There exists several eye tracking technologies (Young and Sheena, 1975) that measure ocular features over time and transform them into a stream of gaze positions. These streams can be analyzed in different ways to identify periods of fixation and saccades. Salvucci and Goldberg (2000) describe five algorithms for such identification: Velocity Threshold Identification (I-VT), Hidden Markov Model Identification (I- HMM), Dispersion Threshold Identification (I-DT), Minimum Spanning Tree Identification (I-MST), and Area-of-Interest Identification (I-AOI). I-VT and I-HMM are velocity-based algorithms. In I-VT, consecutive points are identified as fixations or saccades, based on their point-to-point velocities (Findlay et al., 1995). I-HMM, on the other hand, uses a two-state Hidden Markov Model with hidden states representing the velocity distributions of saccade and fixation points. Compared to fixed-threshold methods like I-VT, I-HMM performs a more robust identification (Salvucci and Goldberg, 2000), since it employs a probabilistic model rather than a fixed velocity threshold, which allows more freedom in identifying points. I-DT and I-MST are dispersion-based algorithms that use a moving window to calculate the dispersion of points. Based on whether the dispersion is above or below the threshold, points are classified as fixations or saccades (Widdel, 1984). In I-MST, a minimum-spanning tree is constructed from gaze points. Edges with lengths exceeding a predefined ratio are labeled as saccades and clusters of points connected by saccades are labeled as fixations (Salvucci and Goldberg, 2000). I-AOI is an area-based algorithm that only identifies fixations within specified target areas (AOIs) (Salvucci and Goldberg, 2000). If a point falls within a target area, it is labeled as a fixation point, and if not, it is labeled as a saccade point. Consecutive fixation points are then grouped together, and groups that do not span a minimum duration are re-labeled as saccade points. A systematic evaluation of the performance of these algorithms are available at Komogortsev et al. (2010).

A tremor (or ocular micro-tremor, or physiological nystagmus) is an aperiodic, wavelike eye movement with a high-frequency ( ∼ 90 Hz) (Carpenter, 1988) and low-amplitude ( ∼ the diameter of a foveal cone) (Riggs et al., 1953). Due to this nature, tremors fall within the range of recording noise, making it challenging to record them accurately (Carpenter, 1988). Tremors allow the retaining of visual acuity during prolonged fixations (Riggs and Ratliff, 1951; Riggs et al., 1953). For instance, Riggs et al. (1953) showed that when tremors are bypassed artificially, the visual acuity diminishes over time.

A microsaccade (or flick, or flicker, or fixational saccade) is a small, fast, jerk–like eye movement that occurs during voluntary fixation (Ditchburn and Ginsborg, 1953; Martinez-Conde et al., 2004). They occur at a typical rate of 1–3 Hz, shifting the line of sight abruptly by a small amount (Ditchburn and Ginsborg, 1953). The average size of a microsaccade is about 6′ arc (i.e., the size of a thumb-tack head, held 2.5 m away from the eye) (Steinman et al., 1973). The dynamics of microsaccades vary with stimuli and viewing task. For instance, the difficulty of a task can be discerned by the number of microsaccades that occurred (Otero-Millan et al., 2008), and their magnitude (Krejtz et al., 2018). Microsaccades also have comparable spatio-temporal properties as saccades (Zuber et al., 1965; Otero-Millan et al., 2008). For instance, microsaccades lie on the saccadic main sequence (Zuber et al., 1965). The refractory periods between saccades and microsaccades are also equivalent (Otero-Millan et al., 2008). Moreover, microsaccades as small as 9′ generate a field potential over the occipital cortex and the mid-central scalp sites 100–140 ms after movement onset, which resembles the visual lambda response evoked by saccades (Dimigen et al., 2009). It is increasingly accepted that microsaccades play an important role in modulating attentional and perceptual processes (Hafed et al., 2015).

A drift (or slow drift) is a low-frequency eye movement that occurs during the intervals between microsaccades and saccades (Steinman et al., 1973). During a drift, the retinal image of a fixated object moves across photoreceptors (Ratliff and Riggs, 1950). Drifts have a compensatory role in maintaining accurate visual fixation; they occur either in the absence of microsaccades, or when the compensation by microsaccades is inadequate (Ratliff and Riggs, 1950). The average size of a drift is about a 6′ arc, with an average velocity of about a 1′ arc/sec (Ratliff and Riggs, 1950; Ditchburn and Ginsborg, 1953; Cornsweet, 1956; Nachmias, 1961).

