Principal components analysis of Laplacian waveforms as a generic method for identifying ERP generator patterns: I. Evaluation with auditory oddball tasks
Introduction
Event-related potentials (ERPs) measure time-locked field potentials extracted from the scalp-recorded electroencephalogram (EEG), and, when embedded in a suitable paradigm, allow the combined study of neuronal activity and information processing within a millisecond time resolution (e.g. Picton et al., 2000). Recent improvements in EEG technology, which enable quick application of even dense electrode montages of 128 or more recording channels, have made ERPs a readily-available, inexpensive, and non-invasive tool, rendering it among the most commonly-used psychophysiological measures for the study of human cognition (e.g. Gevins, 1998). Despite its popularity in both basic and clinical research, by comparison, less attention has been paid to two crucial methodological choices affecting the measurement of an ERP component (or any equivalent construct) and the association to its neuronal generation: (1) the procedure for identifying and quantifying relevant ERP components, and (2) the effects of an active EEG recording reference.
The construct of an ERP component is used to decompose and understand ERP waveforms by both their intracerebral origin (i.e. underlying neuronal generators) and any experimental manipulations (e.g. Picton et al., 2000), thereby associating characteristic ERP constituents with a specific function (i.e. a perceptual, attentional, or cognitive process) and a specific neuronal activation pattern. This theoretical concept of an ERP component must be distinguished from an observational definition of an ERP component (Donchin et al., 1978), ranging from very simplistic (e.g. peak amplitude, peak latency, area measurements) to more sophisticated approaches (e.g. independent component analysis). One frequently-used, systematic approach of reducing the ERP data dimensionality has been principal components analysis (PCA), which decomposes a set of ERP waveforms into a set of orthogonal constituents (e.g. Chapman and McCrary, 1995, Donchin, 1966, Donchin and Heffley, 1978, Glaser and Ruchkin, 1976, van Boxtel, 1998).
While traditional ERP peak and area measures are subject to experimenter bias (e.g. determining area integration or peak detection limits for deflections that invert and shift across scalp recording locations), PCA can instead be used as an objective, heuristic tool to determine ‘data-driven’ ERP components measures (e.g. Donchin and Heffley, 1978, Kayser and Tenke, 2003). This procedure identifies and groups unique variance patterns in the raw data, which are not necessarily evident in grand mean ERP averages, or are impossible to comprehend with ERP visualization tools available to a researcher, who is easily overwhelmed by the temporal and spatial complexity of even a modestly-scaled multichannel data set. Thus, PCA serves a two-fold purpose: to identify ERP components of relevance for a given data set, and to generate efficient measurements for these temporally and spatially overlapping components. The resulting components (i.e. factor loadings or factor waveforms) together with their associated weights (i.e. topography of factor scores) can be interpreted as observational definitions of ERP components, if their characteristics comply with common knowledge of ERP components, vary directly as a function of the experimental manipulation, or can otherwise be meaningfully related to ERP activity that is evident in the averaged waveforms (Kayser and Tenke, 2003). Such an interpretation is possible because of the a priori known organization of the data (i.e. the ERP variables submitted to the PCA are ordered in the temporal and/or spatial domain). Thus, the researcher's subjectivity is reduced to determining the appropriateness of the observed ERP measures, provided in form of PCA component scores, rather than to identifying an ERP measure and justifying its appropriateness.
We have recently shown that unrestricted PCA solutions, when combined with Varimax rotation to achieve simple structure but maintaining factor orthogonality, are particularly helpful in accomplishing this goal (Kayser and Tenke, 2003). Firstly, all restricted solutions converge on an unrestricted PCA solution, independent of the association matrix used for factor extraction (i.e. correlation or covariance matrix). Secondly, unrestricted factor extraction improves the interpretability of high-variance factors and yields stable test statistics typically performed on the factor scores (i.e. F values that are not affected but the arbitrary choice for a factor retention criterion). It is a particular advantage of this unrestricted PCA approach that components can gather variance not systematically related to the experimental manipulations (e.g. stemming from physiological and other systematic artifact sources). The separation of artifactual variance contributions from meaningful ERP variance is a very desirable PCA characteristic, as unsystematic variance that is effectively filtered from the data can no longer obscure effects of primary interest.
