How Does an EMG Machine Work? Signal Acquisition, Filtering, and Biomedical Engineering Guide

An electromyography (EMG) machine records the tiny electrical signals — motor unit action potentials — that muscle fibers generate when a motor neuron tells them to contract. It picks these signals up with either surface electrodes stuck to the skin (surface EMG, non-invasive) or a fine needle inserted directly into the muscle (needle/intramuscular EMG, used for detailed clinical diagnosis), amplifies them tens of thousands of times with a differential instrumentation amplifier that specifically rejects noise common to both electrodes, filters out motion artifact and 50/60 Hz powerline interference, then digitizes and displays the result. Clinically, an EMG machine is almost always paired with nerve conduction studies (NCS) — electrically stimulating a nerve and timing the response — to distinguish nerve damage from muscle disease. The same underlying signal, processed differently, also drives prosthetic hand control, physical-therapy biofeedback, and sports-science muscle analysis.

Everything below unpacks that definition — from the electrophysiology and amplifier/filter design, through nerve conduction pairing and signal processing math, to manufacturer specifications, failure modes, and where wearable/AI-based EMG is heading next.

 

---

 

1. Why EMG matters as a diagnostic and control signal

A muscle that looks weak on physical exam can be weak for very different reasons — a damaged nerve not delivering the signal, a neuromuscular junction not transmitting it, or the muscle fiber itself being diseased — and telling these apart by exam alone is unreliable. EMG solves this by reading the electrical signal at its source: a motor neuron firing produces a measurable action potential in every muscle fiber it innervates, and the pattern, size, and shape of that signal differs characteristically between nerve disease (neuropathy), muscle disease (myopathy), and junction disorders. The same raw signal, once cleaned up and classified instead of clinically read, is also precise enough to drive a prosthetic limb or tell a physical therapist exactly which muscle a patient is failing to activate during rehab.

2. Working principle: motor unit action potentials

 

When a motor neuron fires, it triggers an action potential in every muscle fiber belonging to that motor unit; these individual fiber potentials overlap and sum together into the recordable EMG signal.¹ Key characteristics of this signal:

• Amplitude: roughly 0–10 mV for surface EMG (varies with electrode placement and how deep the active muscle fibers are); 100 µV–5 mV for needle EMG in distal muscles.
• Duration: typical needle-EMG motor unit action potentials (MUAPs) last 5–15 ms.
• Phases: a normal MUAP has 2–4 phases (baseline crossings + 1); more than 4 phases (“polyphasic”) suggests denervation and reinnervation — one of the clearest EMG signs of nerve damage.²
• Firing pattern: motor units recruit at low frequency at the start of a contraction and fire faster, with more units recruited, as effort increases; a reduced number of recruited units at maximal effort is itself a diagnostic sign of nerve loss.

2.1 Surface EMG vs. needle/intramuscular EMG

	Surface EMG (sEMG)	Needle/intramuscular EMG
Electrode	Ag/AgCl disk electrodes on prepared skin	Concentric or monopolar needle inserted into muscle
Invasiveness	Non-invasive	Invasive
Depth sampled	~1–2 cm below skin	Directly at the needle tip, individual motor units
Typical impedance	50–500 kΩ per electrode	100–500 kΩ, depends on tip diameter
Best for	Prosthetics, rehabilitation, biofeedback, sports science	Clinical neuromuscular disease diagnosis, paired with NCS
Main weakness	Motion artifact, cross-talk from nearby muscles	More invasive, patient discomfort

Surface electrode placement follows standardized protocols — skin abrasion to remove dead skin, an alcohol wipe to remove oil, then conductive gel — and the SENIAM/ISEK standard specifies electrode spacing of at least 20 mm to limit cross-talk from adjacent muscles.⁴

3. Signal acquisition architecture

3.1 The differential (instrumentation) amplifier

The front-end amplifier is the component that determines whether the recorded signal is usable at all:

• Design: typically a three-op-amp instrumentation amplifier topology.
• Gain: adjustable, commonly 100–10,000 V/V (40–80 dB); 1,000 V/V is a typical clinical default.
• Input impedance: >10 MΩ, ideally much higher, so the amplifier itself doesn’t distort the weak signal coming off the electrodes.
• Bandwidth: roughly 0.5–10 kHz for full-range EMG recording.

3.2 Common-mode rejection ratio (CMRR)

CMRR measures how well the amplifier rejects noise that appears equally on both electrodes (like powerline hum) while still amplifying the genuine difference between them:

CMRR (dB) = 20 log₁₀(differential gain / common-mode gain)

• Clinical standard: ≥100 dB at 50/60 Hz, per IEC 60601-2-40.
• Practical impact: at 100 dB CMRR with 100× differential gain, 60 Hz noise common to both electrodes is suppressed by a factor of roughly 100,000 relative to the genuine EMG signal — this is what makes clean recording possible next to ordinary building electrical wiring.

