Introduction

Processing faces is a fundamental aspect of human social interaction, emerging early in development (Simion et al. 2011). This ability enables individuals to recognize familiar faces and interpret social intentions and emotional states (Bruce and Young 1986; Schindler and Bublatzky 2020), emphasizing its pivotal role in human interactions. As a result, the underlying neural mechanisms of face processing must operate swiftly.

The temporal dynamics of these neural mechanisms have been extensively investigated using electroencephalography (EEG), with event-related potential (ERP) being the primary technique employed (Olivares et al. 2015; Yovel 2016). ERPs are obtained by averaging epochs of stimuli, revealing distinct components such as P100, N170, P200, and N250/N400 associated with various stages of face processing (Olivares et al. 2015). The P100 component reflects early perceptual face recognition, typically localized in occipital-temporal regions (Herrmann et al. 2005; Herrmann, Ehlis, Muehlberger, & FallgatteHerrmann et al. 2005a, b), while N170 is traditionally associated with face encoding processes (Bentin et al. 1996), particularly in perceiving structural features of the face (Olivares et al. 2015), typically localized in fusiform and inferior temporal regions (i.e., (Corrigan et al. 2009). While the topography of the P100 shows a posterior occipital maximum, an opposite polarity has been described for the N170 (daSilva et al. 2016). Subsequently, the P200 component is associated with second-order spatial configuration processes (Schweinberger and Neumann 2016), characterized by an occipital maximum (Kaufmann and Schweinberger 2012). Following these, further evoked potentials are usually associated with facial representations and semantic processing. The N250, implicated in repetition face effects, presents negativity peaking at posterior temporal sites, with a polarity reversal at anterior sites (Olivares et al. 2015) localized in the fusiform gyrus (Schweinberger and Neumann 2016). Lastly, the N400 is evoked by facial semantic meaning (Balconi and Pozzoli 2005; Yu et al. 2022) and characterized by central-frontal negativity and temporal regions as its primary sources (Kutas and Federmeier 2011).

All these components are modulated by specific task demands and cognitive processes involved in focusing, storing, and recognizing facial information and emotions (Mecklinger and Kamp 2023; Schindler and Bublatzky 2020). Deficits in early, middle, and late latency components during face processing have been widely documented in many psychiatric disorders (Earls et al. 2016; Feuerriegel et al. 2015; Monteiro et al. 2017). However, comparing results across studies can be challenging due to reference selection and temporal investigation windows (Murray et al. 2008).

Microstate analyses, a reference-free approach (Koenig et al. 2014) that reflects the activation of large-scale brain networks (Michel and Koenig 2018), have garnered increasing interest within the EEG research community (Michel et al. 2024; Tarailis et al. 2024). They have been extensively used to study resting-state activity, identifying a discrete number of maps associated with the classical networks identified through functional magnetic resonance imaging (fMRI) (Koenig et al. 2014; Tarailis et al. 2024). Microstate ERP analysis captures rapid, transient changes in brain activity, providing detailed temporal resolution that reveals the nuanced dynamics of cognitive processes (Michel et al. 2024; Schiller et al. 2023). It identifies stable topographical patterns of brain activity, elucidating interactions among different brain regions over time. This approach enables comprehensive analysis without reliance on predefined time windows, facilitating broader and more flexible exploration of brain activity (Murray et al. 2008). Moreover, microstate ERP analysis robustly handles variability, making it well-suited for studies involving heterogeneous populations (Bagdasarov et al. 2024; Perrottelli et al. 2023). This robustness and flexibility highlight its advantages over traditional ERP methods, offering deeper insights into brain function’s temporal and spatial aspects. For these reasons, microstate analysis is often preferred in studies aiming to uncover the complex temporal and spatial dynamics of brain activity associated with cognitive functions (Michel et al. 2024; Schiller et al. 2023). It provides a more detailed and holistic view of the brain’s functional organization and the temporal sequence of cognitive processes.

An illustrative example of the advantages of microstate analysis in making robust inferences about neural processes comes from a study investigating language production (Laganaro 2014). In this study, ERP were calculated both stimulus-aligned and response-aligned and analyzed using classical amplitude and microstate analyses. While amplitude analyses identified modulations related to the age of word acquisition (i.e., early or late), they did not elucidate the temporal dynamics of these neural processes. In contrast, microstate analyses revealed temporal shifts in identical mechanisms (i.e., the same microstates) for early versus late-acquired words, leading the author to conclude that microstate analysis provided a coherent interpretation of the neurophysiological differences underlying the acquisition of early versus late-acquired words.

Applied to ERP analyses, this approach can identify distinctive patterns between experimental conditions (Antonova et al. 2015) and groups (Perrottelli et al. 2023). A procedure that can be employed to extract microstate maps from ERP involves applying clustering analysis to the grand-averages, enabling the identification of specific sensory/cognitive steps through a data-driven approach (see Fig. 1).

Fig. 1
figure 1

Schematic representation of the steps to compute microstates on event-related potentials (ERPs) commonly used in the Cartool software (Brunet et al. 2011). The process begins with processing individual ERPs, followed by calculating the grand averages and then applying clustering algorithms (e.g., k-means) to identify distinct microstates. Hypotheses are operationalized in microstate terms based on the identified microstates corresponding to specific ERP components of interest. For instance, the figure shows the grand averages of 19 subjects during the visual processing of faces: faces with direct gaze (condition 1) and faces with averted gaze (condition 2). Based on prior knowledge and theoretical frameworks, the operationalization involves determining which microstates (number/color/letter) correspond to the ERP components of interest. This is followed by validation through back-fitting of the maps onto the original ERPs and statistical confirmation analyses

Full size image

Despite the use of microstate analyses to examine face processing, a systematic review of these findings is lacking. Therefore, this review aims to (1) systematically summarize ERP findings on face processing using microstate analyses and (2) critically evaluate the potential of microstate analyses to characterize face-related neural representations.

