Tasneem F. Hasan and William O. Tatum, IV
Many simple and complex sources of artifact exist when attempting to record physiological signals from the nervous system. Unwanted artifact is common and may be due to physiological and extraphysiological causes. When artifact obscures the ability to record the targeted neurophysiologic signal, reduction is an important goal to facilitate an accurate interpretation (Figure 14.1). Denoising is a challenge to optimize the physiological signal of interest prior to processing and analysis. All clinical neurophysiological (CNP) procedures use variations of scalp or indwelling electrodes to measure electrical potential and monitor neuronal activity. Signals are often composed of a mixture of frequencies recorded at each electrode site, amplified, and compared to comprise neuronal electrophysiological activity that is interpreted over a predetermined duration (1). Artifact reduction requires preparation during the initial stages of recording to ensure optimal electrical integrity. During the performance of the procedure, it is critical for an experienced technologist to have a heightened awareness of artifact to identify, localize, and eliminate artifact that serves as “interference” to interpretation of the physiological signal. Finally, artifact reduction techniques are essential for online and post hoc signal analysis to remove artifacts from the recorded signal of interest. Some artifacts are common to several CNP procedures (e.g., cardiac, ocular, and myogenic sources), whereas others are procedure-specific, requiring individual techniques (e.g., demagnetizing efforts for magnetoencephalography). Some require simple behavioral maneuvers such as asking the patient to relax their jaw during EEG to eliminate myogenic artifact. Others require simple techniques such as warming an extremity during electromyography (EMG). Electrophysiological techniques (e.g., averaging during recording of evoked potentials [EPs]) may be required to optimally record other CNP waveforms. Nevertheless, innovations in software development using automated techniques have improved our ability to optimize the signal-to-noise ratio (SNR) during CNP recordings. Although many forms are proprietary and evolving, the most common techniques that have been attempted to use or are using for artifact reduction are discussed in this chapter.
Artifact occurs when there is a mismatch of physiological to nonphysiologic signal that is recorded. The presence of artifact may at times be helpful to facilitate interpretation. Ocular and myogenic artifact, for example, may provide insight into the stages of sleep during polysomnography (PSG) and help differentiate wakefulness and rapid eye movement (REM) sleep during EEG and PSG (2). Artifacts are classified according to their source of origin. Those arising from the subject’s body are physiological, whereas those arising externally are nonphysiological (extracerebral) artifacts.
The common physiological artifacts reflect myogenic, ocular, and cardiac generators. Other complex signals seen during movement may reflect a combination of sources.
The impact of myogenic artifact that results from chewing/swallowing, tongue movement (glossokinetic), grimacing, and generalized muscular contraction range from rendering a CNP procedure useless to analyzing patterns that may have usefulness. It is the context of recording that reflects the significance. Muscle frequency may range from 10 to more than 100 Hz and occur with large amplitudes over 100 µV (3). Myogenic artifacts can be easily identified based on their characteristic nonoscillatory morphology and very brief duration (<20 msec), which reflect motor unit potentials (4). Excessive movement by patients during a CNP can substantially contaminate the recording. Physiological signals may closely resemble those artifacts arising from movement disorder patients. Slow (4–6 Hz), rhythmic, and sinusoidal signals in an EEG seen with Parkinson’s disease may be distinguished from faster frequencies (7–12 Hz) typically encountered with essential tremor. Differentiating artifactual from physiologic signals may at times be challenging for the clinician (Figure 14.2). However, similar to sleep staging, patterns of myogenic artifact may be useful in diagnosing psychogenic movement disorders, chiefly movements comprising tremors, jerks, spasms, and dystonia using EEG time-locked with EMG (5). Surface EMG characteristics can differentiate between a convulsive epileptic seizure and a nonepileptic behavioral event even in the absence of rhythmic clonic movements (6). A photomyoclonic response that generates myogenic artifact appearing 50 to 60 milliseconds following intermittent photic stimulation may mimic a seizure and lead to overtreatment (7).
Eye movements during CNP procedures such as EEG can disturb the electric fields in electrodes placed in the frontal head regions by generating large-amplitude alternate currents with fields that create artifact (8). Ocular artifacts range between 0 and 16 Hz and contaminate the physiological activity by introducing stronger deflections within the target signal (9). This may impact interpretation of neural signals produced during EEG, magnetoencephalography (MEG), and PSG. The amount of eye movement helps signify the stage of sleep and level of alertness reflecting a valuable contribution of artifact to overall interpretation of the recording. Fluttering of the eyelids generates a signal frequency in the theta bandwidth, while slower infrequent repetitive eyeblinking occurs within the delta range in the EEG. When excessive eye movements are encountered, they can obscure the CNP recording and interfere with the physiological signal. Challenging aspects of identifying ocular artifact may involve correct identification of the source (Figure 14.3). Therefore, monitoring eye movements and artifact reduction of intrusive eye movements during CNP recordings is essential.