Fixation count is the number of fixations identified within a given time period. Fixations can be counted either over the entire stimuli or within a single AOI (Holmqvist et al., 2011). Fixation count has been used to determine semantic importance (Buswell, 1935; Yarbus, 1967), the efficiency and difficulty in search (Goldberg and Kotval, 1999; Jacob and Karn, 2003), neurological dysfunctions (Brutten and Janssen, 1979; Coeckelbergh et al., 2002), and the impact of prior experience (Schoonahd et al., 1973; Megaw, 1979; Megaw and Richardson, 1979).

Fixation duration indicates a time period where the eyes stay still in one position (Salvucci and Goldberg, 2000). In general, fixation durations are around 200–300 ms long, and longer fixations indicate deeper cognitive processing (Rayner, 1978; Salthouse and Ellis, 1980). Also, fixations could last for several seconds (Young and Sheena, 1975; Karsh and Breitenbach, 1983) or be as short as 30–40 ms (Rayner, 1978; Rayner, 1979). Furthermore, the distribution of fixation duration is typically positively skewed, rather than Gaussian (Velichkovsky et al., 2000; Staub and Benatar, 2013). The average fixation duration is often used as a baseline to compare with fixation duration data at different levels (Salthouse and Ellis, 1980; Pavlović and Jensen, 2009). By comparing average fixation duration across AOIs, one could distinguish areas that were looked at for longer durations than others. In particular, if certain AOIs were looked at longer than others, then their average fixation duration would be higher (Goldberg and Kotval, 1999; Pavlović and Jensen, 2009).

The amplitude of a saccade (see Figure 2) is the distance travelled by a saccade during an eye movement. It is measured either by visual degrees (angular distance) or pixels, and can be approximated via the Euclidean distance between fixation points (Megaw and Richardson, 1979; Holmqvist et al., 2011). Saccade amplitudes are dependent on the nature of the visual task. For instance, in reading tasks, saccade amplitudes are limited to ≈ 2°, i.e., 7–8 letters in a standard font size (Rayner, 1978). They are further limited when oral reading is involved (Rayner et al., 2012). Furthermore, saccade amplitudes tend to decrease with increasing task difficulty (Zelinsky and Sheinberg, 1997; Phillips and Edelman, 2008), and with increasing cognitive load (May et al., 1990; Ceder, 1977).

The direction (or orientation, or trajectory) of a saccade or sequence of saccades is another useful descriptive measure. It can be represented as either an absolute value, a relative value, or a discretized value. Absolute saccade direction is calculated using the coordinates of consecutive fixations. Relative saccade direction is calculated using the difference of absolute saccade direction of two consecutive saccades. Discretized saccade directions are obtained by binning the absolute saccade direction into pre-defined angular segments (e.g., compass-like directions). In visual search studies, researchers have used different representations of saccade direction to analyze how visual conditions affect eye movement behavior. For instance, absolute saccade directions were used to analyze the effect of target predictability (Walker et al., 2006) and visual orientation (Foulsham et al., 2008) on eye movements. Similarly, discretized saccade directions were used to compare and contrast the visual search strategies followed by different subjects (Ponsoda et al., 1995; Gbadamosi and Zangemeister, 2001).

Saccade velocity is calculated by taking the first derivative of time series of gaze position data. Average saccadic velocity is the average of velocities over the duration of a saccade. Peak saccadic velocity is the highest velocity reached during a saccade (Holmqvist et al., 2011). For a particular amplitude, saccade velocity has been found to decrease with tiredness (Becker and Fuchs, 1969; McGregor and Stern, 1996), sleep deprivation (Russo et al., 2003), and conditions such as Alzheimer’s (Boxer et al., 2012) and AIDS (Castello et al., 1998). In contrast, saccade velocity has been found to increase with increasing task difficulty (Galley, 1993), increasing intrinsic value of visual information (Xu-Wilson et al., 2009), and increasing task experience (McGregor and Stern, 1996). Many studies on neurological and behavioral effects of drugs and alcohol (Lehtinen et al., 1979; Griffiths et al., 1984; Abel and Hertle, 1988) have used peak saccade velocity as an oculomotor measure.

