ECG, PPG, EEG, EMG, respiration, EDA

"The body broadcasts on six channels at once, in six different units, and none of them is labeled. I was hired to listen."

A Newly Onboarded AI Agent

Three physical routes out of the body

Every biosignal in this chapter reaches your analog-to-digital converter by one of three physical routes, and the route dictates almost everything about the sensor. Electrical signals are potential differences that ionic currents create as cells depolarize; you read them with electrodes as a voltage. The heart (ECG), the cortex (EEG), and skeletal muscle (EMG) all broadcast this way, which is why they share hardware (differential amplifiers, reference electrodes) yet occupy wildly different amplitude bands. Optical signals are read by shining light into tissue and measuring what returns; photoplethysmography (PPG) is the canonical example, inferring blood volume from how much a green or infrared beam is absorbed. Mechanical and chemical signals cover the rest: respiration is the mechanical expansion of the ribcage or the cooling of exhaled air, and electrodermal activity (EDA) is a chemical-electrical hybrid, the skin's changing conductance as sweat glands fill.

Why insist on this taxonomy first? Because the transduction route, not the organ, predicts the failure modes you will spend Section 28.2 fighting. Electrical channels are corrupted by the same 50/60 Hz mains hum and by electrode-skin impedance shifts. Optical channels are wrecked by motion and by ambient light leaking into the photodiode. Mechanical channels drift with posture. Knowing the route tells you, in advance, which artifact to expect and which second sensor can explain it away.

The cardiovascular pair: ECG and PPG see the same beat differently

ECG and PPG both track the heartbeat, but they measure different physical events separated in time, and confusing them is a classic beginner error. The electrocardiogram records the electrical command: the coordinated depolarization that makes the heart contract, seen as the familiar P wave, QRS complex, and T wave. It is what: a voltage of roughly 1 mV between electrodes, with diagnostic content from about 0.5 to 40 Hz (the full clinical band reaches 150 Hz), typically sampled at 250 to 500 Hz. It is precise in time; the sharp QRS peak marks each beat to within a few milliseconds, which is why ECG is the reference for heart-rate and rhythm work in Chapter 29.

The photoplethysmogram records the mechanical consequence: the pulse of blood volume arriving at the fingertip or wrist a fraction of a second after the electrical command, detected as a dip in transmitted or reflected light. Its pulsatile (AC) component is often only 1 to 2 percent of the steady (DC) light level, so PPG is intrinsically a small-signal-on-large-background measurement, sampled at 25 to 250 Hz and band-limited to a few hertz. The gap between the ECG R peak and the PPG upstroke is the pulse transit time, a quantity that carries blood-pressure information exploited in Chapter 30. When you fuse the two, remember: the same beat produces a crisp electrical spike and, tens of milliseconds later, a soft optical hill.

Neural and muscular electricity: EEG and EMG

EEG and EMG are both electrical, both read with electrodes, and both easy to mistake for each other on a raw plot, yet they sit at opposite ends of the amplitude and frequency map. The electroencephalogram is the faintest signal in routine use: summed postsynaptic potentials of cortical neurons, attenuated by skull and scalp down to tens of microvolts. Its information is spectral, conventionally parsed into bands, delta (\(<\)4 Hz), theta (4 to 8 Hz), alpha (8 to 13 Hz), beta (13 to 30 Hz), and gamma (\(>\)30 Hz), each loosely tied to states from deep sleep to alert attention. Because the signal is tiny and the interesting structure is in relative band power, EEG demands high-input-impedance amplifiers and careful referencing, themes taken up in Chapter 31.

The electromyogram is the opposite temperament: a loud, broadband burst. Surface EMG captures the summed action potentials of firing motor units, reaching several millivolts with meaningful energy from roughly 20 to 450 Hz, which forces sampling at 1000 Hz or higher to avoid aliasing the upper band. Where EEG rewards patient spectral analysis, EMG is often used for fast onset detection: the muscle is either recruiting or it is not, and the envelope of the rectified signal tells you when. That amplitude-and-bandwidth contrast is why the two, despite sharing a transduction route, need different front-ends and different feature pipelines, developed further in Chapter 32.

The slow signals: respiration and EDA

The last pair moves so slowly that beginners often forget they are signals at all. Respiration oscillates at roughly 0.1 to 0.5 Hz (about 6 to 30 breaths per minute) and can be sensed many ways: a chest strain belt, a thoracic impedance measurement, an inductance plethysmography band, or a thermistor reading the warm-cool cycle of airflow at the nose. Because it is slow, respiration is cheap to sample (tens of hertz suffice) but expensive to isolate, since it modulates other signals: it rides on the ECG baseline and on PPG amplitude, an entanglement that lets you sometimes recover breathing rate from a heart sensor alone.

Electrodermal activity, historically called galvanic skin response, is slower still and chemical in origin. Eccrine sweat glands, driven by the sympathetic nervous system, change the skin's conductance measured in microsiemens. EDA splits into a slowly wandering tonic level (skin conductance level) and brief phasic bumps (skin conductance responses) that rise over one to five seconds after an arousing stimulus. Its band is essentially below 1 Hz, so 4 to 32 Hz sampling is ample. EDA is the workhorse of stress and arousal inference precisely because it is an almost direct readout of sympathetic tone, with no fast dynamics to chase. Table 28.1.1 collects the amplitude, bandwidth, and sampling anchors for all six modalities so you can size a front-end at a glance.

The table makes one design rule concrete. Sampling rate is set by the top of the useful band through Nyquist, so EMG at 450 Hz forces at least 1000 Hz sampling while EDA below 1 Hz is comfortable at 8 Hz. The short Python below turns the table into an automatic check: given a signal's band edge, it reports the minimum safe sampling rate and flags any recording that undersampled.

import numpy as np

# (name, upper edge of useful band in Hz, sampling rate the device actually used)
recordings = [("ECG", 40, 500), ("EMG", 450, 500), ("EDA", 1, 8), ("EEG", 40, 250)]

for name, f_high, fs in recordings:
    nyquist_min = 2 * f_high          # theoretical floor
    safe_min = 2.5 * f_high           # practical margin for a realizable anti-alias filter
    status = "OK" if fs >= safe_min else "UNDERSAMPLED"
    print(f"{name:4s}  band<= {f_high:4d} Hz  fs={fs:5d} Hz  "
          f"need>= {safe_min:6.1f} Hz  -> {status}")

As Listing 28.1 shows, the EMG entry that looked fine in isolation is caught: sampling a 450 Hz band at 500 Hz leaves no room for an anti-aliasing filter, so out-of-band energy folds back onto the signal you care about. This is the sampling discipline of Chapter 3 applied to physiology, and it is the first gate every biosignal pipeline should pass.

import neurokit2 as nk
signals, info = nk.ecg_process(ecg, sampling_rate=500)   # cleans, finds R peaks, rates
hr = signals["ECG_Rate"].mean()                          # beats per minute

What's Next

In Section 28.2, we turn from what these six signals are to how they get corrupted: the physiological noise and motion artifacts that ride on every channel, why an optical sensor and an accelerometer end up fighting over the same frequency band during exercise, and the artifact taxonomy that decides which corruption you can filter and which you must instead detect and reject.