Cardiac artifacts are common and may be divided into mechanical and electrophysiological artifact. They may be identified by sharp, regular waveforms synchronized to QRS complexes on the ECG channel. During various types of CNP recording, cardiac artifacts may limit interpretation of the recording or even beguile the interpreter into misinterpreting a nonphysiological cardiac signal as an abnormal one derived from the brain. These artifacts may mimic periodic discharges (PDs; lateralized or generalized) in EEG but can be distinguished by the presence of recurrent time-synched intervals between artifactual discharges and ECG while PDs lack precise periodicity of intervals between discharges (10). Cardiac artifacts may also contaminate MEG, PSG, and EP. Additionally, the presence of arrhythmia may complicate the signal further due to irregular intervals between discharges. Cardiac pacemakers produce a distinct artifact. They can be identified in electrodes as high-frequency, polyphasic potentials of short duration (10). Ballistocardiographic artifact is mechanical artifact from head/body movement during cardiac contractions that is introduced into a CNP recording (11).
Movement of the head, body, and limbs is complex and generates high-voltage signals at irregular intervals and can occur simultaneously during an act of seizure or may be due to hyperventilation/respiration, movement disorders, ocular movements, and cardiac activity.
The electrode–tissue junction is the most critical interface for generating a high-quality physiological signal. Electrode artifacts arise from contact impedance between the electrodes and tissue which may be due to excessive or dried out gel paste. Dry electrodes are superior to their gel counterpart due to patient comfort and ease of setup (e.g., minimal skin preparation time or conductive gel) and have demonstrated better utility during home-EEG recording, long-term monitoring, and during emergencies due to expedited setup (12). However, dry electrodes are subjected to contact impedance and can substantially contaminate the signal. When dry electrodes are preferred, constant monitoring of the electrode–tissue contact is warranted to minimize motion artifact during CNP (12).
Artifact may also arise from change in resistance or electrode potential between the electrode and the tissue, or can be due to a broken lead. 377Electrodes may “pop” and introduce artifact that is characterized by a sudden discharge or change in the electrical potential between the electrode and the tissue and give rise to a distinct waveform with a steep rise and slow fall (10). When several pops result over a few seconds, this suggests technical pitfalls in the recording environment (e.g., loose or nonadherent electrode). Salt bridge artifact results from excessive gel paste in between the electrodes and is channel specific, and appears as attenuated waveforms (10).
The power main artifact is typically produced at 50 or 60 Hz (60 Hz in the United States). Avoiding patient contact with nearby machines, electrical devices, and connectors or cables should be followed to prevent 50/60 Hz alternating current (AC) interference. Unplugging nearby machines in the ICU or the operating room (OR) may help eliminate the source so digital notch filters are not required. The frequencies associated with the power main are not static at a single frequency. Notch filters may produce “ringing” in the time domain and obscure the target signal using an overly narrow filter (seen during EP recording).
QUANTIZATION ERROR ARTIFACT
The ICU and the OR are common settings for quantization error artifact caused by an interaction of sampling rate and high-amplitude periodic patterns, most commonly 50- to 60-Hz artifact. The sampling rate and the high-amplitude signal have to share a common divisor. In the presence of high-amplitude 60-Hz waveforms, it is manifested as 20-Hz activity when the sampling rate is 200/sec, or as 4, 12, and 20 Hz activity when the sampling rate is 250/sec. It occurs more extensively in the ICU setting when the acquisition machine uses a sampling rate of 200/sec (Figure 14.4A). It can be present in a large percentage of recordings, especially as the recordings become increasingly contaminated with high-amplitude, 60-cycle activity (Mark Scheuer, personal communication, March 14, 2018). It may be mistaken as excess cerebral beta activity by EEG readers. Quantization artifact can contaminate both the original waveforms and quantitative EEG (QEEG) metrics and augment beta power or interfere with the alpha delta ratio when the artifactual patterns hit at 4 and 12 Hz at 256/sec sampling rates (Figure 14.4B). However, it is usually easy to recognize in trends because it generates a narrow continuous band of activity to stand out at a precise frequency bandwidth unlike cerebral activity composed of more variation. Quantization artifact is not eliminated through the use of a notch filter because it occurs as an independent signal resulting from the digitization process. It can be eliminated using independent component analysis (ICA) or similar techniques, through the use of narrow notch filters at the appropriate frequency, or by using a sampling rate that does not share a common divisor between the sampling rate and the high-amplitude signal, 60 Hz (or 50 Hz in non-U.S. areas) (Figure 14.5).
APPLICATION OF ARTIFACT TO PATIENT CARE
Most CNP recordings are subject to numerous artifacts. Contamination of CNP recordings with artifact can greatly hamper accurate interpretation and fail to differentiate between neuronal activity and artifact. This can result in compromised interpretation of the procedure and interfere with patient management. A minimum duration of artifact-free recording is suggested to evaluate baseline activity when performing a CNP procedure (13). Current advancements in technology aimed at artifact reduction help facilitate automatic detection and removal of artifact.
TECHNIQUES OF ARTIFACT REDUCTION
Artifacts arising on CNP procedures can be discarded by the use of various techniques; these may be behavioral, manual, or automated methods. Prior to utilizing these methods, employing appropriate preparation technique prior to the CNP procedure is crucial to limit the occurrence of nonphysiological artifact. The American Clinical Neurophysiology Society (ACNS) has established minimum criteria to obtain a technically sound and standardized recording in EEG, EPs, and neurophysiologic intraoperative monitoring (NIOM) that can be interpreted with consistency (14). The American Association of Neuromuscular and Electrodiagnostic Medicine (AANEM) provides evidence-based practice guidelines to enhance the diagnostic accuracy and efficiency of electrodiagnostic studies during the assessment and treatment of muscles and nerve disorders (15). The Standards of Practice Committee of the American Sleep Disorders Association (ASDA) sets practice parameters and indications for PSG in the diagnosis of sleep disorders (16), whereas the American Clinical Magnetoencephalography (MEG) Society (ACMEGS) provides minimum standard practice guidelines for the routine clinical recording and analysis of MEG and EEG in all age groups in attempt to standardize procedures and allow for exchange of recordings/reports among facilities (17).
