Dynamic gaze metrics in virtual reality: an AOI-based methodological framework applied to sexualized virtual characters

1 Introduction

Virtual reality (VR) offers immersive and ecologically valid environments that allow researchers to simulate complex and controlled scenarios while maintaining a sense of natural interaction (Parsons, 2015). These properties make it a powerful tool for studying human behavior in realistic and reproducible conditions. In psychology, VR can be used in therapeutic interventions, such as the treatment of phobias, where it facilitates controlled exposure scenarios (Botella et al., 2017; Freitas et al., 2021). Similarly, in sexology, VR is emerging as a research tool to understand issues such as erectile dysfunction (Huang et al., 2024), sexual aversion (Lafortune et al., 2020; Saint-Pierre-Côté et al., 2024), and even in forensic applications (Trottier et al., 2014). When combined with integrated eye-tracking technology, VR becomes even more relevant, providing direct access to users’ visual attention and perceptual-motor dynamics (Jin et al., 2024).

Traditionally, research on visual perception and attention has relied predominantly on fixation-based eye-tracking metrics (Duchowski, 2002; Holmqvist et al., 2011; Skaramagkas et al., 2021). These aggregate measures are useful for describing where participants look, how often they look, and how long they remain on specific areas. However, these metrics were primarily developed and validated in two-dimensional or screen-based contexts. In immersive VR, gaze behavior unfolds in a three-dimensional environment where visual exploration may involve both, eye and head movements. Head movements continuously change what is visible to the participant, making stationary gaze and fixation-based events more difficult to interpret as direct indicators of attention to fixed spatial locations. As a result, conventional fixation-based measures may not fully describe how gaze is organized in relation to virtual objects or how this organization changes over time.

In immersive settings, eye-tracking analyses are frequently based on the identification of Areas of Interest (AOIs), where gaze samples are classified in a binary manner (inside vs. outside). In studies of visual attention to sexual stimuli, commonly used metrics include the characteristics of the first fixation, the number of fixations on specific AOI, the time spent within them, and the fixation proportions or probabilities relative to predefined AOIs (O’Kane et al., 2023; Lykins et al., 2006; Milani et al., 2023; Rupp and Wallen, 2007). These measures provide important information about gaze allocation, but they remain largely based on discrete AOI event-based summaries. Several approaches have been proposed to better account for the dynamic and continuous nature of gaze behavior in VR (Hwang and Lee, 2025; David-John et al., 2021; Diaz et al., 2013). Among these approaches, gaze entropy offers a complementary way to summarize how gaze is distributed across space and transitions between areas (Shiferaw et al., 2019). It reflects how evenly attention is distributed and shifts across space, providing a global index of gaze exploration and diversity.

Building on previous work, the present study focuses on a set of less commonly used gaze metrics that complement aggregate fixation-based measures by characterizing how gaze is spatially organized relative to AOIs and how this organization varies over time. To this end, immersive VR was used to evaluate these metrics in a three-dimensional viewing context, where gaze behavior could be analyzed in relation to life-sized virtual body regions and natural head-eye exploration rather than only as gaze position on a flat image or video. The framework was applied to an existing dataset involving neutral and sexualized virtual characters to examine its ability to characterize gaze behavior across markedly different viewing conditions. This contrast was particularly useful for illustrating the interpretation of multiple gaze metrics, as prior screen-based studies have shown that visual attention to sexual stimuli varies across observer groups and body regions (Rupp and Wallen, 2007; O’Kane et al., 2023). Gaze behavior is first characterized using Gaze Radial Angular Deviation (GRAD), which measures the angular distance between the gaze direction and the center of a defined AOI. Unlike traditional AOI-based approaches, GRAD provides a continuous measure rather than a binary inside-outside classification (Duchowski et al., 2002; Renaud et al., 2009). From the GRAD signal, several descriptors are derived, including the mean, standard deviation (SD), and coefficient of variation (CV). In addition, the correlation dimension ( D 2 ) is computed directly from the GRAD time series to capture the temporal complexity of gaze behavior relative to the AOI. Higher D 2 values reflect more chaotic and unpredictable gaze paths, while lower values suggest more structured and targeted exploration (Renaud et al., 2009). These metrics are particularly relevant for examining individual differences in perceptual strategies.

The primary aim of this study is methodological. Specifically, it proposes a structured framework for jointly integrating and interpreting fixation-based, AOI-centered dynamic, and entropy-based measures in immersive VR eye-tracking analyses.

More specifically, this study aimed to.

• Develop a unified AOI-based analysis pipeline combining fixation-based, GRAD-based, and entropy-based metrics in immersive VR;
• Specify how these metric families can be interpreted in relation to attentional allocation, spatial organization relative to AOI centers, temporal complexity of gaze fluctuations, and transition structure across AOIs;
• Use sex-related gaze patterns as an empirical case study to illustrate the joint interpretation of fixation-based, GRAD-based, and entropy-based metrics in immersive VR.