Saccadic latency measures the duration between the onset of a stimulus and the initiation of the saccade (Andersson et al., 2010). In practice, the measurement of the saccadic latency is affected by two main factors: The sampling frequency of the setup, and the saccade detection time. The sampling frequency refers to the operational frequency of the eye tracker, where the operational frequency negatively correlates with the introduced error. The saccade detection time is when the device detects a saccade by arriving at the qualifying velocity or the criteria in the saccade detection algorithm. In a study with young and older participants, researchers observed the saccadic latencies to increase significantly in older participants with the decrease in the stimulus size Warren et al. (2013). Further, in comparative studies between healthy subjects and subjects with Parkinson’s disease with and without medication, saccadic latency shows potential as a biomarker for the disease (Michell et al., 2006). A similar study with participants having amblyopia has found the interocular difference between saccadic latencies to correlate with the difference in Snellen acuity (McKee et al., 2016).

Saccade rate (or saccade frequency) is the number of saccadic eye movements per unit time (Ohtani, 1971). For static stimuli, the saccade rate is similar to the fixation rate. For dynamically moving stimuli, however, the saccade rate is a measure of catch-up saccades generated during smooth pursuit (Holmqvist et al., 2011). The saccade rate decreases with increasing task difficulty (Nakayama et al., 2002) and fatigue level (Van Orden et al., 2000). Moreover, subjects with neurological disorders exhibit higher saccadic rates during smooth pursuit (O’Driscoll and Callahan, 2008; van Tricht et al., 2010).

Saccade gain (or saccade accuracy) is the ratio between the initial saccade amplitude and the target amplitude (i.e., Euclidean distance between the two stimuli among which that saccade occurred) (Coëffé and O’regan, 1987). This measure indicates how accurately a saccadic movement landed on the target stimuli (Ettinger et al., 2002; Holmqvist et al., 2011). When the gain of a particular saccade is greater than 1.0, that saccade is called an overshoot, or hypermetric, and when it is less than 1.0, that saccade is called an undershoot, or hypometric. Saccadic gain is probabilistic at a per-individual level; Lisi et al. (2019) demonstrates that biased saccadic gains are an individualized probabilistic control strategy that adapts to different environmental conditions. Saccade gain is commonly used in neurological studies (Ettinger et al., 2002; Holmqvist et al., 2011). For instance, Crevits et al. (2003) used saccade gain to quantify the effects of severe sleep deprivation.

Smooth pursuit direction (or smooth pursuit trajectory) indicates the direction of smooth pursuit movement as the eyes follow a moving stimulus (Holmqvist et al., 2011). The ability to pursue a moving object varies with its direction of motion. Rottach et al. (1996) showed that smooth pursuit gain is higher during horizontal pursuit than during vertical pursuit. This difference in gain can be attributed to most real-world objects naturally being in horizontal rather than vertical motion (Collewijn and Tamminga, 1984).

Smooth pursuit velocity is first moment of gaze positions during a smooth pursuit. Compared to saccade velocity, the velocity of smooth pursuits is low, typically around 20°/s–40°/s (Young, 1971). However, when participants are specifically trained to follow moving stimuli, or are provided with accelerating stimuli, higher peak smooth pursuit velocities were observed (Meyer et al., 1985). For example, Barmack (1970) observed peak smooth pursuit velocities of 100°/s in typical participants, when provided with accelerating stimuli. However, Bahill and Laritz. (1984) observed peak smooth pursuit velocities of 130°/s on trained baseball players. As the velocity of moving stimuli increases, the frequency of catch-up saccades increases to compensate for retinal offset (De Brouwer et al., 2002).

The second moment of the gaze position trace provides the acceleration of a smooth pursuit eye movement. This acceleration is maintained until the eye velocity (smooth pursuit velocity) matches the visual target’s velocity (Kao and Morrow, 1994). Examination of acceleration is typically a part of determining smooth pursuit onset (see Figure 3). Smooth pursuit acceleration has been used to analyze how visual cues (Ladda et al., 2007) and prior knowledge of a visual target’s trajectory (Kao and Morrow, 1994) impact eye movements. Smooth pursuit acceleration is higher when a visual target’s motion was unpredictable.