The electrode–tissue junction, for instance, is the most critical interface for electrical conduction and is most often subjected to high impedance. Although electrodes have evolved with time and are relatively simple to use, they are responsible for significant artifact on CNP recordings. Electrode caps may be used during CNP procedures (e.g., EEG and EP). They may also be 378useful during ICU or ambulatory EEG monitoring for clinical and research purposes (18). Technicians fitting electrode caps must also have prior experience with cap placement to avoid the occurrence of artifact (19). When comparing individual electrodes with cap electrodes, they were found to be of similar quality in one study (18). For most routine CNP recordings, 1-cm diameter, cup-shaped electrodes with a flat rim are utilized and offer good contact in the presence of electroconductive gel (20). Collodion helps maintain electrode attachment for prolonged use and a piece of gauze is used to cover the electrode to establish firm contact (20). Additionally, technologists abrade the skin using an abrasive detergent to remove debris and oil from the surface of the scalp prior to attaching the electrodes. A long-standing understanding that electrodes must have low impedance (i.e., less than 5 kΩ) to eliminate unwanted electrode artifact during the recording is a fallacy. Good recording may be achieved even with high electrode–scalp impedances (21), and subdermal and PressOn electrodes can sustain even higher impedances. Thus, when artifact is difficult to overcome (e.g., in the OR and ICU), subdermal needle electrodes may be helpful to reduce artifact (22). The choice and location of the reference electrode may affect the integrity of the recording and introduce artifact (Figure 14.6). Scalp references are often more subject to artifact than an intracranial source during invasive EEG recording. Signals from the scalp are much lower in voltage than those recorded directly from the brain. Therefore, if a scalp reference is chosen, a secure scalp–electrode interface is essential to provide adequate recording. When a “noisy” scalp reference is present (Figure 14.7), data may be lost. We prefer an intracranial reference to eliminate artifact introduced by scalp recording. The location should be remote from the anticipated region of interest in a “silent” area. Some place an intracranial electrode reference on a subdural strip (usually a 1 × 4) separately at implantation of clinical electrode placement to ensure optimal integrity of the recording.
Behavioral maneuvers are the first line of defense against artifact. Subjects are asked to lie still, close their eyes, breathe regularly, and focus on stimuli when the CNP procedure is performed while awake. On the contrary, provocative behavioral techniques such as sleep deprivation may be used to increase the yield of abnormalities on CNP procedures. For EEG recording, to reduce myogenic artifact, subjects may be asked to relax their jaw or reposition their head. During EMG and nerve conduction studies, temperature-evoked artifact is common due to fluctuation in skin temperature from anxiety, especially in children, or due to air conditioning (23). Measuring limb temperatures and rewarming to 34°C to 36°C for upper limbs and 31°C to 34°C for lower limbs should be done to limit slowing nerve conduction velocities (24). Further, recording external cues during MEG recording can help with artifact reduction, while preserving the spectra and time domain waveforms of the neuromagnetic activity and without reducing the rank of the data (25).
Manual or offline rejection methods may be labor intensive for continuous and multichannel EEG recordings, but are most reliable. The American Society of Electroneurodiagnostic Technologists (ASET) Inc. has defined criteria which technologists must meet to standardize the approach to performing various neurodiagnostic procedures (e.g., EEG, EP, nerve conduction study [NCS], PSG, NIOM) in the clinic or ICU setting (26). During EEG recording, the technologist will manually inspect and eliminate or reduce artifact-affected epochs simultaneously. Manual methods are useful for offline analysis, but cannot be utilized during real-time recording. During EEG monitoring, video recording often sheds light on suspicious artifacts to elucidate waveforms that mimic abnormality.
AUTOMATED TECHNIQUES OF ARTIFACT REDUCTION
Automated techniques utilize mathematical algorithms for artifact removal. In automated subtraction, short epochs or segments of recordings are rejected when exceeding a predetermined threshold has been met. This technique can be used online or offline. However, this method is both time-consuming and results in loss of significant physiological signal. Regression software methods can help focus on specific artifacts (e.g., eye movements, cardiac artifacts) and target them to be subtracted from the CNP recording (27). However, myogenic and electrode artifacts are less amenable to reduction by regression methods due to contamination of several reference channels.
Conventional Trial Rejection
Conventional trial rejection (CTR) is a method that can be used to eliminate high-amplitude artifacts from CNP recordings. It is desirable to perform CTR when there are numerous trials such as during EP and CNP procedures 379performed during NIOM, as those trials containing artifact across all electrodes are eliminated from each channel. Consequently, CTR can be effectively used during procedures when a large number of trials require fewer artifacts. This has more limited use in the pediatric setting.
