashworth Scale (spasticity).
• MDR — EU Medical Device Regulation (2017/745).
• MPK — Microprocessor Knee.
• MPL — Modular Prosthetic Limb (JHU/APL).
• OPRA — Osseointegrated Prostheses for Rehabilitation of Amputees (Integrum).
• POP — Percutaneous Osseointegrated Prosthesis (Utah).
• PR — Pattern Recognition control.
• PMA — Premarket Approval (FDA, Class III).
• PRIME — Neuralink’s first-in-human IDE study.
• PNI — Peripheral Nerve Interface.
• RPNI — Regenerative Peripheral Nerve Interface (Cederna, Michigan).
• SACH — Solid Ankle Cushion Heel (passive foot).
• SCI — Spinal Cord Injury.
• SEA — Series-Elastic Actuator.
• sEMG — surface ElectroMyoGraphy.
• TMR — Targeted Muscle Reinnervation (Kuiken).
• TSR — Targeted Sensory Reinnervation.
• TUG — Timed Up and Go.
• WHO/ISPO — World Health Organization / International Society for Prosthetics and Orthotics.

21. Open problems

• Chronic intracortical stability. Penetrating microelectrodes lose 30–70% of single-unit channels over 6–12 months due to gliotic encapsulation and micromotion. Solving this unlocks closed-loop motor BCIs for prosthetic limb control.
• Sensory-motor latency closure. Bidirectional latency budget for naturalistic touch + grip control is ~50 ms, including transduction, decoding, actuation, stimulation. Current systems hit 200–300 ms.
• Universal socket interface. No “USB” exists between socket and terminal device; mechanical and electrical interfaces are proprietary, locking users into vendor ecosystems.
• Powered-leg battery cliff. Empower-class ankles last ~3000 steps (a few hours of active use); battery weight tax is fundamental.
• Decoder generalisation across days. sEMG decoders re-train every 1–2 days under current clinical use. Continual-learning algorithms (Côté-Allard transfer learning, Hahne adaptation) promising but not deployed at scale.
• Pediatric growth without revision. Modular sockets and growth-tracking liners reduce — but do not eliminate — yearly fitter visits.
• Reimbursement equity. Multi-articulating hand approval under L7259 still varies by MAC and private insurer; appeals burden falls on clinicians and families.
• Retinal cortical resolution. Photoreceptor density (~5 million cones in fovea) far exceeds any prosthesis (<2000 pixels); resolution ceiling probably set by RGC pooling and brain plasticity, not raw electrode count.
• Closed-loop sensory return without surgery. Vibrotactile is the only deployed non-invasive feedback; bandwidth limited; haptic skin patches (Bao group, Stanford) early-stage.
• Soft robot durability. Pneumatic and elastomer hands offer compliance and safety but lifetime is short relative to rigid systems.

22. Adjacent

• prosthetics — sibling: focused myoelectric hand and osseointegration deep-dive
• exoskeletons — assistive and industrial exoskeletons, related actuator and control patterns
• surgical-robotics — implantation systems, robotic neurosurgery for BCI insertion
• impedance-control — back-drivable, assist-as-needed controllers in rehabilitation
• biomechanics — gait phases, joint loading, Perry’s analysis
• bioinstrumentation — EMG/ECoG/EEG front-end design, INA / ADS1299 amplifier chains
• teleoperation-haptics — sensory feedback techniques shared with prosthetic afferent return