Smooth pursuit latency is the delay between when a target object starts to move (i.e., target onset) and when the pursuit begins (i.e., smooth pursuit onset) (Holmqvist et al., 2011). When the direction and velocity of the target object are not predictable, the smooth pursuit latency varies between 100–200 ms (Burke and Barnes, 2006). In paradigms such as step-ramp (Rashbass, 1961) that allow for anticipation, smooth pursuit latency may drop to 0 ms (or less, if pursuit starts before target motion) when its direction and velocity are predictable (Burke and Barnes, 2006; de Hemptinne et al., 2006). If the luminance of the moving object is the same as the background, the smooth pursuit latency may be prolonged by ∼ 50 ms (Braun et al., 2006). Smooth pursuit latency is also affected by distracting motion (Spering and Gegenfurtner, 2007); latency was increased when a distractor moved parallel to the pursuit direction but was decreased when the distractor moved opposite to the pursuit direction.

Fixations are maintained more accurately on stationary targets rather than moving targets. During smooth pursuit, both the eye and the target are in motion, and lag between their positions is expected (Dell’Osso et al., 1992). This error is known as retinal position error, and is formally defined as the difference between eye and target positions measured during fixations.

Smooth pursuit gain (or smooth pursuit accuracy) is the ratio between smooth pursuit velocity and the target velocity (Zackon and Sharpe, 1987; Holmqvist et al., 2011). Typically, the smooth pursuit gain is lower than 1.0 and tends to fall even lower when the target velocity is high (Zackon and Sharpe, 1987). Moreover, smooth pursuit gain is modulated by on-line gain control (Robinson, 1965; Churchland and Lisberger, 2002). Smooth pursuit gain is decreased during conditions that distract user attention (Březinová and Kendell, 1977). Furthermore, smooth pursuit gain is also higher when tracking horizontal motion, compared to tracking vertical motion Rottach et al. (1996). Smooth pursuit gain has also been used in neurologically research. For instance, O’Driscoll and Callahan (2008) analyzed the smooth pursuit gain of individuals diagnosed with schizophrenia, and observed a low smooth pursuit gain in these populations. The smooth pursuit gain can be quantified by the root mean squared distance between the target point and the gaze position (θ) over the span of the experiment of n data samples. θ R M S E = θ 1 2 + θ 2 2 + ⋯ ‥ + θ n 2 n (1)

Blink rate (or spontaneous blink rate, or blink frequency) is typically measured in blinks per minute. In some studies, the time between blinks (or blink interval) is measured instead (Shin et al., 2015). Early studies (Peterson and Allison, 1931; Newhall, 1932) show that blink rate is subjected to factors such as lighting, time of the day (fatigue), temperature, wind, age, and sex. Moreover, while blinks are predominantly involuntary, they are inhibited during engaged visual attention to minimize any blink-induced interruption to visual information (Ranti et al., 2020). More recent studies explore the idea of standard spontaneous eye blink rate in a broader aspect; healthy and non-healthy individuals under single (Doughty, 2002; Doughty and Naase, 2006; Ranti et al., 2020) and multiple (Doughty, 2001; Doughty, 2019) experiment conditions. Their results demonstrate the lack of a common value for blink rate, as they are dependent on experimental conditions. Thus, a standard measure for blink rate remains illusive despite the vast body of studies. Many studies on cognition adopt eye blink rate as a metric due its relative ease of detection, and its ability to indicate aspects of one’s internal state. In a study on the impact of blink rate on a Stroop task, researchers found that blink rate increases during a Stroop task compared with a baseline task of resting (Oh et al., 2012). Further, a study involving movie watching found the blink rate a reliable biomarker for assessing the concentration level (Shin et al., 2015; Maffei and Angrilli, 2019). Studies in dopamine processes also use blink rate as a marker of dopamine function. A review on studies in cognitive dopamine functions and the relationship with blink rate indicates varying results on blink rate in dopaminergic studies, including primates (Jongkees and Colzato, 2016).