The use of PCA for the analysis of ERP data has been disputed because of the risk of a misallocation of variance saliently demonstrated in a simulation study by Wood and McCarthy (1984), although these authors themselves noted that traditional measures are subject to the very same pitfall. In fact, when using a more realistic test power and also simulating a realistic component topography, misallocation of variance is greatly reduced for PCA-derived component measures, and baseline-to-peak measures are equally or even more prone to this problem (Beauducel and Debener, 2003). When directly compared within the same data set, PCA-based component measures yielded larger effect sizes than time window integrals (Kayser et al., 1998) and better reliabilities than peak-based amplitudes (Beauducel et al., 2000). The use of PCA is not a protection against poor ERP data quality stemming from low signal-to-noise ratios, outliers, or temporal or spatial jitter (e.g. Chapman and McCrary, 1995, Dien, 1998a, Donchin and Heffley, 1978, van Boxtel, 1998), although it may alert the researcher to serious data problems, which sometimes may even be counteracted by exploiting the linear properties of PCA (e.g. reducing blink artifacts; Casarotto et al., 2004). Therefore, like traditional ERP measures, PCA solutions are dependent on the characteristics of the raw data, the choice of the recording reference being prominent among many possible methodological variations.
The recording of electrical activity from scalp involves the measurement of a potential difference between at least two sites, with one serving as the reference and, therefore, being arbitrarily set to zero. However, no recording site placed anywhere on the human body can be considered neutral or electrically inactive, including cephalic (e.g. mastoid, nose, ear lobe, vertex, average, etc.) and non-cephalic (sternum, neck, etc.) sites or combination thereof, and any site will be (differentially) affected by a given combination of neuronal generators through volume-conducted activity (e.g. Nunez, 1981, Nunez and Westdorp, 1994). Although equally true for all ERP components, a typical example are the generators of the auditory N1 located in dorsal-superior portions of the temporal lobe (Heschl's gyrus, AII, superior temporal gyrus, planum temporale; e.g. Liegeois-Chauvel et al., 1994, Näätänen and Picton, 1987, Pantev et al., 1995, Simson et al., 1976), which will produce different polarities, amplitudes, and even peak latencies at all recording sites when the recording reference is systematically varied within the EEG montage. The choice of the recording reference is, therefore, essential for identifying both spatial and temporal information of ERP recordings, as the reference like any other recording site will invariably reflect the spatio-temporal activation of ERP generator patterns to a certain degree. Whereas some reference choices may enhance or reduce any particular generator topography, all physically-realizable recording reference schemes, including a montage-dependent average reference, are subject to the very same reference problem (e.g. Desmedt and Tomberg, 1990, Dien, 1998b, Junghöfer et al., 1999, Pascual-Marqui and Lehmann, 1993, Tomberg et al., 1990). By acknowledging the interpretational problems stemming from an arbitrary choice of a recording reference, and to facilitate the comparison of findings across studies using a different reference scheme, ERP waveforms are sometimes rereferenced to two or more common reference schemes (e.g. Kayser et al., 1997, Kayser et al., 2003a). The use of multiple reference schemes may help to improve the appreciation of distinct ERP components, which can be differentially affected by different references.
