cleanline: Cleanline

In gjmvanboxtel/eegr: Electrophysiological data analysis

Cleanline

Description

Estimate and remove sinusoidal (e.g. line) noise from EEG channels using multi-tapering and a Thompson F-statistic.

Usage

cleanline(
  x,
  fs = 1,
  lfreq = 50,
  fbw = 2,
  detrend = c("short-linear", "short-mean", "long-linear", "long-mean", "none"),
  w = 4,
  overlap = 0,
  tbw = 2,
  k = round(tbw * w - 1),
  p = 0.01,
  tau = 100,
  maxIter = 10
)

Arguments

x	input time series, specified as a numeric or complex or matrix vector. In case of a vector it represents a single signal; in case of a matrix each column is a signal. Alternatively, an object of class ctd
fs	sampling frequency of x in Hz. Default: 1. Overruled if x is a ctd object, in which case the sampling frequency is fs(x)
lfreq	line frequencies to remove, either specified as a single numeric value, or as a character string. If a single numeric value is specified, then this value will be expanded to include multiples of that frequency (harmonics) up to the Nyquist frequency. If a character string is specified, then the string will be converted to numeric values without further expansion. Default: 50
fbw	frequency bandwidth. Bandwidth centered on each lfreq to scan for significant lines. Default: 2 Hz.
detrend	character string specifying detrending option; one of: short-linear remove the mean from the data before splitting into segments (default) short-mean remove the mean value of each segment long-linear remove linear trend from the data before splitting into segments long-mean remove linear trend from each segment none no detrending Specifying one of the long forms will result in detrended cleaned data. By contrast, specifying one the of short forms will perform the calculations on detrended segments but this will not affect the returned data.
w	length of sliding window (segments) in seconds. Default: 4
overlap	proportion of overlap between the segments. Default: 0
tbw	taper bandwidth, specified as a positive numeric value. Default: 2 Hz
k	number of tapers to use, specified as a positive numeric value. Default: tbw * w - 1
p	significance level cutoff for F-test. Default 0.01
tau	Window overlap smoothing factor. Default: 100
maxIter	Maximum times to iterate removal. Default: 10

Details

Sinusoidal noise can be a prominent artifact in recorded electrophysiological data. This can stem from AC power line fluctuations (e.g. 50/60 Hz line noise + harmonics), power suppliers (e.g. in medical equipment), fluorescent lights, etc. Notch filtering is generally undesirable due to creation of band-holes, and significant distortion of frequencies around the notch frequency (as well as phase distortion at other frequencies and Gibbs rippling in the time-domain). Other approaches for sinusoidal ("line") noise reduction include adaptive regressive filtering approaches (e.g. RLS, LMS), but these typically require a reference signal (e.g. externally-recorded noise reference), which is often unavailable. Blind-source separation approaches such as ICA may help mitigate line noise, but often fail to completely remove the noise due to spatiotemporal non-stationarities in the noise.

CleanLine uses an approach for line noise removal advocated by Partha Mitra and Hemant Bokil in "Observed Brain Dynamics" (2007), Chapter 7.3.4.

In brief, the data is traversed by a sliding window. Within each window, the signal is transformed to the frequency domain using a multi-taper FFT. The complex amplitude (amplitude and phase) is thus obtained for each frequency. Under the assumption of a deterministic sinusoid embedded in white noise, we can set up a regression of the multi-taper transform (spectrum) of this sinusoidal signal onto the multitaper spectrum of the original data at a given frequency. The regression coefficient is a complex number representing the complex amplitude (phase and amplitude) of the deterministic sinusoid. From this, a time-domain representation of the sinusoid may be constructed and subtracted from the data to remove the line.

Typically, one does not know the exact line frequency. For instance, in the U.S.A., power line noise is not guaranteed to be at exactly 60 Hz (or even to have constant phase over a given period of time). To ameliorate this problem a Thompson F-Test may be applied to determine the statistical significance of a non-zero coefficient in the above regression (indicating a sinusoid with significantly non-zero amplitude). We can then search within a narrow band around the expected location of the line for the frequency which maximizes this F-statistic above a significance threshold (e.g. p=0.05).

Overlapping short (e.g. 2-4 second) windows can be used to adaptively estimate the frequency, phase, and amplitude of the sinusoidal noise components (which typically change over the course of a recording session). The discontinuity at the point of window overlap can be smoothed using a sigmoidal function.

Value

A list consiting of 3 elements:

y

the cleaned data

lfreq

the cleaned line frequencies

niter

the number of iterations executed

Author(s)

Original Matlab code by Tim Mullen (2011); ported to R and adapted by Geert van Boxtel, [e-mail],

See Also

mtspec

Examples

data(EEGdata)
ospec <- mtspec(EEGdata[, 1:28], fs = fs(EEGdata), detrend = "short-linear")
cleaned <- cleanline(EEGdata[, 1:28], fs = fs(EEGdata))
cspec <- mtspec(cleaned$y, fs = fs(EEGdata), detrend = "short-linear")
plot(ospec[, "f"], 10 * log10(ospec[, "Pz"]), type = "l",
      main = "Multitaper Spectrum\nEEGdata - Pz",
      xlab = "Frequency (Hz)", ylab = "Power/Frequency (dB)")
lines(cspec[, "f"], 10 * log10(cspec[, "Pz"]), col = "red")
legend("topright", legend = c("Original", "Cleaned"), lty = 1,
       col = c("black", "red"))

gjmvanboxtel/eegr documentation built on May 20, 2023, 4:26 a.m.