ARTIFACT REJECTION AND DIGITAL FILTERS
Artifact rejection via digital recording involves montage manipulation, filter adjustments, and speed alterations for artifact reduction (28). Algorithms, including least mean squares (LMS), use adaptive filters and require obtaining a suitable reference signal. This reference is then used against the artifact of interest to estimate the acquired signal. Further, commercially available digital filters are available for standard use, eliminating artifacts with cutoff filters varying from direct current (DC) to 512 Hz. During the preprocessing stage of the data for analysis, filtering out data in the frequency range of interest is fundamental for minimizing artifact. Nonetheless, overuse of digital filters, particularly high-frequency low-pass filters, can modify true neurophysiological signals. This is achieved by removing sharp peaks and altering the signal rhythmicity. In this case, after undergoing several degrees of filtration, some artifacts (e.g., muscle) may falsely appear as a physiologic or even pathological signal (e.g., periodic discharges). This warrants documentation and understanding filter use during CNP recordings to ensure the neurophysiologist is aware of the potential for false representation of the signal (Figure 14.8) that might be present during interpretation (27,29). Similarly, high-frequency filters may reduce artifact composed of compound muscle action potentials (CMAP) yet it also eliminates the sensory nerve action potential (SNAP) (23). Consequently, these limitations and loss of significant amounts of data during artifact rejection discourage the routine use of digital filters and warrants development of superior automated techniques.
BLIND SOURCE SEPARATION
The blind source separation (BSS) algorithm is most widely used for artifact reduction in CNP procedures (e.g., EEG, EMG, EP, PSG, MEG). It can use ICA, principal component analysis (PCA), and canonical correlation analysis (CCA) to identify waveform components. Maximum noise fraction (MNF), a relatively new technique, has also demonstrated utility in artifact reduction during EEG recording (30,31). BSS-based algorithms assume that the observed CNP recording signal is linear, composed of a mixture of different sources, and the total number sources are less than or equal to the signal of interest (32). These methods of BSS isolate the observed complex signal and separate it into source-based components. Myogenic, electrode, and ocular movement artifacts when present may obscure important portions of the CNP tracing (Figure 14.9). An effective source separation technique will avoid mixed components comprising physiological and artifactual signals (Figure 14.10). Second-order as opposed to higher-order statistics have been shown to be more effective in removing sources involving ocular artifact (33,34). Additionally, random breaks of the source component into several components should be avoidable by the algorithm when combining ICA with PCA as the preceding step for dimensional reduction (35). ICA and PCA have reported success in artifact removal because artifactual signals tend to be much stronger and therefore easier to isolate when compared to neuronal activity (36). When extracting P300 source components, ICA was not superior to PCA or MNF, and no distinguishable differences were noted between the three methods (37). By contrast, PCA was found to be the simplest technique and required the shortest computation time, whereas MNF was an order of magnitude greater than PCA and ICA (37).
Some proprietary methods of artifact reduction (Persyst Corporation, Prescott, Arizona) use advanced waveform analysis technology (e.g., P13) where BSS divides the original signal into independent source components and subsequently applies advanced machine learning to neural network signals to recognize artifactual components for removal (Figure 14.11) (38). Usage of digital filters may alter the original waveform (Figure 14.12). These algorithms involved with initial BSS continuously analyze recordings and identify common artifact waveforms channel by channel, with subsequent and automated artifact removal (38). Using BSS methods to remove myogenic artifact and neural network identification algorithms to remove eye movement artifact can effectively reduce artifact to facilitate interpretation (Figure 14.13). Because digital filters result in distortion of the original signal, complex BSS algorithms characterize a novel method of artifact reduction. Applying these techniques produces a gray ghosting of frequencies to allow the interpreter to track changes in scalp EEG and QEEG to facilitate greater accuracy in analysis of simple artifacts (Figures 14.14 and 14.15) and those that are more complex (Figures 14.16 and 14.17) (38).
One validation study compared spike detection performance of P13 to three skilled experts and demonstrated comparable results on pairwise evaluation. Mean, pairwise human sensitivities of the three skilled experts were 40.0%, 42.1%, and 51.5%, and the false-positive rates (per minute) were 0.80, 0.97, and 1.99/min, respectively. By contrast, P13 demonstrated a comparable sensitivity of 43.9% ± 29.4% and a false-positive rate of 1.65 ± 2.37/min. The authors concluded that P13 had a sensitivity in the range of human expert assessment and was statistically noninferior in criteria (39). Three iterations of Persyst® spike detector, 2004 (P11), 2012 (P12), and 2016 380(P13) were used in this study, and by contrast, each subsequent iteration also demonstrated improvement in the computational methodology for artifact detection and neural network technology. Twenty feedforward neural network rules (NN rule), where each NN rule is composed of additional 40 individuals’ NNs, help describe the spike morphology. Each NN is further composed of one to three hidden layers, depending on the complexity of the event. The NN rule takes into account the “expert-derived concept,” such that, if the morphology (e.g., amplitude and sharpness) of the spike exceeds that of background activity, it will be identified as a spike. Additionally, NNs are organized in a hierarchy pattern with three absolute thresholds built within to help fasten the removal of nonphysiological events or artifact. Unclear events are assigned a perception value, varying between 0 and 1, where uncertain events will have a value near 0.5 (39). Further, spike activity can be identified based on morphology of half-wave segments (e.g., six-half-wave model). The first two half waves refer to the activity preceding immediately after the spike, the middle two half waves describe negative or positive deflection of the spike, and the last two half waves describe the slow wave, if present. Each half wave is described by its amplitude, duration, and curvature, and can help distinguish spikes from artifact composed of rhythmic activity (39). Spikes from different channels are combined to determine the electric field of the event. Electrodes that include artifact arising from muscle, cardiac, and ocular sources are identified and eliminated (39).