2 Background and related work

2.1 Eye-tracking metrics in virtual reality

Standard eye-tracking metrics, such as fixation count, fixation duration, and saccade-related measures, provide valuable insight into attentional allocation (Moreno-Arjonilla et al., 2024). In virtual reality, their interpretation is challenged by head movements and the dynamic nature of virtual scenes. Despite these constraints, standard measures can still be meaningfully applied, for example by quantifying the time spent within areas of interest, even when gaze is not strictly stationary. Their value has also been demonstrated in predictive applications. For example (David-John et al., 2021), combined established gaze features, including fixation- and saccade-related measures analyzed over short temporal windows, to predict when users were about to interact in VR. However, not all conventional measures are equally sensitive to changes in cognitive state. For example (Hwang and Lee, 2025), found that fixation count and duration did not distinguish sustained-attention states from attentional lapses, whereas the time to the first fixation, the time to the first saccade, and saccade amplitude showed significant differences. This suggests that complementary metrics capturing the spatial and temporal organization of gaze may reveal information not captured by isolated fixation-based summaries alone. Entropy-based approaches offer a complementary perspective and have been applied in VR studies of Presence (Jordan and Slater, 2009), and more recently, to one of its subcomponents, Place Illusion (PI), defined as the sensation of being physically present in a virtual environment (Küçüktütüncü et al., 2025). Entropy provides a global summary of gaze dispersion (Stationary Gaze Entropy) and transition dynamics (Gaze Transition Entropy) across the visual field (Shiferaw et al., 2019). Most conventional implementations compute entropy using a static grid that segments the visual field into fixed spatial areas. While this grid-based method captures how gaze moves spatially within the scene, it does not fully account for the dynamic nature of VR environments, where objects and characters can move during the scenario. In this context, an alternative approach can be used in which entropy is computed relative to gaze-targeted AOIs (Krejtz et al., 2014; 2015). This AOI-based method remains effective in the presence of scene dynamics, as it relies on semantically meaningful and context-dependent areas rather than fixed spatial grids. These two approaches, however, capture different aspects of visual exploration. Grid-based entropy provides a topographic description of eye movements across the visual field, whereas AOI-based entropy characterizes how gaze shifts occur between meaningful targets, emphasizing functional or content-driven gaze dynamics. Although the AOI-based entropy approach allows consideration of moving AOIs, it still provides only a global summary of gaze dispersion and transitions across AOIs, without enabling comparisons of gaze behavior relative to specific areas.

To address this limitation, GRAD provides an AOI-centered metric that complements fixation-based and entropy-based measures. As previously described, GRAD measures the angular distance between the user’s line of sight and the center of an AOI and was introduced by (Renaud, 2006; 2011). A mathematically comparable calculation was later described by (Diaz et al., 2013), who computed the instantaneous angular distance between the gaze vector and a target point in a virtual environment. In their framework, this distance served primarily as a geometric indicator of gaze-to-target alignment and as a criterion for the classification of gaze behavior. Within the GRAD framework, this instantaneous angular-distance measure is further analyzed by extracting summary metrics that characterize different aspects of gaze behavior. The mean GRAD captures the central tendency of the angular deviation from the AOI center, while the standard deviation (SD) reflects the extent of spatial dispersion around that point. Finally, the coefficient of variation (CV = SD/Mean) provides a normalized index of variability, which can be interpreted as an indicator of gaze stability, relative to the center of the AOI. Together, these metrics help identify attentional patterns, such as focused or scattered gaze (Renaud et al., 2009). The correlation dimension D 2 quantifies the complexity of a system’s behavior by analyzing the geometric structure of its trajectory in phase space, a conceptual framework in which all possible states of a dynamic system are represented as points. As the system evolves, its trajectory forms a pattern known as an attractor, which reflects its natural tendencies over time (Wolf et al., 1985). For example, the trajectory of an oscillating pendulum forms a loop in phase space, indicating predictable and repetitive behavior, which corresponds to a low D 2 value. In contrast, the random motion of a butterfly’s flight produces a more complex and dispersed trajectory, resulting in a higher D 2 value. This measure helps distinguish between deterministic systems, which follow structured and predictable patterns, and random systems, which exhibit unstructured and unpredictable behavior. In the context of eye movements, D 2 provides insight into the evolving dynamics of gaze behavior, revealing whether visual exploration is structured or chaotic (Shelhamer and Azar, 1997; Shelhamer, 2005; Renaud et al., 2009).

Taken together, these approaches characterize complementary dimensions of gaze behavior. However, extracting and interpreting them jointly in immersive VR requires a structured methodological framework. The present study proposes such a framework by integrating fixation measures, GRAD-derived metrics, and entropy measures within an AOI-centered analysis of eye-tracking data.

2.2 Gendered visual patterns

Sex-related differences in visual attention provide a useful empirical context for illustrating the contribution of a multi-metric approach. Previous eye-tracking studies using two-dimensional sexual stimuli have shown that gaze allocation varies across body regions and observer groups. For example (Rupp and Wallen, 2007), reported that men spent more time viewing female faces, while studies using video stimuli have shown that women may allocate greater attention to contextual and background elements (O’Kane et al., 2023). These findings suggest that sex-related gaze patterns may involve differences in how visual attention is distributed between anatomically salient regions and contextual elements of the scene. This literature therefore provides a relevant empirical context for examining whether fixation-based, AOI-centered dynamic, and entropy-based metrics offer complementary characterizations of these differences.

3 Materials and methods

3.1 Participants

Participants were recruited through social media, email distribution lists, and word of mouth. The study included 44 participants: 22 cisgender heterosexual women (mean age = 23.27 years, SD = 3.14) and 22 cisgender heterosexual men. One male participant was excluded from the analysis due to continuous fixation on a single point outside the AOI throughout the entire VR exposure, thereby failing to attend to the stimuli as instructed. This reduced the number of male participants to 21 (mean age = 25.00 years, SD = 4.64). To maintain parsimony, only heterosexual and cisgender participants were included in order to minimize confounding variables while maximizing explained variance. We acknowledge the relevance of sexual orientation and gender identity, and recognize the simplification inherent in this approach. Throughout the manuscript, the terms “male or men” and “female on women” refer to biological sex.

3.2 Experimental protocol

Participants were exposed to two types of virtual characters. The first was a neutral, android-like virtual character (VC), and the second was a nude, human-like VC of the opposite sex. The sequence began with a habituation phase in which participants were placed in a virtual room without any VC, allowing them to adapt to the VR environment. They were then shown the neutral VC, followed by the sexualized VC, both displayed in similar bedroom environments. The neutral character was always presented first to avoid inducing any residual sexual arousal in this condition. Following each immersion, participants were asked to complete a questionnaire directly within the headset using the controller. The questionnaire included 14 items assessing subjective sexual arousal, sexual presence, perceived uncanniness, cybersickness, and attention throughout the scenario. Although the stimuli were presented in VR, participants were asked to minimize head movements due to the simultaneous collection of EEG data (BrainVision system, 32-channel EEG with gel-based electrodes). These EEG data are not included in the present analysis. The virtual reality session lasted approximately 45 min. Participants were exposed to four experimental conditions presented in randomized order. In some conditions, elements of the virtual environment and the VC were personalized by the participants. The present study focuses on one standardized condition in which the virtual environment and the VC were identical for all participants. To ensure balanced comparisons, an equal number of male and female participants who experienced this condition at the same position in the sequence were selected for analysis. The study was approved by the ethics committees of the authors’ institutions. Participants received financial compensation (CAD $70).