Methods

Protocol

This review adheres to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses for Protocols 2015 (PRISMA-P 2015). It relies on the corresponding recommendations for review development as suggested by Moher & co-workers (Moher et al. 2015). The criteria outlined below were utilized to guide the selection of research articles for inclusion in this review.

Eligibility

Type of Studies

This review includes only EEG research articles in which microstate analyses were used to study face processing.

Type of Participants

We included studies involving both healthy participants and psychiatric populations. The results are presented first as a review of literature focused exclusively on healthy subjects, followed by a review of studies conducted in the psychiatric field, without detailing the results of individual study.

Type of Tasks

As part of our inclusion criteria, we included studies using face stimuli within their behavioral paradigms. Studies where EEG activity was not examined in relation to face processing and thus not time-locked to it were excluded.

Information Sources

Research articles included in the review were sourced from the electronic databases, with search dates as follows: PubMed (3.3.2024), Google Scholar (3.3.2024), Web of Science (4.3.2024), PsychInfo (APA PsycNet; 5.3.2024), Scopus (6.3.2024). Inclusion criteria were limited to articles in English, with no restrictions on publication dates. Notably, unpublished or not peer-reviewed studies are not included in this review.

Search Strategy

A literature search was performed using the following keywords: (‘microstate’ or ‘microstates’) and (‘ERP’ or ‘EEG’) and (‘face’ or ‘faces’ or ‘facial recognition’ or ‘face memory’ or ‘facial expression’).

Study Records

Data Management

The information obtained from the studies included was recorded in structured tables and categorized into different sections (Table 1 ‘Sample study characteristics and experimental procedures’; Table 2 ‘Microstates (MS) procedures, characteristics & findings’).

Table 1 Sample study characteristics and experimental procedures
Full size table
Table 2 Microstates (MS) procedures, characteristics & findings
Full size table

Selection Process

The selection process involved screening and inclusion stages. Initially, 1734 records were identified through database searching (Fig. 2). Among these, 534 duplicates were removed during the initial search. Screening proceeded in two phases: first by title and abstract, then by full text. Two independent researchers (CB, SK) screened titles and abstracts to identify relevant records, excluding those not pertinent to the topic or methodology (microstate analyses). Subsequently, full-text articles were examined to ensure alignment with specified criteria: microstate analyses and comprehensive scientific contributions. The reviewers, CB and SK, were not blinded to study authors, affiliations, or the scientific journals where the articles were published.

Fig. 2
figure 2

The depiction of the PRISMA flowchart illustrates the outcomes at each categorization step

Full size image

Data Collection Process

Three authors (CB, SK, MFD) conducted the data extraction from the chosen articles and reviewed the entirety of the collected data.

Outcome Measures and Prioritization

The accepted research studies utilized behavioral tasks and ERP analyses. The comparators were tasks, stimuli, microstate features, and procedures.

Risk of Bias in Individual Studies

The risk of bias in individual studies was evaluated using the 14-item Kmet checklist (Kmet et al. 2004). However, as three criteria were deemed irrelevant to the main objective of this review (i.e., random allocation to the treatment group, blinding of investigators, blinding of subjects), only 11 items were considered. Each item received a score of 2 if adequately described, 1 if partially described, and 0 if not addressed. Percentage values were then calculated based on these scores, and studies were classified as strong (scores > 80%), good (70–80%), adequate (50–69%), or limited (scores < 50%). Two authors conducted the assessment independently (SK, MFD).

Results

Twenty-two studies were identified, encompassing a total of 1097 participants. Among them, 16 focused on healthy subjects, while 6 targeted psychiatric disorders. Eight studies were classified as strong (score > 95%), three as good (87–95%), and 11 as adequate (75–87%) in terms of methodological quality, according to the Kmet checklist. Limitations mainly stemmed from small sample sizes and insufficient variance estimates.

In the 16 studies involving healthy subjects, the mean age of the participants was 24.6 years, with 181 males and 181 females. One study (Darque et al. 2012) was excluded from the overall mean age calculation due to insufficient methodological information. Regarding the 6 studies on psychiatric disorders, one focused on attention-deficit/hyperactivity disorder (ADHD), one on bipolar disorder (BD) and their offspring, one on borderline personality disorders (BPD), two on autism spectrum disorder (ASD), and one on post-stress traumatic disorder (PTSD). The two ASD studies investigated effects across a wide age range (i.e., 6 months to 30 years). For the remaining four studies, the mean age was 27.6 years. Across the six studies on psychiatric disorders, 407 participants were males, and 308 were females.

Experimental Procedures and Face Stimuli

Face presentation times ranged from 85 ms to the duration of participants’ response times, with stimuli disappearing only upon the participant’s response. Among the studies, 14 utilized grayscale stimuli, four used black and white, and four employed color stimuli, with three representing realistic face colors. In most cases, face stimuli were obtained from databases, other studies, test stimuli, and the Internet. However, the source of the face stimuli was unclear in six studies. Four studies investigated emotional processing using fearful (n = 2), angry (n = 2), and happy (n = 2) expressions. Experimental paradigms are summarized in Table 1.

Microstates Pipeline

The most prevalent method for extracting microstate maps was k-means clustering applied to grand averages (n = 10). The toolboxes frequently employed were Cartool (Brunet et al. 2011) (n = 10) and Ragu (Koenig, Kottlow et al. 2011) (n = 3). Microstate procedures and characteristics are summarized in Table 2. Microstate computations were conducted within a minimum window of 100 ms up to a maximum of 1024 ms.

