How does Independent Component Analysis Preprocessing Affect EEG Microstates?

Over recent years electroencephalography (EEG) microstates research has grown exponentially in popularity given its potential to contribute to more sophisticated diagnosis, monitoring, prognosis and prevention of mental disorders in various fields, particularly clinical psychology and psychiatry (Kleinert et al. 2023). Microstates allow to investigate at a millisecond scale, the temporal dynamics of large-scale brain networks (Michel and Koenig 2018), which are characterized by a discrete number of patterns of synchrony across the cortex that remain stable for 40–120ms before rapidly transitioning to a different state (Lehmann et al. 1987). One way of computing microstate topographies is to perform k-means cluster analysis (Pascual-Marqui et al. 1995). The temporal dynamics of microstates can then be analyzed by extracting characteristics such as average duration, average number of occurrences of one particular microstate in a period of time, variance explained, and, more recently, microstates syntax and complexity (Artoni et al. 2023; Tait et al. 2020).

An analysis of published work shows remarkable similarities of microstates topographies across studies, which led to the identification of several “canonical” microstates typically observed in resting states EEG data (Koenig et al. 2023; Tarailis et al. 2023; Zanesco 2023). However, a closer analysis of published studies highlighted considerable topographic variance in the template maps assigned to the same microstate class (Michel and Koenig 2018). Such variability may depend on a variety of factors, including the type of algorithm used for microstates extraction or manual mis-classification of the class microstates belong to (Koenig et al. 2023). Another possibly relevant source of variability across studies may be the type of pipelines and artifact removal techniques performed to preprocess the data before microstates extraction. Currently, as long as microstates do resemble those found in literature in comparable studies, data are considered by authors and readers to have been preprocessed well. However, different artifact removal strategies may alter microstate topographies and resulting features.

Independent Component Analysis (ICA) is one of the most popular algorithms for removal of artifacts from EEG. ICA, in fact, allows to disentangle information linearly mixed at the scalp EEG electrodes into maximally temporally independent component (IC) processes that can be used to assess individual EEG effective source dynamics without the requirement of the definition of an explicit electrical forward problem head model (Makeig et al. 1996, 2004). Each IC is represented by a pattern of relative projection to the scalp channel (often referred to as “scalp map”) and by the time varying signed equivalent source signal, often referred to as IC time course (Delorme et al. 2012). Provided enough adequately recorded and preprocessed data are available, ICA has been found to well separate from brain data classes of stereotyped artifacts such electrocardiographic (ECG) signal contamination, scalp and neck muscles electromyographic (EMG) activities, electro-oculographic (EOG) activity as a result of lateral eye movements, eye blinks and ocular motor tremor and single-channel noise, produced by occasional disruption of the connection between scalp and electrodes (Makeig et al. 2004; Onton et al. 2006).

Using ICA artifacts removal as a microstate preprocessing step requires obtaining highly reliable extracted components and their correct interpretation and use in further analysis. Sometimes, however, noise in the data (originating from inadequate data acquisition, low-impedance scalp/sensor interfaces small, irresolvable signal sources), inadequate data sampling (e.g., not enough data points are available) and algorithmic shortcomings (e.g., convergence issues, presence of local minima) may reduce the quality of ICs and make it difficult to reliably identify any but the ICs with the highest explained variance (i.e., ocular activity)(Artoni et al. 2014; Delorme et al. 2007; Jung et al. 2000). A possible approach to preprocessing is that of performing a first, more aggressive filtering (e.g., with higher high-pass frequency to increase data stationarity), extract ICA components, and use them as spatial filters after a more conservative preprocessing procedure (e.g. lower high-pass frequency). This is often an advantage for example when low frequencies (< 1 Hz) should not be removed, e.g., in experiments involving sleep/unconsciousness (Artoni et al. 2022).

To help identify which ICs to retain or discard from further analysis, several measures have been described in literature, for example by evaluating the stability to data resampling (Artoni et al. 2014; Groppe et al. 2009), mutual information reduction (Palmer et al. 2012) or the dipolarity of IC topographies (i.e., the goodness of fit of a single equivalent dipole to the IC scalp topography)(Delorme et al. 2012). Sorting of ICs into classes is further aided by automatic classification methods such as ICLabel (Pion-Tonachini et al. 2019). Even after correct ICs identification, however, there is currently no definite set of rules as to which IC removal strategy is best suited for microstates extraction.

Broadly speaking, it is possible to identify at least four IC removal strategies in decreasing levels of conservativeness, i.e., (i) avoiding ICA preprocessing altogether (i.e., using raw data, after band-pass filtering, bad channel interpolation, and bad epochs rejection), (ii) removing ocular ICs only (blinks and eye movements), (iii) removing all ICs that could be reliably identified as physiological artifacts (e.g. heartbeat, muscles), (iv) removing all ICs except those that can be reliably identified as brain ICs based on ICLabel output probabilities and dipolarity (Delorme et al. 2012) – see Methods. The aim of this work is to test the reliability of the microstate extraction process, i.e., the stability of microstate features against these four strategies of ICA-based EEG data preprocessing with a normative resting state EEG dataset where subjects alternate eyes-open and eyes-closed periods.

