emg_from_lfp 0.1.0

Derive a synthetic EMG from LFP via correlation of high-frequency activity (Schomburg et al. 2014).

emg_from_lfp

Derive a synthetic EMG from LFP (or any multichannel field-potential recording) by measuring the correlation of high-frequency activity across spatially separated channels in sliding windows. When channels share a common signal — e.g. EMG contamination during movement — their high-frequency activity co-varies, so this correlation tracks muscle tone without a dedicated EMG electrode. Useful as a movement/arousal proxy for sleep scoring.

This package originated as a self-contained port of the Buzsaki lab's bz_EMGFromLFP.m, based on Erik Schomburg's method. Further optimizations and extensions of the method have been added.

Citation

If you use this method, please cite:

Schomburg, E. W., et al. (2014). Theta phase segregation of input-specific gamma patterns in entorhinal-hippocampal networks. Neuron, 84(2), 470-485.

Install

pip install git+https://github.com/CSC-UW/emg_from_lfp.git

Usage

import numpy as np
from emg_from_lfp import compute, DEFAULTS

# lfp: (n_channels, n_samples), sf: sampling rate in Hz (must be > 2 * max(ws)).
both = compute(lfp, sf)                      # default method="both" -> dict
pw = compute(lfp, sf, method="per_window")   # exact per-window correlation
gl = compute(lfp, sf, method="global")       # fast amplitude-weighted approx.

print(DEFAULTS)  # target_sf=20 Hz, window_size=25 s, band 300-600 Hz, ...

Methods

• per_window — exact mean pairwise Pearson correlation, re-normalized within each window. Amplitude-independent, bounded to [-1, 1]. Numerically identical to the published method and reference implementation, but computed with an incremental sliding-window numba kernel (visits each sample O(1) times) for huge memory & compute gains.
• global — faster global-normalization approximation (a single boxcar moving average). Tracks the shape/ranking of per_window closely but is amplitude-weighted and not bounded to [-1, 1]. Sometimes this is desirable, since the amplitude-weighting may carry information about the intensity of movement.
• both (default) — returns both as a dict; the band-pass filter (the dominant cost) is computed once and shared.

A (near-)constant channel within a window yields NaN for that window, matching scipy.stats.pearsonr.