The reported minimum number of maps was 2, while the maximum was 18. In seven studies, the number of maps was defined using cross-validation; in two studies, a cross-validation criterion combined with the global explained variance was used; in two studies, a Krzanowski-Lai criterion was applied. Furthermore, in three studies map selection was defined by visual inspection, and three studies reported stability and discrimination criteria as the cut-off level for their selection. Finally, three studies used a combination of multiple criteria to identify the number of maps. See Table 2. For the exclusion of short-duration maps, only two of the reviewed studies reported the smoothing parameters, and three indicated temporal constraints. See Table 2 for detailed information on the parameters used.

Grand averages were utilized for clustering in 18 studies. Additionally, two studies were conducted clustering on individual ERP before fitting the resulting maps to grand averages. A smaller number of studies calculated the maps from individual ERP and then fitted the maps to the grand averages (n = 2). One study also calculated the microstates from a single grand average computed across the experimental conditions and considered groups (i.e., males and females) and then fitted the maps to the individual grand averages (Tanaka et al. 2021). For one study, the levels at which the clustering procedure was applied were not fully specified (Haartsen et al. 2022).

Concerning group comparisons, among the studies focused on healthy individuals only, one study used a single grand average calculated across two groups (i.e., low and high affective attitude) for the calculation of microstates (Pizzagalli et al. 2000). In the comparison between healthy controls (HC) and psychiatric/risk groups, most studies included the grand averages of the experimental groups separately (n = 4), with only one group considering a single grand average of the control group (i.e., No Family History/No ASD (Gui et al. 2021).

Most of the studies considered all experimental conditions together (n = 18), while a smaller number considered two experimental conditions at a time (n = 3), and only one study considered a single experimental condition (Gui et al. 2021).

In interpreting these results, it is important to consider that the distribution of topographic maps indicates the underlying activated networks, with maps of similar topographies suggesting the involvement of similar neural processes (Murray et al. 2008). Regarding the topographic maps identified at the P1 latency, there is notable convergence among studies, typically showing a maximum bilateral posterior distribution (Berchio et al. 2017; Darque et al. 2012; Dering and Donaldson 2016; Gui et al. 2021; Itier and Taylor 2004; Liang et al. 2022; Mauriello et al. 2022; Mueller and Pizzagalli 2016; Perizzolo Pointet et al. 2021; Pizzagalli et al. 2000; Prete et al. 2022; Tanaka et al. 2021; Thierry et al. 2006). However, some studies report a lateralized topography of the P1 microstate (Pizzagalli et al. 1999) or an atypical spatial configuration under certain conditions (Haartsen et al. 2022). Topographies preceding the P1 component are less frequently documented (e.g., (Darque et al. 2012; Dering and Donaldson 2016; Mauriello et al. 2022; Thierry et al. 2006), but three studies consistently identify a microstate at this latency characterized by a maximum posterior central activity (Han et al. 2020; Liang et al. 2022; Mueller and Pizzagalli 2016). For the N170, the literature reports strong convergence in the spatial configuration among studies, typically showing occipito-temporal maxima and centro-frontal positivity (Berchio et al. 2017; Haartsen et al. 2022; Itier and Taylor 2004; Liang et al. 2022; Mauriello et al. 2022; Mueller and Pizzagalli 2016; Perizzolo Pointet et al. 2021; Pizzagalli et al. 2000; Prete et al. 2022; Tanaka et al. 2021). Only one study highlights topographic changes at the N170 between experimental conditions (Darque et al. 2012). In contrast, topographies at the N1 latency exhibit greater variability depending on the tasks or stimuli employed (e.g., (Dering and Donaldson 2016; Pizzagalli et al. 1999; Thierry et al. 2006). Microstates at the P2 latency, characterized by posterior centroids, show good spatial convergence among studies (Berchio et al. 2017, 2019; Mauriello et al. 2022; Tanaka et al. 2021; Thierry et al. 2006). Topographies at the late positive component (LPC) latency, characterized by posterior maxima, also showed consistency (Liang et al. 2022; Perizzolo Pointet et al. 2021). However, it is also important to note that in some studies, topographic maps were reported as subtractions between experimental conditions, particularly at N250 and N400 latencies (Herzmann and Sommer 2007; Schacht and Sommer 2009). Additionally, two studies did not report the topographic maps of the microstates (i.e., (Latinus and Taylor 2006; Thierry et al. 2007). Finally, in one study, these maps were shown only for the time windows where significant group effects were identified (Berchio et al. 2019). These factors may hinder comparisons of maps with results from other studies or analysis of the time course of topographies.

Parameters investigated in microstates varied among studies, with the majority directly analyzing microstate parameters (n = 16), while a minority used the identified temporal windows for the subsequent scalp (n = 4) and source imaging analyses (n = 2). In most cases, microstates were associated with classical ERP components (n = 19). The number of electrodes ranged from 32 to 204, with 64 being the most prevalent (n = 7). Distributed inverse solutions were the most frequently utilized analyses for source localization (n = 12) (i.e., LORETA (Pascual-Marqui et al. 2002), eLORETA (Pascual-Marqui 2007), sLORETA (Pascual-Marqui 2002), LAURA (Grave de Peralta Menendez, Gonzalez Andino, Lantz, Michel, & Landis, 2001).

Microstate Findings Related to Face Processing

In this section, we summarize findings regarding EEG microstate analyses in the context of face processing. Selected studies are categorized by macro-area, with separate subsections discussing findings pertaining to psychiatric disorders. Detailed information regarding stimulus characteristics and microstate procedures and characteristics are summarized in Tables 1 and 2, respectively.