The results show a general stability of microstate features independently of the preprocessing level. Figure 5 shows that except for RAW, where ICA preprocessing is skipped altogether, differences of duration, coverage, occurrence and GEV across preprocessing levels (EYES, ART and HARD), do not reach significance regardless of condition (EO and EC). Relative differences across conditions (EO and EC) are not significantly altered by the level of preprocessing, except in case it is avoided altogether. Comparing EO and EC, MS A Duration, Coverage and Occurrence are significantly different only in the case of EYES, ART, HARD. Similarly, MS C Duration, Coverage and GEV, are significantly higher in EC compared to EO, for all preprocessing levels except for RAW where the values are similar.

Interestingly, removing ocular activity when extracting microstates does not adversely alter eyes-open/eyes-closed comparisons, at least considering microstate features. On the contrary, avoiding preprocessing increases variability in the data, so much so that EO/EC differences are mostly lost (EO/EC Duration, Coverage and Occurrence for MS A, B,C are similar). The only exception might be MS D, where preprocessing increases EO/EC similarity for Duration, Coverage and Occurrence. Furthermore, the metacriteria used to assess the optimal number of microstates for RAW, does not converge to n = 5, but rather to n = 6 and n = 4 respectively for EO and EC (Fig. 3). While this was not directly tested here, it is possible that a greater number of microstates is needed to explain ocular activity variance in EO condition, which is not the case for EC (n = 4). Interestingly, in all preprocessing levels (EYES, ART and HARD), the meta criterion converges to n = 5 regardless of condition, a result closest to the most accepted normative literature (Michel and Koenig 2018), which suggests a “homogenization” effect of artifact removal (at least ocular activity) through ICA, which seems to be beneficial to the process of microstates extraction. The effect is particularly evident when observing microstate topographies (Fig. 4): except for RAW MS E, where topographic correlation is lowest (r = 0.94), topographies are almost identical across conditions, with correlations very close to 1.

The greatest feature differences in microstates are observed when comparing any level of preprocessing (EYE, ART, HARD) with respect to performing no processing at all (RAW). Indeed the microstate extraction process relies on the clustering of repetitive quasi-stable topographies across time: the higher topographical variability, lower rate of occurrence and non-stationarity of most physiological and non-physiological artifacts (such as line noise, electrode displacements, movement artifacts, etc.) likely ensures their exclusion from 2nd level (and perhaps 1st level) clustering (see microstates extraction process described in the methods section), which is not be the case for ocular artifacts. Ocular artifacts have a very precise and stationary generator as they are produced by the rotation of the eye balls, movements of the eye lids and contraction of the extraocular muscles, highly variable across subjects, with magnitudes often 10 times larger (or more) than scalp EEG data, occurring often and with quite stable scalp topographies (with maximum amplitude at frontal EEG electrodes) (Croft and Barry 2000; Dimigen 2020; Hagemann and Naumann 2001). The scalp topography of eye blinks resembles the topography of microstate C. By not excluding eye blinks in the RAW condition, they will probably be labeled as microstate C in the winner-takes-all fitting step. This explains why in RAW, duration and coverage of Microstate C are comparable (if not higher) in the EO condition, while the opposite is true when eye blinks are removed by the ICA. Other artifacts do not produce scalp topographies that resemble the canonical microstates. The fitting procedure with the threshold of > 0.5 correlation will thus not label them with any of the microstates. In this sense, clustering and fitting “naturally” remove these artifacts, making more elaborated ICA artifact removal procedures unnecessary to obtain reliable microstates. However, it may also be interesting in future works to explore these conclusions with different fitting thresholds and/or non-competitive fitting.

The results also highlight that microstate topographies are stable regardless of preprocessing level, with the HARD level of preprocessing resulting in the highest data homogenization with metacriterion convergence to exactly n = 5 microstates (Fig. 3). EO/EC differences are also preserved regardless of preprocessing (EYES, ART, HARD). However, different preprocessing strategies alter absolute values of microstate features (e.g., higher MS-A Duration/Coverage/Occurrence in EO for EYES compared to ART and HARD), although not significantly. This result further underlines the importance of carefully documenting preprocessing steps when comparing absolute values of microstate features across studies. Here convergence was reached for the extraction of 5 microstates: it is possible that the forced extraction of a larger number of microstates will also capture (and will be sensitive to) non-physiological or non-stationary artifacts. In that case, the differences between preprocessing levels might be more pronounced. Also, extracting a number of microstates outside the range suggested by the evaluation criteria (shown in Fig. 3) may reduce the reliability of the microstate extraction process and the comparison across works in literature.

These results have important implications as it becomes apparent that, for good quality resting state data, a simple ICA-based ocular artifact is sufficient to obtain reliable microstates. Such a pipeline, (e.g., by using ICLabel) can be easily automatized given the very distinct and recognizable IC scalp topographies and IC time courses of lateral and vertical ocular eye movements, further paving the way to the full automatization of the microstate extraction process.

It is important to note that the results presented have been extracted from good quality resting state data. Extreme amplitude artifact epochs should be rejected beforehand, and, if bad channels need to be interpolated, ICA should be computed after proper preliminary rank reduction (Artoni et al. 2018). Also, in case strong artifact contamination, it is possible to use a more aggressive preprocessing (ART/HARD) or else exclude such data from further analysis altogether. Indeed, it is always important to carefully consider whether to remove certain types of artifacts from the data. For example, ocular activity might be of interest in experiments involving visual tasks. In that case, ICA might be used to spatially filter the data and isolate ocular activity from other processes.

In conclusion, provided a good-quality dataset is recorded and ocular artifacts are removed beforehand, canonical microstate topographies and features on eyes open / eyes closed resting state data are robust to muscle-related and non-physiological artifacts and can automatically capture brain-related physiological data.