Recently, a reference electrode standardization technique estimating a reference potential at infinity from the recorded EEG has been proposed to solve this problem (Yao, 2001, Zhai and Yao, 2004). While this new reference-free approach, which is based on an equivalent distributed source model, may be appealing and promising, several volume-conduction algorithms have previously been proposed to yield reference-free data transformations (e.g. Hjorth, 1975, Hjorth, 1980, Perrin et al., 1989, Yao, 2002a), thereby circumventing problems associated with the choice of a recording reference. Also known as current source density (CSD) transformations, these algorithms compute an estimate of the current injected radially into the skull and scalp from the underlying neuronal tissue (i.e. the scalp Laplacian) at a given surface location, from a spatially weighted sum of the potential gradients directed at this site from all recording sites (see Tenke and Kayser, 2005, for a detailed discussion). The central transformation common to all CSD algorithms is derived from the negative second spatial derivative of the interpolated scalp surface potentials, which approximates the true scalp Laplacian for low spatial frequencies (Yao, 2002b). CSD maps represent the magnitude of the radial (transcranial) current flow entering (sources) and leaving (sinks) the scalp (Nunez, 1981). The benefits of a CSD transform are a reference-free, spatially-enhanced representation of the direction, location, and intensity of current generators that underlie an ERP topography (Nicholson, 1973, Mitzdorf, 1985). CSD methods have been shown to provide an empirically useful means of simplifying the topographies of ERP components (e.g. Law et al., 1993). By virtue of the algorithm, any surface potential reference montage will produce identical CSD waveforms.
In the context of cognitive ERP research, CSD methods have largely been used to better understand the topography of radial currents that underlie the recorded surface potentials, often only as an additional visualization tool for predetermined ERP components measures (e.g. base-to-peak amplitudes or integrated time windows), thereby focusing on the spatial benefits (i.e. sharper representation, interpolation of undersampled scalp regions). For instance, we have previously applied a local Hjorth transformation to ERP measures based on principal components analysis (PCA) to identify the most representative sites within a given topography (Kayser et al., 2000a). In contrast, intracranial CSD applications have concentrated on the temporal variation of the neuronal origin of the scalp-recorded field potentials to separate the generator contributions of cortical sublaminae (Buzsaki et al., 1986, Holsheimer, 1987, Nicholson and Freeman, 1975, Mitzdorf, 1985, Schroeder et al., 1992), thereby focusing on CSD waveforms rather than CSD topographies. As the CSD algorithm is discrete in the sense that it can be independently applied to any sample point, the resulting temporal (real-time) sequence of sharpened, reference-free current flow topographies could also be used in cognitive ERP research. Although Tenke et al. (1998) used CSD waveforms to study response-related source asymmetries in an auditory oddball task, these topographic analyses were limited to time window integrals.
The present report sought to systematically and more comprehensively evaluate the possibility of combining the methodological advantages of reference-free, topographically-enhanced CSD waveforms with the virtues of unrestricted temporal PCA to identify and measure neuronal generators underlying known ERP components. For this purpose, we revisited the issue of dissociated ERP topographies for tonal and phonetic oddball tasks, and their modulation by different response requirements typical for target detection tasks.
Using a conventional 30-channel (Kayser et al., 1998, Kayser et al., 2001) or a 128-channel (Kayser et al., 2000b) EEG montage, we have repeatedly found that healthy adults show enhanced N2 and P3 amplitudes over the right lateral-temporal region for complex tones, but enhanced N2 and P3 amplitudes over the left parietal region for consonant-vowel syllables, in auditory target detection (oddball) tasks using these stimuli. As these tonal or phonetic stimuli are also known to produce opposite perceptual performance asymmetries in dichotic listening studies (e.g. Berlin et al., 1973, Bruder, 1995, Sidtis, 1981), we interpreted the stimulus-dependent N2/P3 asymmetries as electrophysiological evidence of differentially activated neuronal networks predominantly involved in pitch discrimination (right fronto-temporal) or phoneme discrimination (left parietotemporal). Findings and interpretation are consistent with evidence that N2 and P3 jointly reflect endogenous ERP activity associated with the phonemic categorization of speech stimuli (e.g. Maiste et al., 1995), and that the required cognitive task operations depend on a network of regionalized, functionally-specific subprocessors (cf. Gevins et al., 1995). This tonal/phonetic oddball paradigm has been successfully used to probe lateralized neurophysiologic processes underlying cognitive dysfunctions in psychiatric disorders, such as schizophrenia (Kayser et al., 2001) or depression and anxiety (Bruder et al., 2002).