Independent Component Analysis
ICA operates with the principle that all CNP recorded signals are generated by independent frequency components (IC) forming a linear pattern. ICA decomposes the original signal using the fast Fourier transformation, estimates statistically ICs, and then extracts them. This facilitates quantitative analysis, identification of different source components, and thus facilitates differentiation of artifact from true neuronal activity for removal (40). After eliminating the artifactual ICs, a clear CNP recording is rendered for interpretation. Further, traditional ICA can only be used on multichannel EEG recordings. Recently, a single-channel ICA, with time-embedded data, was introduced and has demonstrated success in separating the signal into source components (41,42).
ICA is widely used for EEG artifact reduction and can be combined with other methods for superior results (e.g., Bayesian classifier, high-order statistics) (43,44). It is frequently used for processing motion artifacts, including those arising from ocular movements, and has also demonstrated utility in reducing myogenic artifact when combined with EMG data (45). The application of ICA to EMG signal for contraction classification demonstrated that the statistical ICs extracted were more prominent and statistic variation was much less as opposed to using EMG alone, thus making ICA-EMG ideal for feature extraction (46). Further, ICA is less reliable when extracting EP data (e.g., P300), as it extracts P300, but also non-P300 components. Nonetheless, a constrained ICA (cICA) algorithm for extracting EPs and relevant components has reported to be fast, reliable, inexpensive, with an accuracy of 98.3% in detecting P300 (47). Additionally, ICA has shown success in differentiating between somatosensory and auditory brain responses during vibratory tactile stimulation (48). Further, extracting essential data from MEG recording is difficult due to large-amplitude artifacts contaminating the signal, but when combined with ICA, particularly FastICA, isolate eyeblinking/movement, cardiac, myogenic, and other artifacts from the MEG signal is made possible (48).
ICA has been effectively used with PSG for analyzing sleep disorders. When PSG data are visually analyzed, epochs grossly containing artifact observed in a sleep disorder such as sleep apnea can be identified and extracted from the data set. However, when these data are subjected to ICA, other sources of electrophysiological signals contributing to the sleep disorder may be encountered that could provide further insight into the overall mechanism of a specific disorder (49).
EEGLAB is a Matlab toolbox that works interactively in processing continuous and event-related EEG data through ICA, time-frequency analysis, or other standard methods. This open-source platform allows researchers to contribute EEGLAB “plug-ins” or extensions that can be added to the EEGLAB toolbox (50). Using these plug-ins, users can build new data processing/visualization functions using structures and conventions of EEGLAB data (51). Several EEGLAB plug-ins are currently available. For example, FIRfilt for applying linear filters to data and automatic artifact removal (AAR) toolbox, which combines state-of-the-art artifact removal methods for ocular and myogenic artifact reduction in EEG recordings. However, they should be used with caution as not all plug-ins have been assessed by EEGLAB developers. Automatic artifact detection based on the joint use of spatial and temporal features (ADJUST) is an ICA-based, automated algorithm that works as an EEGLAB plug-in and is used for artifact identification (51). ADJUST identifies artifactual ICs and combines stereotyped, artifact-specific spatial and temporal features together. By contrast, IC classification by ADJUST was comparable to a manual, expert classification (95.2% agreement of data variance), and reconstructed visual and auditory event-related potentials (ERPs) were recovered from a profoundly artifact-laden EEG recording (51). Further, Fully Automated Statistical Thresholding for EEG artifact Rejection (FASTER) is another ICA-based technique that works as an EEGLAB plug-in and has demonstrated success in detecting and eliminating artifact components 381from EEG recordings arising from functional MRI (fMRI) scanning (52). It was tested on 47 simulated and real-time EEG recordings each, and was compared to expert artifact detection and to a variant of statistical control for dense arrays of sensors (SCADS) method. FASTER demonstrated greater than 90% sensitivity and specificity for detecting ocular and myogenic artifact and identifying contaminated channels (52).
Principal Component Analysis
PCA is a multivariate technique that aims to analyze and extract statistical data by identifying a set of correlated variables; this analysis is focused on second-order statistics (53). Lower-order statistics utilize conventional techniques comprising constant, linear, and quadratic terms such as zeroth, first, and second powers, and are used in the arithmetic mean as first order and variance as second order (54). Higher-order statistics involve functions of the third or higher power of a sample and are used for estimating skewness (measure of asymmetry that can be positive or negative) and kurtosis (sharpness of the peak) in a normal distribution. Unexpectedly, higher-order statistics tend to be less robust than lower order statistics (54). Further, during artifact reduction, singular value decomposition is used to derive principle or major components, which are multiple linear independent components (temporally and spatially noncorrelated) from the CNP recording by finding the greatest variance within the signal. The original signal is then reconstructed by linearly combining these components together and by subtracting the artifactual components to render a new and clean CNP recording. However, PCA may not effectively isolate each signal source into separate components, especially when they have comparable amplitudes (55,56). Thus, ICA is often more effective than PCA in isolating independently active sources for artifact reduction (40). By contrast, when PCA and ICA were compared for ocular artifact reduction, PCA was capable of reducing but not eliminating the artifact, with significant loss in underlying CNP data. In contrast, ICA effectively eliminated the ocular artifact in EEG, including low-frequency ocular activity, with minimal loss in data. In addition, it preserved the theta, alpha, and beta band rhythmic activities in the original EEG signal. When used in EEG, a significant amount of theta activity (4–6 Hz) was removed and false alpha activity (8–10 Hz) was observed after PCA (40). Further, PCA was tested on 800 simulated ERPs, randomly weighted combinations of three 64-point components using covariance PCAs, varimax rotations, and univariate analysis of variance (ANOVAs). Results indicate that PCA was inferior and incorrectly allocated variance across the components, leading to potentially serious misinterpretation (57).