3.3 Virtual characters, immersion software, and equipment

The sexualized VCs used in this study were selected from the Genesis eight collection (Daz Studio) (female - Figure 1B; male - Figure 1C). The morphology and overall appearance of the male and female VCs were kept as similar as possible, although inevitable anatomical differences remained in facial structure, as well as in the breast/chest and genital areas (Figure 1A).

For the sexual scenario, the VCs were animated using motion capture data to depict sexually suggestive gestures. At the beginning of the sequence, participants were placed in the virtual environment without any virtual character present (17 s). The VC then entered the scene and walked through the virtual environment, initially moving with their back to the participant (35 s), then turning to face them (12 s). The VC then engaged in explicit sexual behavior (masturbation) while remaining at a constant distance of 4.5 m from the participant (30 s). Toward the end of the sequence, the VC gradually approached the participant until reaching a distance of 2.5 m (8 s) and remained at that distance while performing sexual movements until the end of the sequence (25 s). For the neutral VC, a gender-neutral “Y Bot” (Mixamo, Adobe) was used (Figure 1D) and animated using stock, non-sexual animations available on the platform. The trajectory and duration of the character’s movements were kept identical across conditions, with only the segments depicting sexual behaviors replaced by neutral animations. While the two scenarios differed slightly in posture and movement characteristics, they were not intended to be directly compared. Rather, each scenario was considered separately to examine whether visual exploration patterns differed between men and women within a given context. Each scenario lasted approximately 2 minutes. For analysis, a 30-s segment from the middle of each scenario was retained: the period depicting explicit sexual activity (masturbation) for the sexualized VC, and a moment when the VC was seated on a bench and performed gentle hand movements resting on its knees for the neutral VC. In both cases, the character was positioned at a fixed distance of 4.5 m from the participant. The AOIs were defined as circular regions centered on the face, upper-body/breast, and genital areas. At the 4.5 m viewing distance used in the analyzed segments, they extended 16.9 cm, 22.0 cm, and 19.7 cm from their respective centers, corresponding to AOI areas of approximately 897 cm 2, 1520 cm 2, and 1213 cm 2. These values refer to the AOIs used in the analyses, not to the anatomical surface area of the virtual character.

Virtual bedroom environments were developed in Unity (Version 2020.3.36) with consistent lighting, experimentally validated for uniformity across conditions. In the sexual condition, the HQ Bedroom (Next Level 3D, Unity Asset Store, Version 1.1) asset was used. In the neutral condition, a similar room layout was recreated within Unity, but the level of realism was intentionally reduced to minimize any emotional elicitation from environmental objects and to maintain a neutral atmosphere.

The experiment was conducted using an HTC Vive Pro Eye HMD equipped with integrated eye tracking running at 90 Hz, offering a spatial accuracy of approximately 0.5 ° to 1.1 ° (Schuetz and Fiehler, 2022). Although the Vive Pro Eye eye-tracker can operate at 120 Hz, pilot testing showed display instability (visual lag and rendering artifacts) at this frequency. The system was therefore set to 90 Hz to ensure stable visual rendering during the VR exposure. Eye-tracking data were collected using the SRanipal SDK, and the experiment was run on a computer equipped with an Nvidia GeForce RTX 3080 GPU, an Intel Core i7-10700K processor, and 32 GB of RAM.

3.4 Eye-tracking metrics

Eye-tracking data were processed to extract quantitative indicators describing participants’ visual behavior in response to the virtual stimuli. An overview of the preprocessing and metric-extraction pipeline is provided in Figure 2, and the corresponding MATLAB scripts are publicly available in a GitHub repository: https://github.com/sstpcote/MATLAB-VR_EyeTracking-Metrics. Prior to analysis, all data lines containing missing or invalid gaze samples were removed from the datasets.

3.4.1 Gaze radial angular deviation (GRAD)

Gaze deviation from AOIs was quantified using the Gaze Radial Angular Deviation (GRAD) metric (Figure 3) (Renaud et al., 2009). The computation was performed in Unity3D (2022.3.20, C#) using a custom playback script that reloaded synchronized datasets containing the participant’s head pose, gaze direction, and the 3D coordinates of the target AOIs. This playback script allowed the visualization of post-experimental gaze behavior relative to the displayed AOIs and the generation of a CSV file containing frame-by-frame angular deviations for quantitative analysis. The custom Unity playback project is publicly available in a GitHub repository: https://github.com/sstpcote/Unity-GRAD-Playback. For each frame, the gaze origin (center point between the participant’s eyes) provided by the SRanipal SDK was expressed in head-local coordinates. This local position was transformed into world space by applying the recorded head rotation, ensuring that the reconstructed eye position accurately reflected head movements within the immersive environment.

At each time point t, a gaze vector g t was defined, representing the participant’s gaze direction in 3D space, and a reference vector r t was defined, extending from the eye position toward the center of the target AOI, namely the head, chest and genital areas, whose placement are shown in Figure 1. The angular deviation between these two vectors was computed at each time point t using Equation 1.

G R A D t = arccos g t ⋅ r t ‖ g t ‖ × ‖ r t ‖ (1)

For the 30-s segment retained for analysis, the mean ( G R A D mean ), standard deviation (G R A D SD), and variation coefficient ( G R A D CV ) of the angular deviation were computed to characterize respectively, the spatial tendency, the absolute dispersion, and the relative variability of visual alignment with the target AOIs.