In our previous study (Kayser et al., 1998), we reported that these task-dependent and region-specific ERP asymmetries are also modulated by response requirements (i.e. a button press to target stimuli with either the left or right hand), and that these effects are not merely due to equal asymmetrical, motor-related negativities contralateral to the response hand (e.g. Kutas and Donchin, 1980). A response-related negativity, superimposing cognitive ERP components (i.e. N2 and P3), was evident in target ERPs particularly over frontocentral brain regions. It thereby had a stronger effect on the regional topographies characterizing the tonal task than on the posterior, parietal asymmetries seen for the phonetic task. For instance, N2 amplitude was greater over left than right hemisphere sites in the phonetic task regardless of response hand, whereas the hemispheric asymmetry of N2 for the tonal task was dependent on response hand (Kayser et al., 1998). Furthermore, responding with the left hand resulted in a greater response-related contralateral negativity than responding with the right hand for right-handed healthy adults (Kayser et al., 1998, Tenke et al., 1998). At the same time, right button presses resulted in greater right-larger-than-left parietal P3 sources compared to left button presses resulting in the opposite asymmetry (Tenke et al., 1998). It would, therefore, be difficult to evaluate task-related topographic effects in the context of a classic oddball paradigm if all responses are made by one hand because motor- or response-related potentials may contaminate the findings. However, it should be carefully noted that in these studies response hand was manipulated between- rather than within-subjects, and that response-related findings may be subject to random selection effects given the relatively small sample size typical for ERP studies.
Frequently, a silent (mental) count instead of a manual response is required in target detection tasks, which avoids motor-related confounds altogether. Several studies have reported an impact of response mode requirements on prominent cognitive ERP components during target detection (e.g. Lew and Polich, 1993, Polich, 1987, Starr et al., 1997), but it is not completely clear whether these differences should be interpreted in terms of attentional resources, movement control, or both. In fact, Salisbury et al., 2001, Salisbury et al., 2004 have argued that button pressing relative to silent counting distorts the typical P3 topography (although only right button presses were studied), and that ‘increase’ in frontal P3 positivity in NoGo as opposed to Go target responses should be interpreted as a motor-related negativity rather than as a NoGo P3 enhancement reflecting active response inhibition (e.g. Fallgatter and Strik, 1999, Fallgatter et al., 2000, Roberts et al., 1994). While these findings may promote the use of a paradigm that avoids a manual response mode, it is important to recognize two main pitfalls of a silent count condition: (1) a grossly reduced insight into participants' performance (response latency and item-related accuracy) preventing the exclusion of error trials when computing ERP waveforms, both of which is particularly concerning when groups or conditions under study differ widely in performance level; and (2) the required verbal memory load of the ongoing target count, adding a lateralized, dual-task component to the oddball paradigm (e.g. Friedman and Polson, 1981).
As previous studies have used between-subjects and/or incomplete response mode comparisons (e.g. Kayser et al., 1998, Kayser et al., 2001, Salisbury et al., 2001, Salisbury et al., 2004), this study directly compared the impact of these distinct response mode requirements (silent count, right press, left press) on the topography of ERP components previously observed in tonal and phonetic target detection tasks. The predominant objective was to explore the usefulness of combining CSD and PCA methodology for disentangling known task- and response-related effects by revealing the temporal-spatial dynamic of their underlying generator patterns.
