Surface vs intramuscular EMG

"One sensor hears the crowd; the other interviews a single voice. Do not ask the crowd to name the speaker, and do not ask the interview to describe the room."

A Discerning AI Agent

One signal source, two ways to eavesdrop

The generator is the same in both cases. A motoneuron in the spinal cord commands a motor unit, one neuron plus all the muscle fibers it innervates, and every time it fires, those fibers depolarize together and emit a motor unit action potential (MUAP). Force is graded by two knobs: recruiting more motor units, and firing each one faster. What differs between surface and intramuscular EMG is not the source but the pickup geometry, and geometry is destiny because muscle tissue is a volume conductor that attenuates and low-pass-filters potential as it spreads.

Surface EMG (sEMG) places two electrodes on the skin over a muscle and measures their voltage difference. Because the electrode is centimeters of tissue away from most active fibers, and because tissue smears high spatial frequencies, each surface electrode integrates over a large detection volume containing many motor units. The result is loud in aggregate (up to several millivolts), broadband from roughly 20 to 450 Hz, and inherently a population statistic. It is noninvasive, painless, repeatable, and wearable, which is exactly why it dominates prosthetics, ergonomics, and consumer neuromotor interfaces, the subject of Chapter 27.

Intramuscular EMG (iEMG) inserts a conductor directly into the muscle: a hypodermic needle electrode for diagnostic recording, or a pair of thin insulated fine wires hooked into the tissue for research and deep-muscle access. Now the electrode sits inside the detection volume, dominated by the few fibers nearest its exposed tip. The recorded MUAPs are sharp, high-bandwidth (energy extends past 1 kHz, so sampling climbs to 5000 to 10000 Hz to respect Nyquist as taught in Chapter 3), and selective enough that a clinician can watch a single motor unit recruit and fire. The price is that it is invasive, uncomfortable, positionally unstable, and cannot be worn home.

Crosstalk: the price surface EMG pays for reach

Because a surface electrode's detection volume is large, it does not respect anatomical boundaries. Activity from a neighboring muscle bleeds into the channel meant for the target muscle, a corruption called crosstalk. When you record the forearm flexors for a hand-close command, the extensors a centimeter away contaminate the trace, and no amount of gain fixes it because the interfering signal is real voltage from a real muscle. Crosstalk is the single most consequential difference in information quality between the two modalities: intramuscular electrodes are largely immune because their tiny detection volume rarely reaches across a fascial plane, while surface electrodes are chronically vulnerable. This is why Section 32.6 devotes itself to artifact and crosstalk handling, and why any sEMG model claiming to read one muscle in isolation deserves suspicion.

The physics is a rapid falloff of potential with distance. Treat each active fiber as a source whose surface contribution decays roughly as an inverse power of the source-to-electrode distance \(r\). The measured channel is the sum over all fibers weighted by that decay,

where \(s_i\) is fiber \(i\)'s action potential, \(\lambda\) is a tissue length scale, and the exponent \(p\) (around 2 to 3 for surface pickup) controls how sharply distant sources are suppressed. For an intramuscular tip, the nearest fibers have \(r_i \to 0\) and dominate the sum; for a surface electrode, thousands of moderately distant fibers each contribute a little, and the many weak, distant terms are exactly the crosstalk. The short model below turns this equation into an experiment you can run.

import numpy as np

rng = np.random.default_rng(0)
# 400 fibers scattered in a muscle cross-section; a second muscle sits 12 mm away.
n = 400
x = rng.uniform(-8, 8, n)          # mm, lateral position within/near target
y = rng.uniform(-8, 8, n)          # mm, depth
in_target = np.abs(x) < 5          # fibers belonging to the target muscle

def channel_weights(ex, ey, lam=3.0, p=2.5):
    r = np.sqrt((x - ex)**2 + (y - ey)**2)     # source-to-electrode distance
    return 1.0 / (1.0 + r / lam)**p

def crosstalk_fraction(ex, ey):
    w = channel_weights(ex, ey)
    return w[~in_target].sum() / w.sum()        # energy share from the wrong muscle

surface = crosstalk_fraction(ex=0.0, ey=10.0)   # electrode on skin, above the muscle
intra   = crosstalk_fraction(ex=0.0, ey=0.0)    # tip inside the target belly
print(f"surface crosstalk fraction: {surface:.2f}")
print(f"intramuscular crosstalk fraction: {intra:.2f}")

As Listing 32.1 makes concrete, moving the electrode from the skin into the belly does not merely raise amplitude; it collapses the crosstalk fraction, because the inverse-power weighting now favors a few close fibers over the distant crowd. That single number, the share of energy arriving from the wrong source, is the quantitative heart of the surface-versus-intramuscular choice.

What each modality lets a model do

The pickup geometry decides which learning problems are even well posed. Because sEMG delivers a smooth, high-signal-to-noise envelope of population activity, it is ideal for estimating effort, timing of muscle onset, and coarse movement intent, the tasks that drive prosthetic and gesture control in Section 32.3 and fatigue estimation in Section 32.4. It is a natural fit for the classical envelope and spectral features of Section 32.2 and for the deep time-series models of Chapter 13. What sEMG cannot natively do is name individual motor units, because they are summed away before the electrode ever sees them.

Intramuscular EMG, by contrast, makes motor-unit decomposition tractable: because MUAPs arrive sharp and separable, algorithms can cluster them by shape and recover each unit's firing train, the fine-grained neural code behind a contraction. This is the reference modality for neurophysiology and for validating what surface methods only estimate. The frontier reunites the two: high-density surface EMG (HD-sEMG), a grid of dozens to hundreds of closely spaced skin electrodes, samples the spatial voltage field densely enough that blind-source-separation methods can decompose some motor units noninvasively, recovering a slice of the intramuscular view without a needle.

import neurokit2 as nk
clean = nk.emg_clean(emg, sampling_rate=2000)      # band-pass + mains handling
signals, info = nk.emg_process(emg, sampling_rate=2000)  # amplitude envelope + onsets

Choosing between them

The decision reduces to a short checklist. Reach for surface EMG when you need noninvasive, repeatable, wearable recording; when the question is about whole-muscle effort, timing, or gross intent; and when the target muscle is superficial and reasonably isolated from crosstalk. Reach for intramuscular EMG when you need single-motor-unit resolution, when the muscle is deep or small (where a surface electrode would record mostly its neighbors), or when crosstalk immunity is non-negotiable, accepting that the recording is invasive, position-sensitive, and unrepeatable across days. And reach for HD-sEMG when you want to push toward intramuscular-grade information (a spatial field, partial decomposition) while staying on the skin, accepting more electrodes, more compute, and a decomposition you must validate. The rest of this chapter mostly lives in the surface world, because that is where deployable neuromuscular AI lives, but every surface claim should be read against what an intramuscular recording would have shown.

What's Next

In Section 32.2, we take the surface signal this section has characterized and turn it into numbers a model can use: the amplitude envelope, the spectral features whose median frequency shifts as a muscle tires, and the time-domain descriptors that survive the noisy, non-stationary reality of skin-mounted electrodes.