Canonical Correlation Analysis
CCA is a multivariate analysis technique to test relationships between two sets of variables. Canonical is a term specifically used in statistics to describe the analysis of latent variables that are not directly observed but which characterize multiple variables that are directly observed. CCA has been used in CNP procedures and may be especially useful for EP recording. This includes steady-state visually evoked potential–brain computer interface to detect frequency components present in EEG. It has also been suggested for the creation of spatial filters to improve SNR of EEG, in addition to analyzing ERPs (58,59). One study proposes an algorithm combining CCA and wavelets with random forests (an ensemble approach to improve performance) for the automatic removal of muscle artifact from continuous EEG recording, and report excellent performance using this approach, which eliminates the need for expert marking, reference signal recording, and visual inspection (60). When techniques of BSS are compared, ICA-based techniques appear to be widely used in CNP procedures and have demonstrated reliable and consistent results for artifact reduction (40,61–63).
EMPIRICAL MODE DECOMPOSITION
Empirical mode decomposition (EMD) is a time–frequency analysis tool used to decompose nonlinear, nonstationary signals into a set number of intrinsic mode functions (IMFs), which are basic modes directly derived from the data (64). Irregular components and artifact in the CNP recording are identified as IMFs. IMFs are composed of a large number of artifacts, on a scale from fine to coarse, are statistically determined and removed, and the remaining IMFs are combined together to recreate a clear signal (64). Traditional methods, including wave-based techniques, rely on a priori knowledge to effectively utilize these techniques, while the same is not true for EMD. One study simulating EMD-based algorithms on several different cases demonstrated success in reducing muscle artifact and those that were related to power interference (reference). Additionally, results from various other experimental studies conclude that EMD-based techniques are data driven, adaptive, highly efficient, and can be effectively applied to complex data (64). Although EMD and ICA algorithms have been utilized in analyzing MEG single-trial data and have demonstrated high performance in reducing artifact (65), EMD has also shown success in frequency band analysis of MEG data during volitional sensorimotor tasks, where extraction of IMFs has provided insight into cognitive activity, often limited by classic methods of frequency band analysis (66).
Several wavelet techniques exist, including wavelet transform (WT), discrete wavelet transform (DWT), wavelet packet transform (WPT), and stationary wavelet transform (SWT). WT is performed under the assumption that time complexity is linear, and the wavelet is an oscillatory waveform with a limited duration of low value or zero (67). The input signal undergoes wavelet decomposition where the low-pass filter produces a waveform called approximation and the high-pass filter produces a waveform called detail. The “multiresolution decomposition” role of WT thus separates the input signal into details at various scales and approximations (67). Artifact reduction can be achieved by setting a threshold for wavelet coefficients and then separating these coefficients, while preserving important data (67). MEG data from 274 sensors were effectively subjected to wavelet decomposition using three different types of wavelets (daubechies, coiflet, and adjusted Haar). Results indicated this technique to be robust for MEG denoising (68). Further, another study reported using wavelet-based technique as an adjunct to PSG to extract the correct stage of sleep for diagnosis of sleep disorder from EEG data, in addition to assisting the diagnosis of an arousal by correlating wavelet-extracted data from various CNP procedures (e.g., oximetry in PSG, EMG) (69).
Several wavelet-based techniques, developed in combination with other reduction techniques to target specific artifacts (e.g., ballistocardiographic artifact), or special studies like pervasive EEG recording, where monitoring takes place in a more natural environment with greater degrees of movement, are discussed in the following sections.
Wavelet-Based Nonlinear Noise Reduction
EEG and functional MRI (EEG–fMRI) are collectively used to achieve spatiotemporal resolution of the neuronal activity. However, interference occurs between both systems and poses challenges for its routine use. Good spatial, but poor temporal resolution can be obtained on fMRI, while the reverse is true with EEG (70). Thus, combining both systems together provides the ideal spatiotemporal resolution required for auditory perception studies (70). Although artifact-free fMRI is easy to obtain, the same is not true with EEG–fMRI recordings because the EEG system is highly sensitive to magnetic fields within the machines used to obtain fMRI (11). fMRI has the tendency to induce cardiac pulsation within a highly static magnetic field, and thus, can severely compromise EEG recordings when performed simultaneously (70).