Section snippets
Participants
EEG data recorded from 66 right-handed, healthy adults (25 men) were selected for this report. These individuals had volunteered to participate in one of two ongoing research studies at the Psychophysiology Laboratory at New York State Psychiatric Institute, which had been approved by the institutional review board, for a monetary compensation of $15/hr. The experimental protocol, which was undertaken with the understanding and written consent of each participant, was identical in these two
Behavioral data
Mean response latency for correct button press responses was 20 ms faster for tones (M=476.6 ms, SD=127.8) compared with syllables (M=496.0 ms, SD=126.0; task main effect, F[1,64]=6.18, P=.02), which is in accordance with our previous findings for healthy adults (Kayser et al., 1998, Kayser et al., 2001). Conversely, the mean hit rate was approximately 2% lower for tonal (M=96.9%, SD=7.2) compared with phonetic stimuli (M=98.7%, SD=2.6; task main effect, F[1,64]=4.65, P=.03); however, this
Discussion
The current study evaluated reference-free CSD transformations of ERP surface potentials as an intermediate processing step to further improve the PCA-ERP analytic approach (e.g. Chapman and McCrary, 1995), building on the benefits of using unrestricted, covariance-based, temporal PCA with Varimax rotation (Kayser and Tenke, 2003). Since this evaluation was based on a parallel analysis of CSD/ERP data obtained from tonal and phonetic oddball paradigms known to produce distinct effects of
Acknowledgements
This research was supported by grants MH058346, MH36295, MH50715, and MH59342 from the National Institute of Mental Health (NIMH).
We greatly appreciate the assistance of Nil Bhattacharya, Carlye Griggs, Paul Leite, Mia Sage, Stewart Shankman, and Barbara Stuart with data collection, storage, and preprocessing.
Waveform plotting software was written by Charles L. Brown, III, who also provided helpful insights during the implementation of the spherical spline algorithm, benefitting from gracious
References (132)
- et al.
Cortical differences in tonal versus vowel processing as revealed by an ERP component called mismatch negativity (MMN)
Brain Lang
(1993) - et al.
Recognition of imagined hand movements with low resolution surface Laplacian and linear classifiers
Med Eng Phys
(2001) - et al.
Human auditory and somatosensory event-related potentials: effects of response condition and age
Electroencephalogr Clin Neurophysiol
(1987) - et al.
Misallocation of variance in event-related potentials: simulation studies on the effects of test power, topography, and baseline-to-peak versus principal component quantifications
J Neurosci Methods
(2003) - et al.
The neuroanatomical substrate of sound duration discrimination
Neuropsychologia
(2002) - et al.
Dichotic right ear advantage in children 5 to 13
Cortex
(1973) - et al.
Intracerebral somatosensory event-related potentials: effect of response type (button pressing versus mental counting) on P3-like potentials within the human brain
Clin Neurophysiol
(2003) - et al.
Combined event-related fMRI and intracerebral ERP study of an auditory oddball task
Neuroimage
(2005) - et al.
Laminar distribution of hippocampal rhythmic slow activity (RSA) in the behaving rat: current-source density analysis, effects of urethane and atropine
Brain Res
(1986) - et al.