The blink amplitude is the measure of the distance traveled by (the downward distance of upper eyelid) in the event of a blink (Stevenson et al., 1986). The amplitude measures the relative distance of motion to the distance the eyelid travels in a complete blink. This measure can be obtained using a video-based eye tracker (Riggs et al., 1987) or using electrooculography (Morris and Miller, 1996). Blink amplitude is often used in conjunction with research associated with fatigue measurement (Morris and Miller, 1996; Galley et al., 2004) and task difficulty (Cardona et al., 2011; Chu et al., 2014). Cardona et al. (2011) assessed the characteristics of blink behavior during visual tasks requiring prolonged periods of demanding activities. During the experiment, the authors noted a larger percentage of incomplete blinks. Another study involving fatigued pilots in a flight simulator showed that blink amplitude served as a promising predictor for level of fatigue (Morris and Miller, 1996).

Vector and string-based editing approaches have been developed to compute the similarity of scan paths (Jarodzka et al., 2010; De Bruin et al., 2013). In particular, De Bruin et al. (2013) introduced three metrics: 1) Saccade Length Index (SLI), 2) Saccade Deviation Index (SDI)—to assist in faster analysis of eye tracking data, and 3) Benchmark Deviation Vectors (BDV)—to highlight repetitive path deviation in eye tracking data. The SLI is the sum of the distance of all the saccades during the experiment. The SLI can be obtained through following equation, where s is the starting position of the saccade, e ending position of the saccade, and n is the total number of saccades. S L I T o t a l = ∑ i = 1 n s x − e x 2 + s y − e y 2 (2)

Jarodzka et al. (2010), on the other hand, proposed representing scan paths as geometrical vectors, and simplifying scan paths by clustering consecutive saccades directed at most Tϕ radians apart.

Time-to-First-Fixation on AOI (or time to first hit) refers to the time taken from stimulus onset up to the first fixation into a particular AOI (Holmqvist et al., 2011). This measure may be useful for both bottom-up stimulus-driven searches (e.g., a flashy company label) as well as top-down attention driven searches (e.g., when respondents actively decide to focus on certain elements or aspects on a website or picture). This metric is particularly useful for user interface evaluation, as a measure of visual search efficiency (Jacob and Karn, 2003). For instance, it has been used to evaluate the visual search efficiency of web pages (Ellis et al., 1998; Bojko, 2006). It is also influenced by prior knowledge, such as domain expertise. For instance, studies show that when analyzing medical images, expert radiologists exhibit a lower Time to First Fixation on AOIs (lesions, tumors, etc.) than novice radiologists (Krupinski, 1996; Venjakob et al., 2012; Donovan and Litchfield, 2013).

Revisit (or re-fixation, or recheck) count indicates how often the gaze was returned to a particular AOI. It can be used to distinguish between the AOIs that were frequently revisited, and the AOIs that were less so. A participant may be drawn back to a particular AOI for different reasons, such as its semantic importance (Guo et al., 2016), to refresh the memory (Meghanathan et al., 2019), and for confirmatory purposes (Mello-Thoms et al., 2005). The emotion perceived through visual stimuli also affect the likelihood of subsequent revisits, and thereby, the revisit count (Motoki et al., 2021); yet this perceived emotion is difficult to interpret purely through revisit count. Revisits are particularly common in social scenes, where observers look back and forth between interacting characters to assess their interaction (Birmingham et al., 2008). Overall, revisit count is indicative of user interest towards an AOI, and can be used to optimize user experiences.

Dwell time is the interval between one’s gaze entering an AOI and subsequently exiting it (Holmqvist et al., 2011). This includes the time spent on all fixations, saccades, and revisits during that visit (Tullis and Albert, 2013). For a typical English reading task, a lower dwell time (e.g., < 100 ms) may imply limited information processing, and a higher dwell time (e.g., > 500 ms) may imply otherwise (Tullis and Albert, 2013). Dwell times are dependent on factors such as the size and complexity of content within an AOI (Goldberg and Kotval, 1999), the level of interest towards an AOI (Fisher et al., 2017), situational awareness (Hauland and Duijm, 2002), and task difficulty (Goldberg and Kotval, 1999). In some cases, a higher dwell time could be associated with motivation and top-down attention, as goal-driven respondents may refrain from looking at contextually irrelevant stimuli (Mohanty and Sussman, 2013).

