Detection of muscle artefact in the normal human awake EEG

Abstract

Objectives: A study was performed to investigate automatic detection of muscle artefact, using time domain and frequency domain methods. The evaluation focussed on epoch length and performance of detection.

Methods: EEG data were recorded in 21 normal adult subjects for 50 min during awake state. Investigated positions included central, temporal and parietal scalp electrodes. Expert annotation of muscle artefact was performed by accurate visual marking in a randomised test-set of the data, which allowed for intra-expert comparison. For time domain detection, the parameter set consisted of slope and maximum/minimum amplitude. Parameters in the frequency domain were absolute and relative `high beta' power (>25 Hz) and spectral edge frequency. Distributions as calculated from a reference period in each subject were used to investigate the statistics of the parameter ranges. Detection thresholds were calculated from these distributions per subject, and performance was compared to constant (empirical) thresholds for the entire data set.

Results: Results indicate a 1 s epoch length as optimal for detection of muscle artefact. The analysis using a slope threshold or absolute `high beta' power showed the best results in sensitivity (80%) and specificity (90%), matching the expert's performance.

Conclusions: Constant threshold settings performed better than statistical thresholds per subject.

Introduction

Validation of the signals recorded in electroencephalography is needed before any clinical investigation can be performed. A human expert will evaluate the recording based on his or her clinical expertise, ignoring periods of invalid signal in subsequent visual analysis. Still, the presence of artefact may have a serious effect on how well the evaluation is performed. In quantitative processing methods, results can differ significantly if artefact periods remain in the recording.

One of the major contaminations in EEG recording is the electrical signal induced by muscle activity. Most of this activity is generated in muscles in the face, neck and on the scalp, and is caused by movement, chewing, swallowing, muscle twitches, anxiety, tremor or general muscle tension. Muscle artefact mainly affects the frontal and temporal regions, but can obscure the signal on any head electrode. Analysing the EEG is not reliable in recordings that contain a substantial amount of muscle artefact. For example, Brunner et al. (1996)showed that it is important to detect short periods (4 s epochs) of muscle artefact for proper inspection of spectral features during sleep, in order to reduce the confound between cortical and muscle activity.

Relative to normal EEG activity, muscle artefact is characterised as a high-frequency phenomenon, with frequencies dominantly over 15 Hz. Filtering of the high-frequency components (e.g. >15 Hz) in general is undesirable as some underlying EEG activity is often observed by clinicians, e.g. `high' beta activity. Moreover, muscle artefact can be obscured through filtering, introducing erroneous sharp or fast wave activity. Filtering may be applied, but the original signal should always be available (Gotman et al., 1981; Barlow, 1986). In this respect, a non-linear filter described by Panych et al. (1989)was concluded as robust, and still allowed for identification of contaminated periods. However, tuning of this filter for optimal results was considered difficult.

This paper investigates the detection of muscle artefact using both frequency and time domain methods. The evaluation focussed both on detection performance and on accuracy of detection, by matching automatic marking to accurate human marking as performed by an experienced neurologist.