3.4.1.1 Correlation dimension ( D 2 )

The D 2 of the GRAD signal was computed using the Grassberger-Procaccia algorithm (Renaud et al., 2010; Benbouriche et al., 2016). This algorithm evaluates how data points in a phase space cluster across different spatial scales. The idea is to take an embedded time series x n and, for a given point x i, count how many other points x j lie within a distance ε. This process is repeated for all points and for various values of ε, where N is the total number of points in the time series, ‖ x ‖ is the Euclidean norm, and θ ( x ) is the Heaviside step function.

The correlation function C ( ε ) is defined as follows (Equation 2):

C ε = 2 N N − 1 ∑ i = 1 N ∑ j = i + 1 N θ ε − ‖ x i − x j ‖ (2)

The Heaviside function is defined as follows (Equation 3):

θ x = 0 if x < 0 1 if x ≥ 0

The double summation in the equation above counts only the pairs ( x i , x j ) whose distance is less than ε. For small ε and large N, the function C ( ε ) is expected to follow a power law. Thus, a plot of log ( C ( ε ) ) versus log ( ε ) should yield a line.

The slope of this line gives the D 2, defined as follows (Equation 4):

D 2 = d ⁡ log C ε d ⁡ log ε (3)

To verify that the observed D 2 reflects deterministic structure rather than random fluctuations, we compared the D 2 values obtained from the original gaze time series against those derived from surrogate data. Surrogates are synthetic time series that preserve selected linear properties of the original signal (e.g., amplitude distribution and power spectrum). This allows us to assess whether the observed complexity reflects structured gaze dynamics rather than random fluctuations (McSharry, 2011). A significant deviation of D 2 orig from the surrogate distribution supports the presence of structured, non-random gaze dynamics.

3.4.2 Fixations

Gaze direction vectors were first transformed into angular coordinates (in degrees) using trigonometric projection (Imaoka et al., 2020). Horizontal and vertical gaze angles were then calculated using Equations 4, 5:

θ x = arctan x z × 180 π (4)

θ y = arctan y z × 180 π (5)

where θ x and θ y represent the horizontal and vertical gaze angles, respectively (in degrees). The values x, y, and z correspond to the components of the normalized gaze direction vector (SRanipal SDK). A moving average filter with a window of three samples was then applied to the angular gaze signals to reduce high-frequency noise(Mejia-Romero et al., 2020). Fixations were identified using angular dispersion thresholds applied over time. Specifically, sequences of gaze samples were evaluated based on the minimum and maximum angular positions observed during a given segment. A segment was classified as a fixation when the angular dispersion, defined as the difference between the maximum and minimum angular values, remained below 1° on both horizontal and vertical axes for a continuous period of at least 250 ms. These thresholds were selected based on previous calibration work in immersive VR, which identified 1° of dispersion and 250 ms as the optimal parameter combination for fixation detection (Llanes-Jurado et al., 2020). To determine fixation location, each fixation was classified based on its angular position in 3D space, using the angular deviation of gaze relative to the center of the AOIs. Since, for the 30 s segment selected for analysis, the VCs were displayed at a fixed distance from the viewer, the thresholds used to assign fixations to a given AOI were empirically defined based on the angular size of each AOI, ranging from 2.15° to 2.8°. The total number of fixations on each area was then calculated.

3.4.3 Dwell time

In this paper, dwell time refers to the total duration for which gaze was located within an AOI. Following (Holmqvist et al., 2011), a dwell is defined as one visit within an area of interest, from entry to exit. Dwell time was therefore computed by summing all periods during which gaze remained within a given AOI, regardless of whether a fixation was occurring. In dynamic contexts such as video or virtual reality, this metric can be particularly informative, as it captures sustained visual attention without requiring the eye to remain stationary. Gaze samples were assigned to AOIs using the same GRAD-based localization procedure applied to fixations.

3.4.4 Gaze entropy metrics

Each gaze sample from the continuous eye-tracking signal was assigned to one of three semantically and anatomically defined AOIs: Face, Chest, and Genital or to an Out-of-AOI category when it fell outside all predefined AOIs. These AOIs move coherently with the VC and therefore provide a stable reference frame for quantifying gaze behavior under dynamic conditions relative to meaningful anatomical regions. Inspired by entropy-based analyses of gaze sequences proposed by (Krejtz et al., 2015), two complementary entropy metrics were computed from this AOI sequence to characterize (1) the spatial dispersion of gaze across AOIs, referred to as spatial gaze entropy (SGE), and (2) the complexity of transitions between them, referred to as gaze transition entropy (GTE).

3.4.4.1 Spatial gaze entropy

SGE quantifies how evenly visual attention was distributed across the four mutually exclusive gaze-location states over time: Face, Chest, Genital, and Out-of-AOI. Let P i denote the proportion of gaze samples classified into state i. Shannon entropy was computed as described by (Shannon, 1948), using Equation 6:

H SGE = − ∑ i = 1 4 P i ⁡ log 2 P i , (6)

and normalized to the maximum entropy of a four-state system, as shown in Equation 7:

H SGE norm = H SGE log 2 4 . (7)

Lower SGE values indicate that gaze is concentrated on a limited number of states, whereas higher SGE values reflect a more even distribution of gaze across states.

3.4.4.2 Gaze transition entropy

GTE captures the temporal complexity of gaze shifts between AOIs. Because consecutive gaze samples often remain within the same AOI during a fixation, the continuous AOI sequence was reduced by collapsing consecutive identical labels into a single state, such that only changes in gaze allocation between areas were retained. This step reduces the influence of repeated self-transitions, which are known to dominate transition-based metrics in high-frequency eye-tracking data (Duchowski et al., 2024).