The gaze transition matrix is an adaptation of the transition matrix of Markov models into eye movement analysis. In a Markov model, a transition matrix (or probability matrix, or stochastic matrix, or substitution matrix) is a square matrix, where each entry represents the transition probability from one state to another. This concept was first applied for eye movement analysis by Ponsoda et al. (1995) to model the transition of saccade direction during visual search. Similarly, Bednarik et al. (2005) used this concept to model the transition of gaze position among AOIs and study the correlation between task performance and search pattern. The gaze transition matrix is calculated using the number of transitions from the X ^th AOI to the Y ^th AOI. Based on the gaze transition matrix, several measures have been introduced to quantify different aspects of visual search.

Goldberg and Kotval (1999) defined transition matrix density (i.e., fraction of non-zero entries in the transition matrix) in order to analyze the efficiency of visual search. Here, a lower transition matrix density indicates an efficient and directed search, whereas a dense transition matrix indicates a random search.

When comparing gaze transition matrices, Vandeberg et al. (2013) performed an element-wise comparison of transition probabilities by modelling eye movement transitions as a multi-level Hidden Markov Model. Similarly, Jayawardena et al. (2020) utilized gaze transition matrices to analyze the probabilities of transition of gaze between four AOIs. Figure 4A shows the AOIs used in this study. Figure 4B shows a sample scanpath of a participant, and Figure 4C shows its corresponding gaze transition matrix. Here, each cell in the matrix represents the probability of gaze transition from one AOI to another.

Gaze transition entropy is a measure of predictability in AOI transitions and overall distribution of eye movements over stimuli (Krejtz et al., 2014; Krejtz et al., 2015). Compared to transition matrix density, gaze transition entropy is a histogram-based estimation, which, in effect, takes into account the AOI size. The concept of entropy used here is that of information theory; it describes the amount of information required to generate a particular sequence, as a measure of its uncertainty. Krejtz et al. (2014), Krejtz et al. (2015) computes gaze transition entropy by first modelling eye movements between AOIs as a first-order Markov chain, and then obtaining its Shannon’s entropy (Shannon, 1948). They obtain two forms of entropy: 1) transition entropy H _t (calculated for individual subjects’ transition matrices) and 2) stationary entropy H _s (calculated for individual subjects’ stationary distributions). H t = − ∑ i π i ∑ j p i j ⁡ log 2 p i j H s = − ∑ i π i ⁡ log 2 π i (3)

H _t indicates the predictability of gaze transitions; a high H _t (∀ _ij |p _ij → 0.5) implies low predictability, whereas a low H _t (∀ _ij |p _ij → {0, 1}) implies high predictability. It is calculated by normalizing the transition matrix row-wise, replacing zero-sum rows by the uniform transition probability 1/s (Here, s is the number of AOIs), and obtaining its Shannon’s entropy. H _s , on the other hand, indicates the distribution of visual attention (Krejtz et al., 2014); a high H _s indicates that visual attention is equally distributed across all AOIs, whereas a low H _s indicates that visual attention is directed towards certain AOIs. It is estimated via eigen-analysis, and the Markov chain is assumed to be in a steady state where the transition probabilities converge. In one study, Krejtz et al. (2015) showed that participants who viewed artwork with a reportedly high curiosity, yielded a significantly less H _t (i.e., more predictable transitions) than others. Moreover, participants who viewed artwork with a reportedly high appreciation, yielded a significantly less H _s (i.e., more directed visual attention) than others. In another study, Krejtz et al. (2015) showed that participants who reportedly recognized a given artwork, yielded a significantly higher H _t and H _s (i.e., less predictable gaze) than others. Jayawardena et al. (2020) performed entropy-based eye movement analysis on neurotypical and ADHD-diagnosed subjects during an audiovisual speech-in-noise task. They found that ADHD-diagnosed participants made unpredictable gaze transitions (i.e., high entropy) at different levels of task difficulty, whereas the neurotypical group of participants made gaze transitions from any AOI to the mouth region (i.e., low entropy) regardless of task difficulty. These findings suggest that H _t and H _s are potential indicators of curiosity, interest, picture familiarity, and task difficulty.

In eye-tracking neuroscience, researchers have jointly analyzed neuronal activity and oculomotor activity to study the physiological organization of the vision system, and their effect on cognition and behavior. Studies of neuronal activity during phenomena such as attention (Blair et al., 2009; Kimble et al., 2010), scene perception (Duc et al., 2008), inattentional blindness (Simons and Chabris, 1999), visual engagement (Catherine and James, 2001), and covert attention processing (i.e., visually fixating on one location while diverting attention to another) (Posner et al., 1980) had revealed important facts about cognition and behavior.