Response selection and motor areas: a behavioural and electrophysiological study
Clin Neurophysiol
(2004)
Principal component analysis for reduction of ocular artefacts in event-related potentials of normal and dyslexic children
Clin Neurophysiol
EP component identification and measurement by principal components analysis
Brain Cogn
Cognitive psychophysiology: the endogenous components of the ERP
The NoGo-anteriorization as a neurophysiological standard-index for cognitive response control
Int J Psychophysiol
The future of electroencephalography in assessing neurocognitive functioning
Electroencephalogr Clin Neurophysiol
Regional modulation of high resolution evoked potentials during verbal and non-verbal matching tasks
Electroencephalogr Clin Neurophysiol
Spatiotemporal maturation of the central and lateral N1 components to tones
Brain Res Dev Brain Res
Intracerebral potentials to rare target and distractor auditory and visual stimuli. I. Superior temporal plane and parietal lobe
Electroencephalogr Clin Neurophysiol
Intracerebral potentials to rare target and distractor auditory and visual stimuli. II. Medial, lateral and posterior temporal lobe
Electroencephalogr Clin Neurophysiol
Dorsal and ventral streams: a framework for understanding aspects of the functional anatomy of language
Cognition
An on-line transformation of EEG scalp potentials into orthogonal source derivations
Electroencephalogr Clin Neurophysiol Suppl
The polar average reference effect: a bias in estimating the head surface integral in EEG recording
Clin Neurophysiol
Optimizing PCA methodology for ERP component identification and measurement: theoretical rationale and empirical evaluation
Clin Neurophysiol
Trusting in or breaking with convention: towards a renaissance of principal components analysis in electrophysiology
Clin Neurophysiol
Principal components analysis of Laplacian waveforms as a generic method for identifying ERP generator patterns: II. Adequacy of low-density estimates
Clin Neurophysiol
Brain event-related potentials (ERPs) in schizophrenia during a word recognition memory task
Int J Psychophysiol
Event-related potentials (ERPs) to hemifield presentations of emotional stimuli: differences between depressed patients and healthy adults in P3 amplitude and asymmetry
Int J Psychophysiol
Event-related brain potentials (ERPs) in schizophrenia for tonal and phonetic oddball tasks
Biol Psychiatry
Event-related brain potentials during auditory and visual word recognition memory tasks
Brain Res Cogn Brain Res
Evaluating the quality of ERP measures across recording systems: a commentary on Debener et al. (2002)
Int J Psychophysiol
Pre-movement parietal lobe input to human sensorimotor cortex
Brain Res
Overlap between P300 and movement-related-potentials: a response to Verleger
Biol Psychol
Preparation to respond as manifested by movement-related brain potentials
Brain Res
Improving spatial and temporal resolution in evoked EEG responses using surface Laplacians
Electroencephalogr Clin Neurophysiol
P300, habituation, and response mode
Physiol Behav
Evoked potentials recorded from the auditory cortex in man: evaluation and topography of the middle latency components
Electroencephalogr Clin Neurophysiol
Regulating action: alternating activation of midline frontal and motor cortical networks
Clin Neurophysiol
Frontal midline theta and the error-related negativity: neurophysiological mechanisms of action regulation
Clin Neurophysiol
Principal component analysis of event-related potentials: a note on misallocation of variance
Electroencephalogr Clin Neurophysiol
Hemispheric asymmetry of the auditory evoked N100m response in relation to the crossing point between the central sulcus and Sylvian fissure
Electroencephalogr Clin Neurophysiol
The assessment and analysis of handedness: the Edinburgh inventory
Neuropsychologia
The five percent electrode system for high-resolution EEG and ERP measurements
Clin Neurophysiol
Specific tonotopic organizations of different areas of the human auditory cortex revealed by simultaneous magnetic and electric recordings
Electroencephalogr Clin Neurophysiol
Spherical splines for scalp potential and current density mapping [Corrigenda EEG 02274, Clin Neurophysiol 1990;76:565]
Electroencephalogr Clin Neurophysiol
Response mode and P300 from auditory stimuli
Biol Psychol
Mapping P300 waves onto inhibition: go/no-go discrimination
Electroencephalogr Clin Neurophysiol
Neurophysical theory of coherence and correlations of electroencephalographic and electrocorticographic signals
J Theor Biol
Intracerebral P3-like waveforms and the length of the stimulus-response interval in a visual oddball paradigm
Clin Neurophysiol
Button-pressing affects P300 amplitude and scalp topography
Clin Neurophysiol
The NoGo P300 ‘anteriorization’ effect and response inhibition
Clin Neurophysiol
Cited by (498)
Changes in information integration and brain networks during propofol-, dexmedetomidine-, and ketamine-induced unresponsiveness
2024, British Journal of AnaesthesiaDisruption of alpha oscillation propagation in patients with schizophrenia
2024, Clinical NeurophysiologyThe neurophysiology of continuous action monitoring
2023, iScience
☆A preliminary summary of this report has been presented at the 42nd Annual Meeting of the Society for Psychophysiological Research (SPR), October 2002, Washington, DC.