Wavelet-based nonlinear noise reduction (WNNR) method has demonstrated success in reducing ballistocardiographic artifact during EEG–fMRI use, with preservation of a temporal relationship to the EEG signal (11). WNNR is composed of three components: (a) wavelet transformation, (b) nonlinear noise reduction, and (c) spatial average subtraction, and undergoes two stages of analysis. During the first stage, wavelet transformation separates EEG data in each channel into multiple bands. This is followed by embedding wavelet coefficients on each individual band (11). Wavelet transformation utilizes a multiresolution decomposition system to separate the data, while simultaneously preserving the local time–frequency features (11). Ballistocardiographic artifact undergoes nonlinear noise reduction through manipulation on space embedding, and after achieving reduction, the data are inversely transformed back within the time domain (11). In the second stage, the residual artifacts in the corrected data across all EEG channels are reduced using spatial average subtraction (11).
WNNR is limited by its slow speed; it requires nearly 2 hours for processing the data on MATLAB 6.0 (11). However, the use of fast K-D tree algorithm may be helpful in reducing the computational time required for WNNR data processing (71).
Combined Artifact Reduction Techniques
Complex signals from multiple sources are present in EEG and contain artifacts that have more recently been approached through combined techniques for artifact reduction. In one study, a framework for automatic rejection for eyeblink, hand movement, and head shaking artifact was developed and used an SNR-like metric calculating signal cleansing performance (72). Results indicate that the hybrid algorithms were more effective in reducing artifact compared to individual techniques alone. As per the SNR-like criterion, WPT–EMD hybrid was superior to WPT–ICA for artifact reduction in all subjects (72).
Linear techniques through adaptive filtering, including finite/infinite impulse response (FIR) filters, can be used when the source of artifact is able to be measured (e.g., electrooculogram for eye movement artifact or EMG for muscle artifact) (73–75). In pervasive EEG recording, artifact source measurement is not possible and linear techniques cannot be employed. Thus, BSS-based methods, including ICA and PCA, have demonstrated success and have been utilized in combination with WPT for artifact reduction (72).
In both WPT–ICA and WPT–EMD, WPT is first applied to all EEG channels where both the approximation and detail space are further decomposed into frequency sub-bands. Although WPT cannot extract single-frequency components, it brings it to the frequency of interest. The wavelet coefficients of all channels are then considered and the maximum standard deviation of the wavelet energy is used to identify the node with the maximum captured artifact within the decomposition (72). The coefficients of the identified node 383are rejected and a clean signal is thus produced, which is then analyzed by ICA or EMD.
In WPT–ICA, artifact components contaminating the EEG have channels separated via fast-ICA algorithm. Within an area of artifact, the temporal standard deviation of each component as identified by the amplitude threshold is used as a baseline to detect components composed of the artifact. The highest standard deviation of the artifact component is the most powerful constituent and is thus removed from the signal (72). In WPT–EMD, irregular oscillations and large-amplitude signals, which are features of all artifacts, are matched with the nonlinear signals from EMD for decomposition. Similar to ICA, the IMF with the highest temporal standard deviation and greatest entropy is considered for evaluating the hybrid index of each channel in the decontaminated and resting state EEG. The IMF with the greatest hybrid index (most significant area with artifact) is rejected during signal reconstruction to reduce artifactual interference in interpretation (72).
PEDIATRIC CLINICAL NEUROPHYSIOLOGY
CNP procedures in children are fraught with numerous challenges regarding artifact. Electrode placement during EEG in a child requires an experienced, well-trained technologist to affix electrodes using collodion (76,77). Wrapping the child’s head after electrode attachment may be done to reduce high sensor impedances during a CNP procedure (78). ACNS recommends the use of inverted silver–silver chloride disc electrodes in the pediatric setting (79). During EMG, limited tolerance to pain may necessitate topical anesthetic creams, sedation, or even general anesthesia during the procedure. The smallest commercially available needle (30-gauge, 25-mm disposable concentric needle) is often used (80). Children may be referred to sleep specialists for the evaluation of sleep disorders using PSG. Although PSG is a painless, noninvasive technique in adults, it may be terrifying for children and requires a child-friendly sleep laboratory (81). NIOM in children follows the same basic principles as in adults; however, the effect of the inhalant anesthetics can be more profound in children and produce a greater reduction in the amplitude of cortical potentials than in adults (82).
Significant artifact on CNP recording is commonly encountered in children and therefore, conventional artifact-reduction techniques used in adults cannot be extrapolated to the pediatric setting. This is due to limited duration of time in cooperating for CNP procedures, and heavy contamination of physiological signals caused by prominent motion artifact. Two automated artifact-reduction techniques, independent channel rejection (ICR), and artifact blocking (AB) algorithm, have demonstrated success in children and are discussed further.
ERPs measure the cerebral response to various stimuli (e.g., sound, speech, and music) in preverbal infants and populations where obtaining verbal response may be challenging (83). ERP estimation is obtained as an average value over numerous trials. However, the ERP amplitude is classically less than 5 µV and approximately 10 times smaller than background EEG activity and poses a challenge during ERP extraction and assessment (84). Background EEG activity in pediatrics consists of high-amplitude, slow-frequency waveforms, chiefly compromised by motion artifact (83). During ERP estimation, artifact removal is necessary but results in the loss of a large number of trials and rejection of data.