P ( i , j ) denotes the transition probability from AOI i to AOI j in the compressed sequence, and π i denotes the empirical probability of initiating a transition from AOI i. Transition entropy was computed as shown in Equation 8:

H GTE = − ∑ i = 1 4 π i ∑ j = 1 4 P i , j log 2 P i , j , (8)

and normalized using the theoretical maximum for four states, as shown in Equation 9:

H GTE norm = H GTE log 2 4 . (9)

Higher GTE values indicate more complex and less predictable transitions between AOIs, while lower values reflect more structured scanning patterns.

3.5 Statistical analyses

A central methodological challenge in this study is that men and women viewed opposite-sex VCs. Although the 2 VCs were matched in overall geometry, unavoidable differences remained, particularly in the chest and genital areas (Figure 1). These geometric differences were not explicitly controlled for, as the present work does not aim to disentangle geometric from region-specific salience effects or to provide definitive psychological interpretations. Instead, such variability is considered acceptable, and even informative, within the scope of this study. Accordingly, the analyses focus on evaluating how different eye-tracking metrics respond to systematic variations in visual exploration. In this context, geometric differences between stimuli serve as a practical test case to assess the sensitivity of spatial and dynamic gaze metrics under realistic conditions.

Another source of variability arises from baseline differences in how observers distribute their gaze across body regions. Prior eye-tracking studies have reported sex-related differences in attentional allocation when viewing sexualized stimuli (Rupp and Wallen, 2007; Nummenmaa et al., 2012; O’Kane et al., 2023). In this context, these baseline differences between male and female particpants in which AOIs are preferentially explored can affect the underlying AOI-to-AOI transition structure and introduce noise into the GRAD-based dynamic metrics, even before evaluating the effects that these metrics are intended to capture. Because GRAD is computed continuously relative to each AOI’s center, any shift of gaze toward one area simultaneously alters the GRAD values of all other AOIs. As a result, baseline differences in which areas are more frequently inspected can propagate across all GRAD-based metrics, affecting not only the AOI being viewed but also the remaining ones. This means that group differences in basic AOI exploration can produce widespread differences in GRAD measures, even when the underlying gaze dynamics are otherwise similar.

To illustrate the complementary information provided by the different metric families, the analyses proceeded in three steps: (1) examining each AOI-specific metric separately to assess its incremental contribution beyond global scanpath organization; (2) evaluating Condition × Sex effects by jointly analyzing the set of raw gaze measures within each AOI, thus considering the metrics collectively; and (3) repeating the multivariate analyses after removing variance associated with global gaze distribution and AOI-to-AOI transitions, to determine whether the observed effects reflected AOI-specific gaze characteristics rather than broader scanpath organization. Analyses were conducted separately for each predefined AOI. AOI size remained constant within each model and therefore could not account for the observed effects of Sex or Condition. Area normalization was not applied, as the primary analyses did not involve direct statistical comparisons across regions.

3.5.1 Step 1- sensitivity analysis of gaze metrics

We first performed a hierarchical logistic regression limited to the sexualized condition, which maximizes between-group variability and tests each metric’s ability to discriminate between groups.

In Block 1 of the regression, SGE and GTE, both computed from the continuous sequence of AOI-to-AOI transitions without auto-transitions, were entered to account for global scanpath complexity. In Block 2, each AOI-specific metric (e.g., fixation count, dwell time, GRAD_mean, GRAD_SD, GRAD_CV, or GRAD D 2) was evaluated in a separate model.

This step identifies which gaze metrics contribute unique information beyond the global transition structure of the scanpath.

3.5.2 Step 2 - repeated-measures MANOVAs on gaze metrics

To examine how the gaze metrics collectively characterized differences across Condition and Sex, repeated-measures MANOVAs were conducted separately for each AOI: the face, chest (upper-body area), genital area, and out-of-AOI space. Analyzing AOIs separately reduced the number of dependent variables per model, thereby improving statistical stability given the sample size ( N = 43 ).

For the face, chest, and genital AOIs, each MANOVA included six dependent variables: fixation count, dwell time, GRAD_mean, GRAD_SD, GRAD_CV, and GRAD D 2. For the out-of-AOI area, only fixation count and dwell time were included, as GRAD metrics require a spatial reference anchored to a specific AOI and are therefore not defined outside the AOIs.

Wilks’ Lambda was used to test multivariate effects. Significant multivariate outcomes were followed by univariate ANOVAs and pairwise comparisons.

Distributional assumptions were assessed within each sex group and condition using Shapiro-Wilk tests. Most variables did not show significant departures from normality, including all face AOI metrics across conditions and all genital AOI metrics in the sexualized condition. Departures from normality were observed for some variables, particularly for the chest AOI and out-of-AOI measures across conditions, and to a lesser extent for genital AOI metrics in the neutral condition.

For variables showing departures from normality, aligned rank transform analyses were conducted as robustness checks for the corresponding univariate Condition × Sex effects (Wobbrock et al., 2011; Elkin et al., 2021). These analyses yielded a pattern of results consistent with the parametric follow-up tests. Therefore, the interpretation of the parametric analyses was retained.

3.5.3 Step 3 - repeated-measures MANOVAs on residualized gaze metrics

To determine whether the interaction effects of Condition and Sex persisted once variance associated with the global transition structure of the scanpath (SGE and GTE) was removed, we repeated the AOI-specific MANOVAs using residualized metrics. In the repeated-measures analyses comparing neutral and sexualized conditions, SGE and GTE were condition-specific and varied within participants. A standard repeated-measures MANCOVA was therefore not appropriate, as theses covariates could not be treated as between-subject adjustments (Miller and Chapman, 2001). In principle, a general multivariate mixed-effects model could accommodate covariates that vary by Condition (Bolker et al., 2009), but in our context, these models did not converge reliably given the present sample size and the number of dependent variables.

Given these constraints, each gaze metric was residualized by regressing it on SGE and GTE. The resulting residuals represent the portion of variance not explained by global scanpath structure and were used as adjusted inputs for the MANOVAs. This two-step procedure, removing the effect of a varying factor through multivariate regression before conducting group-level comparisons, is conceptually similar to established practices in geometric morphometrics, where variables of shape are regressed on covariates such as size and the adjusted values are subsequently submitted to multivariate analyses to test group differences (Zelditch et al., 2004; Boussali and Cavaré, 2026). Residual-based multivariate strategies of this form have also been applied in behavioral science (Miller and Chapman, 2001).