Sustained/Passive Attention

Four of the reviewed studies included experimental conditions requiring responses to ensure participants actively (or at least passively) attended to the face stimuli.

In the study by Darque et al. (Darque et al. 2012), participants performed a dual task involving the identification of shapes and faces presented upright or inverted. They were asked to detect the position of the stimulus’s eyes. In a single task, participants only detected face targets. Microstate analysis revealed differences in map duration between tasks, with upright faces inducing 6 microstates in both dual and single tasks, and inverted faces inducing 7 microstates in the dual task and 5 in the single task. Furthermore, only the single task elicited two maps resembling the N170 and the P3 topographies. Notably, differences in map duration disappeared when target stimuli were presented with a long temporal lag. LAURA source imaging showed comparable activations for upright and inverted faces, with occipital and right temporal lobes being activated in various microstates.

In another study (Dering and Donaldson 2016), faces with full frontal view and butterflies were presented in an oddball paradigm, where deviant stimuli were distinguished based on their color. Microstates during the P1 component distinguished faces from non-face stimuli (i.e., butterflies), with one specific microstate explaining variance better for target faces and another showing a higher proportion of variance for butterfly target stimuli.

Latinus and Taylor (Latinus and Taylor 2006) investigated upright and inverted faces in a face vs. non-face detection task using various types of stimuli (i.e., grayscale photographic faces, schematic faces, mooney faces, scrambled non-face stimuli). Four microstates differentiated among the conditions, with consistent microstate maps for P1 and N170 across all conditions. The temporal windows identified by the microstates were used to explore brain sources, with LAURA inverse solution indicating distinct activation patterns for different stimuli types.

In the study by Itier and Taylor (Itier and Taylor 2004), upright and inverted faces were presented alongside non-face stimuli, with participants passively viewing the stimuli. Participants were instructed to press a key upon the appearance of a checkerboard stimulus. Microstate analysis identified five microstates for each stimulus category, with passive attention to faces resulting in an additional microstate at the latency of the N170, showing delayed and larger global field power (GFP) amplitudes for inverted faces compared to upright faces.

Repetition Effects

Four studies investigated repetition effects and face processing using microstate analyses.

In a study by Herzmann and Sommer (Herzmann and Sommer 2010) participants underwent a learning session followed by a priming experiment one week later. During the priming experiment, familiar faces (i.e., faces learned with or without biographical knowledge, and famous faces) and unfamiliar faces were presented. The repetition effect for familiar faces induced 4 microstates, while unfamiliar faces induced to 2 maps, associated with early repetition effects (‘ERN’/N250) and late repetition effects (‘LRN’/ N400). Old/new effects identified 3 microstates for familiar faces learned either with biographical knowledge or without and for famous faces. A first microstate was identified at 250–350 ms, followed by two segments believed to correspond to the FN400 and the LPC. Old/new effects were significant across all three conditions. While this study did not directly analyze the parameters of microstates, the results of ERP amplitudes indicated priming effects on N250/N400 microstate time windows for familiar faces compared to unfamiliar ones.

Effects of long-term memory on face-related microstates were investigated using a learning and recognition design (Herzmann and Sommer 2007). Participants performed a recognition task one week after a learning session, where familiar and unfamiliar faces/names were presented. Target stimuli were preceded by either a primed condition (the same stimulus) or an unprimed condition (an unrelated stimulus). Four microstates were identified post-target onset, associated with the N250, N400, NF400, and LPC components. Topography and amplitude analyses revealed differences between learned and unlearned stimuli with priming effects observed for learned faces on the time window of the N250.

Mueller & Pizzagalli (Mueller and Pizzagalli 2016) investigated fear conditioning using neutral faces as conditioned stimuli. After exposure to the fear conditioning task, face stimuli were presented again 24 h after (recent recall task) or approximately 1 year later (remote recall task). Faces elicited 5 microstates in both recall tasks with microstates 3, 4 and 5 were associated with the C1, P100 and N170 components, respectively. Each microstate was studied using LORETA, revealing that fear-conditioned faces elicited rapid activation in the proximity of the fusiform gyrus in both the recent recall tests (i.e., 41–55 ms) and the remote recall tests (i.e., 45–90 ms).

Thierry and colleagues (Thierry et al. 2007) compared ERP microstates during the viewing of faces, cars, and butterflies while manipulating interstimulus perceptual variance. They conducted two 1-back experiments. Five microstates were identified across the experimental conditions. In the first experiment, distinct microstates were found at the P1 for faces and cars, with the same microstate observed for the N170 but appearing earlier for faces. Differences between P1 microstates were confirmed by assessing the amount of explained variance. Furthermore, the GFP of the N170 microstate was higher with low interstimulus perceptual variance stimuli than with high interstimulus perceptual variance stimuli. In the second experiment, different microstates were found between faces and butterflies at the P1, confirmed by overall variance assessment. For the N170, microstate maps were similar across conditions, yet the GFP was higher for stimuli with slow interstimulus perceptual variance.

In another study (Thierry et al. 2006), participants performed a 1-back task during which they viewed faces, bodies and objects. Seven microstates were identified across conditions: the first two microstates were related to the P1 and N1 components. The variance explained by microstates was compared across experimental conditions, revealing differences between categories at the N1. For the N1 microstate, the LAURA inverse solution revealed that faces and bodies elicited stronger activations in the right temporal-occipital areas, although the generators were more ventral for faces compared to bodies.

Affective/Emotional Valence

Seven studies investigated affective values (n = 3) or emotional processing (n = 4) using microstate analyses.