For instance, studies on covert attention processing show that attention cannot be inferred solely from whether an object was looked at (Hafed and Clark, 2002; Ebitz and Moore, 2019), and doing so would lead to false positives (Posner et al., 1980). Similarly, studies on scene perception show that information is processed in two forms: top-down (based on semantic importance) and bottom-up (based on visual significance such as color, brightness, etc.) (Duc et al., 2008). Moreover, studies on visual attention show that attention is divided among AOIs through a sustained cognitive state, from which relevant visual objects become available to influence behavior Duncan et al. (1994).

Studies have also revealed that pupillary activity is correlated with cognitive load (Hess and Polt, 1964; Hyönä et al., 1995), and also with neural gain (Eldar et al., 2013) and cortical activity (Reimer et al., 2016). While cognitive load can be inferred from pupillary activity, studies show that such inference becomes challenging in fast-paced cognitive tasks (Wierda et al., 2012), temporally overlapping cognitive tasks (Wierda et al., 2012), and in surroundings with varying ambient luminance (Zénon, 2017). Cherng et al. (2020) showed that pupil diameter is regulated by sympathetic activation (arousal-induced pupil dilation) and parasympathetic inhibition (saccade-induced pupil dilation), both of which are affected by ambient luminance. Wierda et al. (2012) showed that a deconvolved pupil response signal is indicative of cognitive load, with a high temporal resolution. However, (Zénon, 2017), showed that this method does not account for low-frequency pupil fluctuations and inter-individual variability of pupil responses, and instead proposed using auto-regressive models (Cho et al., 2018) with exogenous inputs to analyze pupillary activity. Similarly, Watson and Yellott (2012) proposes a generalized formula to analyze pupillary activity, which accounts for ambient luminance, the size of the adapting field, the age of the observer, and whether both pupils are adapted.

Researchers of eye tracking neurscience have utilized various measures to study neurodevelopmental disorders. Some of these studies are exclusively based on fixation measures. For instance, He et al. (2019) examined the fixation count and fixation duration in preschoolers with Autism Spectrum Disorder (ASD), and found that preschoolers with ASD had atypical gaze patterns in a facial emotion expression task compared to typical individuals. Moreover, they found that deficits in recognizing emotions from facial expressions in ASD correlated with the social interaction and development quotient. In contrast, some studies are based exclusively on saccadic measures. For instance, MacAskill et al. (2002) found that the saccade amplitude was modified during adaption of memory-guided saccades and this adaptive ability was impaired in individuals with Parkinson’s disease. Similarly, Barbosa et al. (2019) identfied that saccadic direction errors are associated with impulsive compulsive behaviors of individuals with Parkinson’s Disease, as they had difficulty in suppressing automatic saccades to a given target. Patel et al. (2012) showed that saccade latency is correlated with the severity of Huntington Disease (HD), and suggested its possiblity of being a biomarker of disease severity in HD. In another study, Jensen et al. (2019) found that reduced saccade velocity is a key indicator of progressive supranuclear palsy and other disoders of mid-brain. Similarly, Biscaldi et al. (1998) found that individuals with Dyslexia had significantly higher regressive saccade rate in a sequential-target task. Similarly, Termsarasab et al. (2015) stated that abnormalities in saccade gain could aid diagnosis of hyperkinetic and hypokinetic movement disorders. Similarly, Fukushima et al. (2013) used a memory-based smooth pursuit task to examine working memory of smooth pursuit direction in individuals with PD.

Certain studies performed their analysis using visual search measures. Rutherford and Towns (2008), for instance, demonstrated scan path similarities and differences during emotion perception between typical individuals and individuals with ASD. Similarly, Bours et al. (2018) demonstrated that individuals with ASD had increased time to first fixation on the eyes of fearful faces during emotion recognition task. Duret et al. (1999) showed that refixation strategies in macular disorders was dependant on location of the target relative to the scotoma, spatial characteristics of the disease and the duration of the disorder. Guillon et al. (2015) showed that gaze transition matrices could potentially reveal new strategies of visual scanning followed by individuals with ASD. Moreover, Wainstein et al. (2017) showed that pupil diameter could be a biomarker in ADHD based on the results from a visuo-spatial working memory task.