Importantly, the residualized metrics are not interpretable as gaze measures. Because residualization removes the portion of each metric that is predictable from SGE and GTE, the resulting values no longer correspond to meaningful angular deviations or to gaze behavior within a specific AOI. Instead, they represent statistically adjusted values whose sole purpose is to assess whether Condition × Sex effects observed in the previous step persist after controlling for global scanpath complexity.

3.5.4 Multiple-comparison control

Within each MANOVA, univariate tests were Bonferroni-corrected. Because four separate MANOVAs were conducted (one per AOI), an additional Bonferroni adjustment was applied, resulting in a final significance threshold of p < . 0125 for all univariate tests.

4 Results

4.1 Sensitivity analysis: discriminative value of AOI-based gaze metrics

Hierarchical logistic regressions were conducted to assess whether each gaze metric explained variance not accounted for by the global scanpath structure captured by SGE and GTE. Model improvement (Block 2 vs. Block 1) was assessed using the change in model fit ( Δ χ 2 ), the increase in explained variance ( Δ R Nag 2 ), the gain in classification performance ( Δ A U C ), and the unique effect size controlling for SGE and GTE (partial η 2). The partial η 2 values reported in Table 1 were derived from ANCOVA models conducted within the sexualized condition only. This differs from the repeated-measures analyses, where SGE and GTE varied by condition and were therefore handled through residualization. Improvements in these indices indicated the unique discriminative contribution of each metric. Table 1 summarizes those results.

TABLE 1. Hierarchical logistic regression sensitivity analysis and ANCOVA-derived effect sizes for gaze metrics discriminating sex groups within the sexualized condition.

Block 1 included SGE and GTE. Each Block 2 row represents a separate model in which one gaze metric was added to the Block 1 model. Thus, Block 2 rows should not be interpreted as a single multivariable model. * p < . 05, p < . 01, * p < . 001.

Overall, SGE and GTE alone (Block 1) yielded an AUC of 0.835, indicating good discrimination, although substantial additional information remained to be explained. Adding each gaze metric separately in Block 2 improved model performance for several AOIs.

For the face AOI, GRAD-based metrics showed the strongest additional contribution. GRAD_mean, GRAD_SD, and especially GRAD_CV produced significant improvements in model fit (Δ χ 2 = 11.08 to 18.85, all p ≤ 0.013). All three metrics increased AUC relative to SGE + GTE alone ( Δ A U C = 0.087 to 0.113 ), but only GRAD_CV produced a statistically significant AUC improvement ( p = 0.029 ). These effects were accompanied by medium-to-large partial effect sizes (η partial 2 = 0.209 to 0.305, all p ≤ 0.003). In contrast, fixation count, dwell time, and GRAD D 2 did not provide meaningful additional information for this AOI. When considered separately, the face-related metrics indicated that differences between men and women were mainly reflected in the spatial properties of gaze relative to the AOI center, rather than in the amount of time or number of fixations allocated to the face.

For the chest AOI, several gaze metrics produced large and highly significant improvements. Fixation count, dwell time, GRAD_mean, and GRAD_SD all produced strong increases in model fit (Δ χ 2 = 12.64 to 22.69, all p < 0.001), accompanied by substantial gains in explained variance ( Δ R Nag 2 = 0.223 to 0.359 ). Each of these metrics also increased AUC relative to SGE + GTE alone ( Δ A U C = 0.091 to 0.135 ), with fixation count, dwell time, and GRAD_SD showing statistically significant AUC improvements (p = 0.037, 0.016, and 0.016, respectively). Partial effect sizes were medium to large (η partial 2 = 0.288 to 0.395, all p < 0.001). GRAD D 2 also significantly improved model fit (Δ χ 2 = 6.01, p = 0.014) and showed a modest increase in explained variance ( Δ R Nag 2 = 0.114 ), together with a partial effect size (η partial 2 = 0.159, p = 0.010), but it did not produce a significant AUC improvement (Δ A U C = 0.065, p = 0.071). In contrast, GRAD_CV showed minimal incremental value, with small and non-significant changes across indices. This pattern suggested that, when metrics were evaluated individually, differences between men and women in the chest AOI were reflected in attentional allocation, spatial gaze organization, and, to a lesser extent, dynamic complexity.

For the genital AOI, GRAD_CV showed by far the strongest additional contribution. It produced the largest improvement in model fit (Δ χ 2 = 23.93, p < 0.001), the greatest increase in explained variance ( Δ R Nag 2 = 0.374 ), and the only statistically significant gain in AUC (Δ A U C = 0.135, p = 0.010). The corresponding partial effect size was large (η partial 2 = 0.423, p < 0.001). Fixation count, dwell time, GRAD_mean, and GRAD D 2 also produced significant but more moderate improvements in model fit (Δ χ 2 = 5.94 to 9.50, all p ≤ 0.015), with small-to-medium increases in explained variance ( Δ R Nag 2 = 0.113 to 0.174 ). However, none of these metrics yielded significant AUC improvements. Their partial effect sizes were in the small-to-medium range (η partial 2 = 0.102 to 0.200, all p ≤ 0.042). In contrast, GRAD_SD did not contribute meaningfully to model performance. Overall, at the individual-metric level, GRAD_CV was the dominant contributor for the genital AOI, suggesting that relative variability in gaze distance from the AOI center was more informative than attentional allocation, average gaze distance, or dynamic complexity.

Together, these sensitivity analyses identified which gaze dimensions carried information about differences between men and women when each metric was evaluated separately within each AOI. The next analyses examined how these metrics behaved jointly across conditions and sex groups using AOI-specific repeated-measures MANOVAs.