In the study of Han and colleagues (Han et al. 2020), participants were asked to judge the attractiveness of new faces or faces already presented. The authors identified six microstates that were not associated with specific components. The duration and occurrence of a specific class of microstate, occurring before 200 ms with frontal positive polarity, were found to be negatively correlated with attractiveness judgment. For highly attractive faces, sLORETA revealed activity in the supramarginal gyrus and the superior temporal gyrus, while for moderately attractive faces, it revealed activity in the middle frontal gyrus.

In a pioneering study, Pizzagalli and colleagues (Pizzagalli et al. 1999) used grayscale faces of psychiatric patients to induce affective judgments. Face stimuli were passively presented either to the right or left of a fixation cross. After ERP recordings, participants evaluated the affective appeal of each face, indicating whether they liked or disliked it. Microstate analyses were employed to differentiate between liked and disliked faces, as well as left and right visual stimulations. Three distinct microstates were identified within a 250-ms time window, with the first two labelled as P1 and N1. Liked vs. disliked faces significantly modulated microstates centroids at P1 after right hemisphere stimulation and at N1 after left hemisphere stimulation.

Grayscale faces of psychiatric patients were investigated in a passive task, with participants subdivided into two groups based on their affective attitudes: those with negative affective attitudes and those with positive affective attitudes (Pizzagalli et al. 2000). Group assignment was determined through a post-rating task assessing the affective valence of faces. Nine microstates were identified, albeit not labeled according to classical ERP literature, and compared on their centroids. Differences between groups emerged in microstates occurring at 132–196 ms and 196–272 ms. LORETA analyses additionally showed that individuals with negative attitudes exhibited stronger activations than those with positive attitudes in temporal, postcentral, and cingulate regions at 132–196 ms, and in a right temporal-parietal lateralized network at 196–272 ms.

In another study (Liang et al. 2022), participants engaged in two experiments: one where they judged facial expressions (neutral or fearful), and another where they had to evaluate whether a bar was positioned below or above the eyes or the mouth of facial stimuli. This study also investigated the effect of continuous or interrupted sounds during face processing. Six microstates were identified, with maps 2, 3, and 5 labeled as P1, N170, and LPC, respectively. Similar microstates were found in both experiments within the first 400 ms, but divergence occurred afterward in the emotional judgment task. Variations in the temporal parameters (i.e., onset, offset and duration) of microstates 1, 3, 4 and 5 were also observed among faces with neutral content and fearful faces accompanied by interrupted or uninterrupted sounds. For the map observed after 400 ms, sLORETA showed stronger activity in the right insula and right superior temporal gyrus in the interrupted than in the continuous condition.

In the study conducted by Prete and colleagues (Prete et al. 2022), participants were presented with happy and angry faces, either unilaterally or bilaterally. Following a 400-ms stimulus onset, four microstates were identified although they were not classified according to classical ERP literature. Notably, differences in duration, time coverage, and occurrence were detected only for happy faces. In the unilateral presentation, emotional faces induced a higher occurrence of the first microstate when presented in the left than in the right field. Additionally, happy faces induced a higher occurrence in the second microstate than angry faces. Using eLORETA, source localization revealed the first two microstates in the right and left temporal areas, the third in bilateral temporo-parietal regions, and the fourth in the frontoparietal network.

Schacht & Sommer (Schacht and Sommer 2009) investigated microstate modulation in a face and lexicality decision task. Participants were presented with face stimuli exhibiting happy, angry or neutral emotions, and were tasked with determining whether the faces were intact. Stimuli disappeared upon participants’ responses. Five microstates were identified for face stimuli, and subsequently used for ERP amplitude analyses. The authors hypothesized that these first microstates might overlap with the early posterior negativity (EPN) and the subsequent with the LPC. Effects of emotions were identified in early and in later microstates. For the EPN microstates, the topographies showed an enhanced frontal positivity for happy faces compared to angry and neutral faces. Conversely, the late microstate, associated with the LPC, showed an enhanced posterior positivity for angry faces compared to neutral and happy ones.

In a study conducted by Tanaka et al. (Tanaka et al. 2021), an affective subliminal priming task was employed to compare male and female participants. They were presented with a face prime stimulus, either neutral or fearful, followed by a target displaying expressions of neutrality, fearfulness, or ambiguous fear. The task was to determine whether the target exhibited a neutral or fearful expression. Across both male and female participants, 18 microstates were identified. Interestingly, two microstates were consistently identified at the P1 latency regardless of facial expression. Fearful faces elicited distinct microstates in male and female participants at the N170. Moreover, variations among all face stimuli were observed at the P2 latency in male participants, whereas differences were identified solely between neutral target faces in female participants. Microstate time windows were leveraged to investigate and statistically assess ERP features such as amplitudes and latencies of the P1, N170, and P2. The ERP findings revealed differences in N170 and P200 microstate time windows associated with biological sex and neutral, ambiguous fearful, and fearful facial expressions.

Face Processing in Psychiatric Populations

Six studies investigated face processing in psychiatric populations and at-risk individuals using microstate analyses. Three studies utilized the same experimental paradigm and stimuli: a 2-back working memory task, with faces displaying both direct and averted gaze (i.e., (Berchio et al. 2017, 2019; Mauriello et al. 2022)).

In the study by Berchio et al. (Berchio et al. 2019), individuals diagnosed with BD and BD offspring were compared with healthy controls (HC) matched for age and gender. Group comparisons revealed six microstates, and differences between the two BD groups and HC were observed for the P2 microstate in terms of occurrence and explained variance. LAURA inverse solutions revealed that for faces with direct gaze, BD patients had less activation in the left superior frontal gyrus and supplementary motor area compared with HC, while BD offspring showed greater activation compared to HC in the left orbital frontal cortex. For faces with averted gaze, BD patients showed reduced activation in the left pre-central lobe, medial cingulate cortex, and caudate nucleus whereas BD offspring showed lower activation in the ventral premotor cortex compared to HC.