Sudden and continuous movement during EEG recording generates high-amplitude artifact, disrupts the electrode–tissue interface, and frequently introduces artifact. Challenges caused by limited monitoring time warrants usage of alternate techniques for artifact reduction in the pediatric setting. During pediatric nerve-conduction studies, stimulus artifact is obtained in nearly every SNAP recording. The absence of artifacts suggests a technical pitfall such as disconnection of the recording leads from the amplifier (23).
Furthermore, children are more susceptible to stimulus artifact due to the short distance between the recording and stimulating leads, which can be reduced by isolating the stimulator and shielding the cables, as seen with modern EMG machines (23). Additionally, avoiding short distances, drying the skin to limit impedance prior to testing, changing the electrode, and using a ground electrode between the recording and stimulating leads can help limit artifact during EMG (23). Additionally, myogenic artifact or compound motor action potential (CMAP) during EMG is common in children due to smaller limbs and close proximity of nerves to muscles and can contaminate the true signal or SNAP. Reducing the stimulating current or performing orthodromic sensory nerve testing can help remove the CMAP artifact from the recording (23).
PSG in children is not only challenging, but subjected to artifact. Pulse oximetry, used to assess oxygenation, may be contaminated with motion artifact during unattended periods overnight. Thus, correlating the pulse waveform with the saturation value would be insightful in differentiating motion artifact from true desaturation (81). Further, oximeters incorporating artifact reduction algorithms have reported good accuracy and fewer false-positives due to motion artifact (85).
384Conventional trial rejection (CTR) is often employed in adult CNP recordings and rejects trials that are composed of significant background artifact. Nonetheless, it is of limited use in the pediatric setting due to fewer trials and greater time constraint. Similarly, ICA is highly effective in eliminating ocular artifacts in adults, but due to ill-defined distribution of these artifacts in children, its utility is limited in the pediatric setting. Further, one study testing ICA on infant EEG recordings found that the ICA algorithm spread the artifacts across 14 different sources. These 14 independent sources were composed of both neuronal activity and artifactual signals at various points in time. After obtaining an ICA-corrected EEG, artifacts still persisted and removal of the 14 independent components resulted in loss of significant data (84). Similarly, utilization of other standard methods of artifact reduction in children is limited due to similar challenges. Fujioka et al. report two pediatric automated artifact-reduction techniques during CNP procedures, with favorable results (84).
INDEPENDENT CHANNEL REJECTION
Pediatric CNP procedures are subjected to significant artifact and tend to affect one or a few electrodes on any specific trial. ICR can eliminate artifact arising from a single electrode (e.g., FP1 and FP2 in EEG) or affecting a specific trial during EPs, and thus, data in the remaining channels can be further utilized (84). Thus, depending on the number of electrodes affected, each channel will offer a varied number of trials, and these trials are then averaged together. A drawback of this technique is spatial distortion due to electrode-specific elimination of trials, thus resulting in a mixed number of trials per electrode and selective representation of the EEG study (84). This effectiveness of ICR has been previously tested and has reported good results (86–89).
ARTIFACT BLOCKING ALGORITHM
An AB algorithm is largely data driven and does not depend on the number of components as in ICA (84). AB algorithms applied to EEG have two phases. The first is composed of developing a reference matrix from the data matrix and adjusts all samples of the EEG data matrix to zero as a baseline, with absolute amplitude levels surpassing a predetermined threshold. Choosing an appropriate threshold value would thus allow identification of high-amplitude artifacts. The second phase uses the reference and EEG data matrix to estimate a “smoothing matrix.” When the smoothing matrix is multiplied by the initial EEG data matrix, a new EEG data matrix is produced, minimizing artifact (84). The smoothing matrix thus also blocks high-amplitude artifacts from the initial EEG data matrix. Further, AB algorithms are a linear transformation of the original data and are similar to interpolation methods. When the reference data matrix is projected onto the initial EEG data matrix, EEG data coinciding with zero is obtained (84). An AB algorithm can be applied to continuous data to study steady state activity, or activity in various frequency bands (as opposed to ICR), and is subjected to less spatial distortion (84). It has an advantage over other EEG interpolation methods because no previous knowledge of the volume conductor model or three-dimensional scalp surfaces is necessary (90–93). Finally, an AB algorithm is especially appropriate for infants with anatomical variations including open fontanelles, as the presence of a skull breach can significantly influence the degree of volume conduction of the electrophysiological signal leading to false representation (94).
Neurophysiological function in neurobiological disorders reflects age-specific and technology-based application of artifact reduction. Several techniques and methods of artifact reduction exist, applicable in different settings (Figures 14.18 and 14.19). None are perfect in eliminating every form of artifact for all CNP procedures and are especially challenging with complex and multiple source generators. Prior techniques of artifact subtraction and elimination are increasingly being replaced with state-of-the-art techniques for simple and complex artifact reduction. Technology is advancing our ability to reduce artifact in CNP procedures though it should not replace the technologist’s and neurophysiologist’s role for artifact prevention, identification, and elimination.
The authors thank Mark Scheuer MD, Persyst® for his review and comments to improve the manuscript and for his donation of selected examples of artifact involving QEEG and blind source separation.
ADJUST—Automatic artifact detection based on the Joint Use of Spatial and Temporal features (ADJUST) is an ICA-based, automated algorithm that works as an EEGLAB plug-in.