生理信号处理与研究分析｜个人思维图谱小站

正文

references/emg.md

Checked 2026-10-01 against NeuroKit2 0.2.13 stable runtime/source, the official EMG API, and human surface-EMG guidance.

Acquisition contract

Record muscle and task, electrode type/size/orientation/inter-electrode distance, reference, skin preparation, amplifier gain/range, hardware filters, unit, sampling rate, synchronization, posture/contraction, and artifact annotations. Follow a muscle-specific placement protocol such as SENIAM when applicable.

EMG amplitude depends on geometry, subcutaneous tissue, cross-talk, electrode contact, hardware, and contraction history. It is not a direct universal force scale.

Stable processing pipeline

signals, info = nk.emg_process(emg, sampling_rate=1000)

Pinned 0.2.13 default columns:

EMG_Raw, EMG_Clean, EMG_Amplitude,
EMG_Activity, EMG_Onsets, EMG_Offsets

info contained event-level EMG_Activity, EMG_Onsets, EMG_Offsets, and sampling_rate. This is a default observation, not a universal schema.

Cleaning and amplitude

clean = nk.emg_clean(emg, sampling_rate=1000, method="biosppy")
amplitude = nk.emg_amplitude(clean)

Important stable details:

emg_clean() currently offers biosppy and none.
The BioSPPy path uses a fourth-order 100 Hz high-pass Butterworth filter followed by constant detrending.
emg_amplitude() takes only the cleaned vector; it has no sampling_rate argument in 0.2.13.

The last point is a scientific constraint: its private envelope routine defaults to 1000 Hz (10–400 Hz band-pass, rectification, 8 Hz low-pass), and the public wrapper does not forward the acquisition rate. emg_process(..., sampling_rate=500) changes cleaning/detection timing but still computes that envelope as if sampled at 1000 Hz. The example above uses 1000 Hz. At other rates, implement and validate an explicit envelope with the actual rate or use a justified resampling workflow; do not report the hard-coded nominal cutoffs as the physical cutoffs of that recording.

A 100 Hz high-pass may be unsuitable for some surface-EMG endpoints and impossible at low sampling rates. Do not treat the package default as a universal acquisition standard. Set sampling rate above twice the highest retained frequency with transition band margin; surface EMG is commonly acquired around 1000–2000 Hz, but the study, hardware, muscle, and endpoint determine requirements.

Activation detection

Stable signature:

emg_activation(
  emg_amplitude=None, emg_cleaned=None, sampling_rate=1000,
  method="threshold", threshold="default", duration_min="default",
  size=None, threshold_size=None, **kwargs
)

Use the correct input for the method:

threshold or mixture: pass emg_amplitude;
pelt, biosppy, or silva: pass emg_cleaned.

activity_signals, activity_info = nk.emg_activation(
    emg_amplitude=amplitude,
    sampling_rate=1000,
    method="threshold",
    threshold="default",
    duration_min=50,  # samples: 50 ms at 1000 Hz in stable 0.2.13
)

Return order is (activity_signals, info). The DataFrame has same-length EMG_Activity, EMG_Onsets, and EMG_Offsets; the dict contains event indices. An explicit duration_min is passed to sample-count event filtering in stable 0.2.13; the docstring incorrectly says seconds. Only "default" converts 0.05 seconds using the supplied rate. Pass round(minimum_seconds * sampling_rate) for an explicit minimum. size is seconds for biosppy, but samples for silva; Silva's threshold_size is also samples. Verify detected run lengths at the chosen rate.

Automatic thresholds are hypotheses, not ground truth. Tune/validate on independent labeled contractions or a prespecified baseline. Report threshold, window, smoothing, minimum duration, and false/missed activation performance.

Missing data and artifacts

Inspect and annotate:

clipping/saturation and disconnected/flat channels;
motion/cable artifacts and low-frequency transients;
powerline interference;
ECG contamination, especially trunk/proximal sites;
cross-talk from adjacent muscles; and
baseline changes from contact/sweat.

Do not interpolate across a burst or activation onset. Segment long gaps, preserve a validity mask, and reject affected trials/windows under prespecified rules. Filtering cannot establish that a deflection came from the target muscle.

Event and interval features

epochs = nk.epochs_create(
    signals,
    events,
    sampling_rate=1000,
    epochs_start=-0.1,
    epochs_end=1.0,
    baseline_correction=False,
)
event_features = nk.emg_eventrelated(epochs)
interval_features = nk.emg_intervalrelated(signals)

Documented event-related features include EMG_Activation, EMG_Amplitude_Mean, EMG_Amplitude_Max, EMG_Amplitude_SD, EMG_Amplitude_Max_Time, and EMG_Bursts. Interval analysis returns EMG_Activation_N and EMG_Amplitude_Mean in the pinned official example. Columns are conditional; inspect output.

For startle or rapid onset work, hardware latency, synchronization, filter group delay, rectification/smoothing, and onset-definition error can dominate the result.

Normalization

Raw/envelope amplitude is usually not comparable across participants or sessions. Possible denominators include MVC, reference contraction, or within-participant standardization, but each changes the estimand.

If using MVC:

acquire it with a validated, safe protocol;
inspect fatigue, pain, effort, clipping, and target/cross-talk;
define which statistic and window represent MVC;
report repetitions and reliability; and
never divide by a near-zero/invalid reference.

Normalization does not make electrode/site differences disappear.

Interpretation boundary

NeuroKit2 EMG can support research on amplitude and detected activity. It does not provide validated motor-unit decomposition, neuromuscular diagnosis, fatigue monitoring, prosthetic control safety, rehabilitation decisions, or sleep scoring. Those require endpoint-specific acquisition, algorithms, standards, and independent validation.

生理信号处理与研究分析

正文