Berchio et al. (Berchio et al. 2017) investigated BPD women and HC, finding six microstates. Differences between groups were found at the N170 and P200 microstates (i.e., duration), regardless of gaze direction. LAURA inverse solution revealed that compared to HC, BPD women showed increased activation in the anterior cingulate and prefrontal regions for the N170 microstates, and reduced activation in the right medial temporal lobe at the P200.

Comparing ADHD patients and HC Mauriello et al. (Mauriello et al. 2022) identified six microstates. Discrepancies between the groups during face processing with direct gaze were observed at the P200 and N250 microstates in terms of mean correlation and duration. LORETA inverse solution showed that for the P200, ADHD patients showed reduced activations in cerebellar regions compared with HC. For the N250, ADHD patients showed reduced activations in the left posterior cingulum, left calcarine, left and right lingual and cerebellar regions.

In the study conducted by Gui and colleagues (Gui et al. 2021) a combination of top-down and bottom-up data-driven approaches was employed to investigate face processing and familial risk for ASD. EEG data were collected from infants with familial risk for ASD and infants without familial risk. The task involved passive observation of faces with direct gaze, averted gaze, and a control condition. Four microstates were identified on the grand-average of face stimuli with direct gaze in infants without familial risk, and these were used for the back-fitting in all groups. The analysis focused on the time window of the negative component ‘Nc’ (300–794). Using a support vector machine classifier, the authors demonstrated that microstate features, specifically duration and GFP, were capable of predicting a later diagnosis of ASD.

Haartsen and colleagues (Haartsen et al. 2022) investigated passive observation of upright and inverted faces in a large sample of children, adolescents and adults with and without ASD. Within a temporal window of 800 ms post-stimulus onset, 7 microstates were identified for children, 5 for adolescents and 6 for adults. Early microstates were associated with P1 and N170 components, but subsequent stages were not labelled. In response to face inversion, autistic children showed weaker modulation of microstates across the entire time window. Autistic adolescents showed weaker modulation than HC in an early-stage microstate at the P1, in response to face inversion. Similarly, autistic adults showed weaker modulation of an early-stage microstate (i.e. P1 latency) in response to face inversion, and they did not show additional involvement of a later microstate as seen in HC.

Finally, ERP data were recorded from 16 women with lifetime PTSD and 14 HC during a face-evaluation task (Perizzolo Pointet et al. 2021). The stimuli consisted of male face avatars varying in their degree of threat as rated along dimensions of dominance and trustworthiness. Eight microstates were computed for five levels of dominance and trustworthiness, and linked to the P1, N170, and late positive potential (LPP). Differences between groups were observed in explained variance for an N170 microstate with non-dominant avatars, and in the number of time frames at the LPP for ‘relatively dominant’” avatars. Differences between PTSD and HC were also found on number of time frames at the N170 for ‘relatively trustworthy’ avatars. LAURA inverse solution showed increased activation in the limbic system of PTSD women in response to non-threatening male avatars. Moreover, in response to ‘relatively trustworthy’ avatars, PTSD women showed dysfunctional involvement of the anterior prefrontal cortex, in comparison to HC.

Discussion

The reviewed studies have provided substantial evidence supporting the efficacy of microstate analysis methodologies in delineating various stages of face processing, including attentional, priming or learning, affective, and emotional processes. Moreover, these analyses have proven their utility in identifying analogous processing stages across different psychiatric disorders.

By integrating the findings from these studies, a comprehensive model of facial perception can be constructed, considering its relationship with attention, repetition, affective, and emotional processes. This model can be informed by established frameworks in the field of face perception, such as those proposed by Bruce and Young (Bruce and Young 1986), Haxby et al. (Haxby et al. 2000), Schweinberger and Burton (Schweinberger and Burton 2003), and Schweinberger and Neumann (Schweinberger and Neumann 2016). Refer to Fig. 3 for a visual representation. Overall, the evidence reviewed in this paper suggests that various microstates are consistently associated with distinct ERP stages involved in face processing providing valuable insights into the underlying neural mechanisms of facial perception.

Fig. 3
figure 3

Figure 1: Cognitive model of face perception. (adapted from Bruce and Young 1986; Haxby et al. 2000; Schweinberger and Burton 2003; Schweinberger and Neumann 2016). On the left, temporal processes related to face processing are depicted; in the center, the time course is shown; on the right, the potential involvement of microstates highlighted by studies in healthy individuals and psychiatric populations (autism spectrum disorders, ASD; bipolar disorder, BD; borderline personality disorder, BPD; attention deficit hyperactivity disorder, ADHD; post-traumatic stress disorder, PTSD)

Full size image

Indeed, the microstate findings related to passive or sustained attention to faces have yielded several noteworthy insights. Across various studies, microstate analyses have successfully discriminated specific global patterns involved in facial processing, revealing the following key insights:

  1. 1.

    Low-level visual analysis: Microstate analysis has demonstrated the ability to track stable patterns corresponding to early visual processing stages, such as the P1 component (Dering and Donaldson 2016; Pizzagalli et al. 1999; Thierry et al. 2007). This analysis provides more detailed information on the spatial and temporal dynamics of low-level visual processing than can be inferred from the P1 component alone.

  2. 2.

    Stimuli encoding: Studies by Darque et al. (Darque et al. 2012) and Itier & Taylor (Itier and Taylor 2004) have shown that microstate analysis captures specific and distinct states that span the duration of components related to the encoding of facial stimuli. This analysis highlights how the brain selectively focuses on specific stimuli while ignoring others, providing insights into the N170 component associated with facial features and configuration processing.

  3. 3.