3.3 Filtering: the three-stage pipeline

Filter stage	Typical cutoff	Purpose
High-pass	5–10 Hz (surface EMG), 100 Hz (needle EMG)	Removes motion artifact, DC offset, baseline drift
Notch	50 Hz or 60 Hz, 2–5 Hz bandwidth, >40 dB attenuation	Removes powerline interference and its harmonics
Low-pass (anti-aliasing)	400–500 Hz (surface EMG, per SENIAM/ISEK)	Removes high-frequency noise before digitization

Full EMG bandwidth is typically 20 Hz–2,000 Hz for surface EMG and 50 Hz–10 kHz for needle EMG, which is why the two use different filter cutoffs.⁴

3.4 Sampling and digitization

• Surface EMG: minimum 2 kHz sampling by the Nyquist theorem for a ~1 kHz signal; 4–5 kHz is the typical clinical standard, with some systems (e.g., Natus Nicolet EDX) offering up to 48 kHz.
• Needle EMG: 10–20 kHz minimum, given the higher-frequency content of individual motor unit signals.
• ADC resolution: 12–16 bit is standard in clinical systems; 12-bit gives roughly 70 dB signal-to-noise ratio, 16-bit roughly 90 dB.
• Channels: 4–32 channels in typical clinical surface EMG systems, up to 64–128 in high-density research systems.

 

4. Nerve conduction studies (NCS): EMG’s clinical partner

Needle EMG is almost always paired with nerve conduction studies to build a complete neuromuscular diagnosis:

• Stimulation: supramaximal electrical pulses (20–50 mA) applied to a peripheral nerve.
• Recording: response measured at proximal and distal points along the nerve’s path.
• Measured parameters:
• Latency (ms) — time from stimulus to the first response deflection.
• Amplitude (µV for sensory responses, mV for motor responses) — reflects how many axons are still functioning.
• Conduction velocity (m/s) — distance between the two recording points divided by the difference between proximal and distal latency.
• F-wave latency — an antidromic motor response used to assess the proximal (closer to the spine) segment of a nerve.
• Clinical value: the combination distinguishes demyelination (damage to the nerve’s insulating sheath, which slows conduction velocity but often preserves amplitude) from axonal loss (damage to the nerve fiber itself, which drops amplitude more than velocity) — a distinction that changes both prognosis and treatment.²

5. Signal processing: turning raw voltage into a usable number

• Rectification: full-wave rectification (|x(t)|) converts the naturally biphasic EMG waveform into a one-directional signal representing muscle activation magnitude — the ISEK-recommended standard step before further analysis.
• RMS (root mean square): RMS = √(mean of squared samples over a window); this is the most common time-domain feature in both clinical and prosthetic-control applications because it correlates strongly and roughly linearly with the force a muscle is producing.
• Frequency analysis: a Fast Fourier Transform over 256–1,024 sample windows (typically with 50% overlap) yields the power spectral density; mean frequency (MNF) and median frequency (MDF) are the two standard summary metrics, and both measurably decrease — by roughly 10–30% over 5–10 minutes of sustained contraction — as a direct, quantifiable signature of muscle fatigue.
• MUAP parameter analysis (needle EMG specifically): duration, amplitude, and phase count are measured against the ranges above; AAEM/AANEM quality standards require electrode impedance under 500 kΩ and baseline noise under 50 µV peak-to-peak before a recording is considered diagnostically valid.²

6. Clinical applications

• Neuromuscular disease diagnosis: myopathy typically shows short-duration, low-amplitude, abundant polyphasic MUAPs; neuropathy/denervation typically shows large-amplitude, long-duration MUAPs with reduced recruitment at maximal effort.
• Neuromuscular junction disorders (e.g., myasthenia gravis): show characteristic “decrement” or “blocking” patterns of transmission failure under repetitive stimulation.
• Rehabilitation and biofeedback: real-time EMG feedback helps retrain muscle activation after stroke or nerve injury, and combined with an electrogoniometer, tracks muscle-joint coordination during gait rehab.
• Prosthetics control: multi-muscle surface EMG, decoded through pattern recognition, drives prosthetic hand and arm commands; conventional machine learning (SVM/LDA) reaches roughly 85–95% classification accuracy, while modern deep-learning models (CNN/LSTM) reach roughly 92–97%, with real-time latency under 50 ms considered necessary for the control to feel intuitive.⁵
• Sports and biomechanics analytics: multi-muscle EMG clustering detects asymmetries and fatigue patterns during athletic movement.

7. Manufacturer landscape (representative, non-exhaustive)

Manufacturer	Representative system	Sampling rate	Channels	CMRR	Notes
Natus Medical	Nicolet EDX	Up to 48 kHz	Up to 32 (surface)	>100 dB	IEC 60601-2-40:2025 compliant; hospital neurology standard
Cadwell Industries	Cascade	10–20 kHz	Up to 16	Built-in artifact rejection	Primary care neurology, intraoperative monitoring
Medtronic	Keypoint	4 kHz standard, up to 30 kHz	2–8	—	Portable and desktop configurations
Noraxon USA	DTS EMG	1,500–2,000 Hz	Wireless modules, 2.4 GHz	—	Wireless wearable, biomechanics/PT focus
Delsys Inc.	Trigno IM	20 kHz	Up to 16 wireless	—	<5 ms wireless latency, 6+ hour battery, research/rehab robotics