In HCI research, eye tracking has been used to evaluate the usability of human-computer interfaces (both hardware and software). Here, the primary eye movement measures being analyzed are saccades, fixations, smooth pursuits, compensatory, vergence, micro-saccades, and nystagmus (Goldberg and Wichansky, 2003). For instance, Farbos et al. (2000) and Dobson (1977) had attempted to correlate eye tracking measures with software usability metrics such as time taken, completion rate, and other global metrics. Similarly, Du and MacDonald (2014) had analyzed how visual saliency is affected by icon size. In both scenarios, eye movements were recorded while users navigated human-computer interfaces, and were subsequently analyzed using fixation and scan-path measures.

Pupil diameter measures are widely used to assess the cognitive load of users when interacting with human-computer interfaces. For instance, Bailey and Iqbal (2008) used pupil diameter measures to demonstrate that cognitive load varies while completing a goal-directed task. Adamczyk and Bailey (2004) used pupil diameter measures to demonstrate that the disruption caused by user interface interruptions (e.g., notifications) is more pronounced during periods of high cognitive load. Iqbal et al. (2004) used pupil diameter measures to show that difficult tasks demand longer processing time, induces higher subjective ratings of cognitive load, and reliably evokes greater pupillary response at salient subtasks. In addition to pupil diameter measures, studies such as Chen and Epps (2014) have also utilized blink rate measures to analyze how cognitive load (Sweller, 2011) and perceptual load (Macdonald and Lavie, 2011) varies across different tasks. They claim that pupil diameter indicates cognitive load well for tasks with low perceptual load, and that blink rate indicates perceptual load better than cognitive load.

In psychology research, eye tracking has been used to understand eye movements and visual cognition during naturalistic interactions such as reading, driving, and speaking. Rayner (2012), for instance, have analyzed eye movements during English reading, and observed a mean saccade duration and amplitude of 200–250 ms and 7–9 letters, respectively. They have also observed different eye movement patterns when reading silently vs aloud, and correlations of text complexity to both fixation duration (+) and saccade length (−). Three paradigms were commonly used to analyze eye movements during reading tasks: moving window—selecting a few characters before/after the fixated word (McConkie and Rayner, 1975), foveal mask—masking a region around the fixated word (Bertera and Rayner, 2000), and boundary—creating pre-defined boundaries (i.e., AOIs) to classify fixations (Rayner, 1975). Over time, these paradigms have been extended to tasks such as scene perception. Recarte and Nunes (2000) and Stapel et al. (2020), for instance, have used the boundary method to analyze eye movements during driving tasks and thereby assess the effect of driver awareness on gaze behavior. In particular, Recarte and Nunes (2000) have observed that distracted drivers had higher fixation durations, higher pupil dilations, and lower fixation counts in the mirror/speedometer AOIs compared to control subjects.

In psycholinguistic research, pupil measures have been used to analyze the cognitive load of participants in tasks such as simultaneous interpretation (Russell, 2005), speech shadowing (Marslen-Wilson, 1985), and lexical translation. For instance, Seeber and Kerzel (2012) measured the pupil diameter during simultaneous interpretation tasks, and observed a larger average pupil diameter (indicating a higher cognitive load) when translating between verb-final and verb-initial languages, compared to translating between languages of the same type. Moreover, Hyönä et al. (1995) measured the pupil diameter during simultaneous interpretation, speech shadowing, and lexical translation tasks. They observed a larger average pupil diameter (indicating a higher cognitive load) during simultaneous interpretation than speech shadowing, and also momentary variations in pupil diameter (corresponding to spikes in cognitive load) during lexical translation.

Moreover, studies in marketing and behavioral finance, have used eye-tracking and pupillometric measures to understand the relationship between presented information and the decision-making process. For instance, Rubaltelli et al. (2016) used pupil diameter to understand investor decision-making, while Ceravolo et al. (2019), Hüsser and Wirth (2014) used dwell time. Further, studies in marketing have identified relationships between the consumer decision process through fixations on different sections in the product description Ares et al. (2013), Menon et al. (2016). The utility of eye-tracking in marketing extends beyond product descriptions, to advertisements, brands, choice, and search patterns (Wedel and Pieters, 2008).