    Analytical processing: Microstate analysis has further elucidated large-scale brain networks involved in more analytical processing of facial stimuli beyond basic visual encoding. For instance, microstates identified at the P2 latency have been linked to different brain generators for schematic faces compared to photographic faces (Latinus and Taylor 2006). This highlights the complex cognitive processes underpinning facial perception.

Overall, these findings underscore the utility of microstate analysis in dissecting global functional states of facial perception, revealing distinct stages of processing from low-level visual analysis to more analytical cognitive processes.

Investigations of face repetition effects using microstate analysis have yielded significant discoveries, shedding light on various stages of face processing:

  1. 1.

    Early priming effects: Microstates are sensitive to early priming effects, manifesting as differences in the activation patterns of microstates following repeated exposure to face stimuli. This finding indicates the rapid modulation of neural processing in response to familiar faces (Mueller and Pizzagalli 2016).

  2. 2.

    Discrimination of face vs. non-face categories: Microstate analysis has been instrumental in distinguishing between faces and other stimulus categories, particularly at early stages of visual analysis such as the P1 component (Dering and Donaldson 2016; Thierry et al. 2007).

  3. 3.

    Facial recognition and person identity: Microstate analysis has also been effective in identifying stages related to facial recognition and person identity, particularly associated with the N250 and N400 components. These components reflect processes related to memory retrieval and recognition of familiar faces, as revealed in studies by Herzmann & Sommer (Herzmann and Sommer 2007, 2010).

  4. 4.

    Controversial findings at face detection stages: Some studies have reported controversial findings regarding microstates associated with face detection stages, such as the N1 and N170 components. Thierry et al. (Thierry et al. 2006, 2007) observed inconsistencies in the microstate patterns corresponding to these components, suggesting potential specificity under certain task conditions. These findings may also suggest an increased sensitivity at these stages of microstates in capturing the cognitive process of selectively focusing on specific stimuli, while showing reduced sensitivity in identifying and distinguishing individual faces from other categories. Further research is needed to clarify the circumstances under which microstates reliably reflect face detection processes.

In our investigation, we explored evidence highlighting the sensitivity of various microstates to different types of affective and emotional information conveyed by facial expressions. The research findings we examined reveal:

  1. 1.

    Influence of facial attractiveness: Modulations of microstates to low-level visual processes and face encoding (Han et al. 2020; Pizzagalli et al. 1999) suggest that the aesthetic appeal of faces can modulate neural activity at distinct stages of visual perception.

  2. 2.

    Influence of affective attitudes: Modulations of microstates to emotional preferences (Pizzagalli et al. 2000) indicate their influence on the neural processing of facial stimuli at early stages of perception.

  3. 3.

    Gender-specific effects: Microstate variances in response to fearful facial expressions between males and females (Tanaka et al. 2021) hold potential for distinguishing gender-specific emotional face processing stages.

  4. 4.

    Differentiation of emotional expressions during face representation stages: A series of evidence (Liang et al. 2022; Prete et al. 2022; Schacht and Sommer 2009) highlighted the ability of microstates to differentiate among various emotional expressions, including anger, happiness, and neutrality, during stages of facial representation for recognition. These observations imply that microstate analysis is adept at capturing neural patterns underlying to the processing of emotional content conveyed by facial stimuli.

By identifying specific microstates associated with different stages of emotional face processing, this approach contributes to our understanding of how the brain responds to facial expressions and emotional cues.

In a dedicated section collecting studies involving psychiatric populations and at-risk individuals, our review reports the following findings:

  1. 1.

    Dysfunctional attentional mechanisms in autism spectrum disorders (ASD): In children, adolescents, and adults with ASD, passive observation of face inversion elicited weaker modulation of microstates associated with low-level visual analysis and face detection (i.e., P1, N170; (Haartsen et al. 2022). Additionally, in a study focusing on infants at familiar risk of ASD, characteristics of microstates during semantic representation in direct gaze processing were found to predict later ASD diagnosis (Gui et al. 2021). These findings shed light on how microstates can identify sequential impaired attentional neural mechanisms at both visual and higher-order stages of face processing in ASD.

  2. 2.

    Facial recognition deficits in emotional dysregulation disorders: The results on facial repetition effects indicate that microstate topographies can differentiate differences between individuals with emotional dysregulation disorders, those at high risk, and healthy controls (HC) (Berchio et al. 2017, 2019; Mauriello et al. 2022). Specifically, studies investigating bipolar disorder (BD), borderline personality disorder (BPD) and attention deficit hyperactivity disorder (ADHD) have revealed distinct microstate patterns at various stages of face processing compared to HC. For instance, individuals with BD and those at high risk for BD, exhibited alterations in facial feature processing (P2 stage), while differences in face detection (N170) and structural encoding (P2) stages were noted in women with BPD.

  3. 3.

    Face encoding and semantic dysfunction in post-traumatic stress disorder (PTSD): Microstates associated with face detection (N170) and semantic information processing (LPC) are capable of distinguishing between women with and without PTSD during affective judgments (Perizzolo Pointet et al. 2021), providing insights into the pathological processes related to PTSD.

The reported evidence sheds light on the potential of ERP microstate analysis, but certain clarifications are needed, especially regarding the diversity of paradigms used (See Table 1). Variations in tasks, stimulus duration, and perceptual attributes can lead to different ERP modulations and microstate characteristics. This diversity poses challenges in directly comparing EEG topographic maps across studies and computing spatial similarities. An important consideration is the feasibility of such comparison across studies (Koenig et al., 2024; Michel et al. 2024). To tackle this issue, one strategy is to provide microstate map templates on online repositories as reference maps for future studies, allowing researchers to compare their findings with a standardized template (Koenig et al., 2024). Nevertheless, this approach remains underutilized within the ERP community and should be considered as a forthcoming challenge for bolstering the implementation of microstate ERP analysis methodologies.