4.2 Multivariate effects of sex × condition on raw gaze metrics

The MANOVAs were conducted to assess whether the Sex × Condition interaction produced multivariate differences in the gaze metrics for each AOI (Table 2).

TABLE 2. Repeated-measures MANOVA results for sex differences in gaze metrics across neutral and sexualized conditions.

Values are presented as means (SE). * p < . 001, p < . 01, * p < . 0125 (Bonferroni-corrected significance threshold), † p < . 05.

Across all AOIs, the multivariate Sex effect was not significant in the neutral condition, with all p-values well above the Bonferroni-corrected threshold ( p = . 0125 ) and the nominal significance level ( p = . 05 ), indicating that men and women did not differ in their gaze patterns under neutral viewing conditions. In contrast, a significant multivariate Sex effect emerged consistently in the sexualized condition, indicating that sex-related differences in gaze behavior, in our context, became apparent only when participants viewed the sexualized VC. Accordingly, detailed AOI-specific MANOVA results are described only for the sexualized condition.

The multivariate test revealed a significant Sex effect for the face AOI (Wilks’ Λ = 0.361, F ( 6,36 ) = 10.625, p < 0.001, η 2 = 0.639). Significant Sex differences emerged for the same three GRAD-based metrics identified in the hierarchical regression analysis. GRAD_mean was higher in men ( p < . 001 ), indicating that their gaze remained, on average, farther from the center of the face AOI. GRAD_SD and GRAD_CV were higher in women (p < . 001 for both), reflecting greater variability and less stability in their distance from the AOI center. Fixation count, dwell time, and GRAD D 2 did not differ between groups.

For the chest AOI, the MANOVA revealed a significant multivariate Sex effect in the sexualized condition (Wilks’ Λ = 0.285, F ( 6,36 ) = 15.051, p < 0.001, η 2 = 0.715). Univariate tests showed significant Sex differences for all metrics except GRAD_CV. Men exhibited higher fixation counts ( p < . 001 ) and longer dwell times ( p < . 001 ), indicating greater visual attention to the chest. GRAD_mean was also higher in men ( p < . 001 ), meaning that their gaze tended to remain farther, on average, from the center of the chest AOI. In contrast, women showed higher GRAD_SD ( p < . 001 ), reflecting greater variability in their distance from the AOI center. GRAD D 2 was higher in men ( p < . 001 ), indicating more dynamically complex gaze trajectories around the center of the chest area. GRAD_CV did not differ between groups ( p = . 362 ). This pattern suggested that sex differences for the chest AOI involved multiple dimensions of gaze behavior, including visual engagement, spatial organization around the AOI center, and temporal complexity of gaze fluctuations.

For the genital AOI, Sex differences again emerged in the sexualized condition (Wilks’ Λ = 0.405, F ( 6,36 ) = 8.824, p < 0.001, η 2 = 0.595). Univariate tests showed that women had longer dwell times than men ( p = . 010 ). They also exhibited greater relative variability in their distance from the AOI center, as reflected by higher GRAD_CV values ( p < . 001 ), meaning that gaze distance from the genital AOI center fluctuated more strongly relative to the average distance from that center. Women also showed higher dynamical complexity of gaze fluctuations, indexed by GRAD D 2 ( p = . 011 ), indicating greater temporal complexity in gaze-distance fluctuations around the genital AOI. Other metrics did not reach the corrected significance threshold ( p < . 0125 ): GRAD_mean showed only a slight difference ( p = . 014 ), GRAD_SD did not differ between groups ( p = . 052 ), and fixation count showed no Sex effect ( p = . 246 ).

For the out-of-AOI area, the MANOVA revealed a significant Sex effect in the sexualized condition (Wilks’ Λ = 0.821, F ( 2,40 ) = 4.360, p = 0.019, η 2 = 0.179). Univariate tests showed that only dwell time reached the corrected significance threshold: women spent more time looking outside the defined anatomical AOIs than men ( p = . 009 ). This indicated that women distributed a larger portion of their viewing time toward regions not captured by the predefined anatomical AOIs. Fixation count showed no significant Sex difference after applying the Bonferroni-corrected threshold ( p = . 048 ).

4.3 Multivariate effects after controlling for global scanpath structure

To determine whether Condition and Sex interaction effects persisted once variance associated with global scanpath structure was removed, the same AOI-specific MANOVAs were repeated on residualized metrics (Table 3). These residuals represent the portion of each measure not explained by SGE and GTE.

TABLE 3. Repeated-measures MANOVA results for sex differences in residualized gaze metrics across neutral and sexualized conditions after controlling for global scanpath structure.

Values are presented as means (SE). * p < . 001, p < . 01, * p < . 0125 (Bonferroni-corrected significance threshold), † p < . 05.

For the face AOI, all three GRAD-based metrics that were significant in the raw data remained significant after residualization: GRAD_mean ( p = . 008 ), GRAD_SD ( p = . 015 ), and GRAD_CV ( p = . 003 ). These results indicated that Sex differences in gaze behavior relative to the face AOI persisted even after controlling for the global AOI-transition structure.

For the chest AOI, most of the Sex differences observed in the raw metrics persisted after residualization. Only GRAD D 2 ( p = . 035 ) did not reach the corrected significance threshold ( p < . 0125 ). This suggested that most chest-related Sex differences were not solely attributable to global scanpath organization, whereas the dynamic complexity indexed by GRAD D 2 overlapped more strongly with global transition structure.

For the genital AOI, only GRAD_CV remained significant after residualization ( p < . 001 ). All other metrics, including dwell time, GRAD_mean, and GRAD D 2, no longer reached the corrected significance threshold once global scanpath variance was removed. This pattern indicated that most Sex differences in gaze toward the genital area are explained by global scanpath structure, while GRAD_CV reflected an AOI-specific dynamic component that persists beyond these global effects.

For the out-of-AOI area, no residualized metric reached the corrected significance threshold. This indicated that the raw out-of-AOI differences were largely accounted for by global scanpath structure.