8. Common failure modes and artifacts

Problem	Mechanism	Typical mitigation
Motion artifact	Electrode-skin impedance changes with movement, producing 0.5–10 Hz noise	Higher high-pass cutoff (5–10 Hz), stable reference electrode placement, good adhesive gel
Electrode impedance rise	Perspiration, gel drying, adhesive degradation; signal loss above ~1 MΩ	Impedance-check function before recording; re-gel or replace electrode
Powerline interference (50/60 Hz)	Capacitive/inductive coupling from nearby equipment or wiring	Notch filter, star-point grounding, >100 dB CMRR amplifier, shielded twisted-pair leads
Cross-talk	Adjacent muscle’s signal picked up by a nearby electrode via volume conduction	Electrode placement over the motor point, ≥20 mm inter-electrode spacing (SENIAM), high-density spatial filtering
Saturation/clipping	Strong contraction exceeds the ADC’s input range	Automatic gain control or manual range selection
Electrode detachment	Poor skin contact	Automated impedance/baseline monitoring; abrupt signal loss triggers a re-placement

9. Regulatory standards

• IEC 60601-2-40:2025 — the core safety and performance standard specific to electromyographs, requiring CMRR ≥100 dB at 50/60 Hz, input impedance ≥10 MΩ, patient auxiliary leakage current under 5 mA, and — in its 2025 update — added cybersecurity and data-integrity logging requirements for wireless EMG systems.³
• AAEM/AANEM guidelines — technique standards covering minimum muscle sampling per limb, standardized MUAP measurement methodology (duration base-to-base, amplitude peak-to-peak), and quality thresholds (electrode impedance <500 kΩ, baseline noise <50 µV peak-to-peak).²
• ISO 13485:2016 — the quality management system standard EMG device manufacturers must operate under, requiring a design history file, ISO 14971 risk management, and ISO 10993 biocompatibility testing for electrode materials.
• SENIAM/ISEK standards — the reference protocol for surface electrode placement across 27+ commonly assessed muscles, plus the standard 20 Hz–450 Hz surface EMG bandwidth recommendation.⁴

10. Recent developments (2023–2026)

• Wireless/wearable systems: Bluetooth Low Energy and ultra-wideband transmission with under 5 ms latency now support real-time prosthetic feedback; onboard flexible/stretchable electrode materials (graphene, PEDOT:PSS) reduce impedance variation from skin movement.⁵
• Multimodal sensor fusion: combining surface EMG with an inertial measurement unit lifts gesture-recognition accuracy from roughly 85–90% (EMG alone) to over 95%, since the IMU disambiguates limb position independent of the muscle signal itself.⁵
• AI-based classification: the field has moved from classical machine learning (SVM/LDA) through CNNs and CNN-LSTM hybrids (94–98% gesture accuracy) toward transformer-based architectures reporting 95–99% accuracy with better generalization across recording sessions, plus early graph neural network research treating electrode geometry explicitly as a graph.⁵
• Domain adaptation for real-world use: transfer learning fine-tuned on 10–20 calibration gestures per user reports 15–25% accuracy gains on a new recording session, addressing the long-standing problem that EMG signals drift day to day with electrode placement and skin condition.⁵
• Clinical AI: a 2026 study evaluating AI classification of needle EMG signals (distinguishing genuine muscle activity from background noise) reported meaningful improvement in diagnostic consistency and reduced operator-to-operator subjectivity.⁶

11. Conclusion

An EMG machine’s entire value comes down to recovering a very small, very noisy biological signal cleanly enough to either diagnose disease or drive a device — and every design choice described above, from the >100 dB CMRR requirement to the specific 5–10 Hz high-pass cutoff, exists to solve one part of that noise problem. The clinical version pairs the signal with nerve conduction timing to separate nerve disease from muscle disease; the engineering version strips away the diagnostic interpretation and asks only whether a machine-learning model can turn the same signal into a reliable, low-latency command — and by 2026 the honest answer for controlled settings is yes, with real accuracy numbers now published rather than assumed.

FAQ

Is needle EMG painful?
It involves inserting a fine needle directly into the muscle, so it causes brief discomfort, but it’s typically well tolerated and is the standard method for detailed neuromuscular diagnosis when surface EMG isn’t specific enough.

Why does an EMG machine need such a high common-mode rejection ratio?
Because the genuine muscle signal is measured in microvolts to low millivolts, while ordinary building electrical wiring radiates 50/60 Hz noise that can be much larger — a CMRR below about 100 dB lets that noise swamp the real signal.³

Can EMG tell the difference between a nerve problem and a muscle problem?
Yes — a large-amplitude, long-duration, reduced-recruitment pattern points to nerve damage (neuropathy), while a short-duration, low-amplitude, abundant polyphasic pattern points to muscle disease (myopathy); this is one of EMG’s core diagnostic jobs.²

How accurate is EMG-based prosthetic control today?
Published results for multi-muscle pattern-recognition control report roughly 85–95% accuracy with classical machine learning and 92–99% with modern deep-learning and transformer models, though accuracy commonly drops on a new day or a new user without recalibration.⁵

References