Another aspect pertains to the procedures of the EEG microstate analysis that are evolved over time, reflecting advancements in methodology and technology (Khanna et al. 2015). Initially, studies focused on the location of centroids (Pizzagalli et al. 1999, 2000) with maps primarily identified through Global Map Dissimilarity (GMD) (Herzmann and Sommer 2007, 2010; Mueller and Pizzagalli 2016; Schacht and Sommer 2009). However, there has been a gradual transition towards more sophisticated and data-driven approaches for identifying microstate periods. Currently, k-means clustering stands out as one of the most widely employed methods for microstate analysis across various EEG analysis toolboxes (e.g., (Darque et al. 2012; Liang et al. 2022; Tanaka et al. 2021)). This approach allows for the identification of distinct microstate periods based on their temporal characteristics, providing a more nuanced understanding of brain dynamics during cognitive tasks.

In this review, we observed that the most validated procedure to date is calculating maps from the grand averages and then using these maps for back-fitting (see Fig. 1; Table 2). Our review highlights variations in these procedures, which could stem from methodological choices or the toolboxes used. For instance, utilizing a single grand average across groups or an average of the HC group for subsequent back-fitting could help obtain less variable distributions in microstate parameters among datasets. However, in such cases, it is important to report measures of spatial correlations between the maps of the various groups considered to fully assess the presence or absence of potential microstate differences between them. A similar aspect concerns considering experimental conditions together or in discrete groups. This approach can offer advantages in terms of reducing variability in comparisons. However, if adopted, it should be justified by valid methodological reasons.

Methodological choices must also be addressed in the selection of maps. While some of the earlier pioneering studies used more arbitrary procedures, most studies have employed cross-validation procedures or the use of multiple combined criteria (see Table 2). For microstate resting-state data, it has been suggested that using combined criteria could lead to more stable solutions across studies (Michel and Koenig 2018). In line with this reasoning, the optimal number of ERP clusters should be estimated for each dataset using robust optimization criteria rather than arbitrary decisions.

In a similar manner, different temporal smoothing parameters used in various studies may influence the results of microstate analysis. If no smoothing parameters are applied, this can lead to shorter global durations of the microstates (Michel and Koenig 2018; Poulsen et al. 2018). Not using temporal smoothing parameters could introduce biases, such as including maps characterized by physiologically implausible durations. Furthermore, various maps identified across different studies may simply reflect the choice of different temporal smoothing parameters. Among the studies reviewed, only two reported the smoothing parameters, and three indicated temporal constraints for small maps rejection. Several where such information was not reported likely used the default criteria of the toolbox utilized. However, for transparency, this type of information should be systematically reported.

Moreover, significant variability persists in the features reported across different studies. Some studies only report a subset of microstate parameters (Berchio et al. 2017; Thierry et al. 2006), while others provide several measures (i.e., (Han et al. 2020; Perizzolo Pointet et al. 2021). In this framework, it’s also important to note how a study has shown that specific microstate parameters could be useful clinical predictors (Gui et al. 2021). Therefore, this analysis could represent an effective dimensionality reduction strategy for machine learning algorithms. Nevertheless, in some cases, microstate parameters are not directly examined, but the identified temporal windows are used for subsequent scalp (Schacht and Sommer 2009; Tanaka et al. 2021) and source imaging analyses (Latinus and Taylor 2006; Mueller and Pizzagalli 2016)).

The lack of clear guidelines on microstate analysis procedures may pose challenges for the reproducibility of findings and hinder comparisons across studies. Moving forward, efforts to establish standardized protocols and reporting guidelines for microstate ERP analysis could help address these challenges and enhance the reliability and validity of research findings in the field.

Furthermore, it is essential to consider that identical topographies may underlie different brain generators (Michel and Koenig 2018). Several studies discussed in this review have applied linear and distributed inverse solutions to estimate the underlying neural sources associated with microstate activity. These source imaging results have revealed networks consistent with those identified using functional magnetic resonance imaging (fMRI) during face processing (Han et al. 2020; Latinus and Taylor 2006; Mauriello et al. 2022; Mueller and Pizzagalli 2016; Prete et al. 2022; Thierry et al. 2006). Whenever feasible, source imaging data should be reported to facilitate cross-comparisons with different neuroimaging techniques, such as fMRI or magnetoencephalography (MEG). Implementing microstate analyses directly in inverse space represents a promising approach to further address the inverse problem (Tait and Zhang 2022). By integrating microstate analysis with source imaging techniques, researchers may achieve more accurate localization of brain activity and enhance the interpretability of EEG and ERP findings. This approach, if well validated, would represent a robust methodology to address the so-called ill-posed problem in solving the inverse problem.

The limited number of studies focusing on developmental ages, biological sex differences, and a comprehensive range of emotions represents a limitation of the current review. While insights into developmental age effects can be gleaned from control groups in clinical studies, dedicated research is needed to fully understand the nuances of microstate dynamics in these populations.

In summary, the review underscores the value of ERP microstate analysis in investigating face processing, demonstrating a refinement of methodologies over time towards more automated and data-driven approaches. Overall, microstate analysis emerges as a valid method for studying global brain patterns related to cognitive functions and distinguishing processes in clinical populations. However, it also highlights the imperative to develop standardized analysis procedures to facilitate cross-study comparisons. Moreover, recent research suggests its potential in informing machine learning applications (Gui et al. 2021). Future challenges include leveraging microstate features to predict clinical outcomes and validating their utility directly in the inverse space. Overcoming these challenges will further enhance the applicability and validity of microstate analysis in understanding brain function and clinical conditions.