5 Discussion

Standard gaze measures primarily capture how often and how long an area is viewed, providing information related to visual engagement and interest. In contrast, AOI-centered dynamic metrics inform how gaze is organized and regulated relative to a spatial reference over time, an aspect that is particularly interesting in immersive virtual environments characterized by dynamic, three-dimensional, and interactive scenes. The present framework illustrates how AOI-centered dynamic measures can complement standard gaze metrics by describing spatial and temporal properties of gaze organization. To establish the contribution of individual metrics, variance related to global scanpath organization was first accounted for using a sensitivity analysis based on hierarchical regression. This step allowed the discriminative value of each gaze metric to be evaluated independently. Building on this foundation, AOI-specific repeated-measures MANOVAs examined how these metrics behaved jointly across AOIs. Finally, residualized MANOVAs were conducted after removing variance associated with SGE and GTE, allowing the robustness of the AOI-specific effects to be evaluated independently of global gaze distribution and transition structure. Together, these three analytical steps provided the basis for an integrative interpretation of gaze organization.

5.1 Disentangling AOI-centered gaze dynamics from global scanpath organization

A central challenge in the analysis of gaze as a dynamic, continuous signal arises when comparing groups whose visual exploration may be structured differently by socially and sexually salient features of the stimulus, as is the case when men and women view sexualized VCs of the opposite sex. In such cases, group differences may primarily emerge at the level of overall visual exploration rather than from how gaze is organized relative to specific AOIs, influencing the structure of the continuous gaze signal and, in turn, the AOI-centered metrics derived from it. Entropy-based measures of global scanpath organization (SGE, GTE), computed from gaze transitions between AOIs, provide a way to account for broader exploration patterns, whether driven by semantic, affective, or geometric properties of the stimulus. Importantly, this approach captures how such properties shape visual exploration rather than directly controlling for geometric differences, which are discussed in the limitations section. Although global scanpath organization alone provided good discriminative performance between groups, several metrics explained additional variance.

5.2 Differential sensitivity of gaze metrics across AOIs

The sensitivity analysis clarified how individual gaze metrics contributed information beyond global scanpath organization, revealing that their discriminative value varied across AOIs. Rather than reflecting a uniform pattern, different metrics emerged as most informative depending on the reference area.

For the face and genital areas, sex-related differences were primarily carried by relative variability measures (GRAD_CV), which provided the strongest individual contribution beyond global scanpath organization. The fact that GRAD_CV yielded the largest and only significant increases in AUC for these areas indicates that, when metrics are considered individually, sex differences are better characterized by how stably gaze is regulated around a reference distance from the AOI center than by how long the area is viewed or how widely gaze is dispersed.

In contrast, for the chest area, the largest individual gains in classification performance were observed for dwell time and GRAD_SD. Significant increases in AUC for these metrics indicated that information at this level was primarily carried by sustained visual engagement and absolute spatial dispersion relative to the area, rather than by relative variability. This pattern suggests that, when metrics are examined individually, discriminative information for this area is mainly distributed across attentional allocation and spatial organization of gaze.

Taken together, these findings showed that, when viewing sexualized virtual characters, sex-related discriminative information was not carried by a single gaze dimension across all anatomical regions. Instead, the most informative metric varied by region: relative variability around the AOI center was most informative for the face and genital areas, whereas dwell time and absolute spatial dispersion were more informative for the chest area. This pattern illustrates the methodological value of examining fixation-based and GRAD-based metrics jointly, as relying on a single metric family may obscure which dimension of gaze behavior carries the relevant information in each region.

Importantly, significant gains in AUC at the individual-metric level indicate where discriminative information is concentrated for each area, without implying a specific gaze strategy. This sensitivity analysis therefore identifies which dimensions of the gaze signal carry information prior to their joint interpretation in the subsequent analyses.

5.3 Sex differences as a case study: what the metrics reveal

In this section, sex-related differences are used as a case study to illustrate how the proposed set of gaze metrics jointly characterizes patterns of gaze organization in immersive virtual environments. As the sensitivity analyses indicated that different metrics carried discriminative information depending on the AOI examined, the discussion is organized by AOI. Before turning to area-specific profiles, we briefly contrast gaze organization across neutral and sexualized conditions.

5.3.1 Contextualizing the case study: neutral vs. sexualized conditions

The absence of sex differences in the neutral condition indicates that gaze organization was largely comparable across groups when the stimulus lacked sexual relevance. This pattern may first reflect the absence of sexual salience, a context in which sex-related differences in gaze behavior are typically attenuated. Similar context-dependent reductions in sex differences have been reported in prior eye-tracking studies using non-sexual or low-salience stimuli (Lykins et al., 2008; Nummenmaa et al., 2012). The fact that this condition-dependent pattern was also observed with GRAD-based and entropy-based metrics suggests that these measures are sensitive to changes in gaze organization, consistent with the patterns typically detected using standard eye-tracking measures. Because sex-related differences were primarily observed in the sexualized condition, the remainder of the discussion focuses exclusively on this condition.

5.3.2 Face AOI in sexualized condition

In the present VR setting, fixation count and dwell time did not reveal sex-related differences in facial attention. These results differed from previous eye-tracking studies using static sexual images, which reported greater facial viewing in men (Rupp and Wallen, 2007). Moreover, the lack of group differences in GRAD D 2 suggested that the observed sex-related differences were not associated with a more complex temporal organization of gaze trajectories and did not arise from changes in how gaze unfolded over time. Instead, sex-related differences emerged at the level of spatial organization. Women showed lower GRAD mean values, indicating that their gaze was, on average, positioned at a shorter distance from the facial reference point. This pattern is consistent with a general upward shift in gaze positioning within the scene. Given that women’s gaze was, on average, positioned closer to the facial reference point, shifts toward other parts of the body resulted in larger variations in gaze distance relative to that reference, as reflected in higher GRAD SD and CV values. Taken together, this combination suggests a gaze distribution that is shifted closer to the facial reference on average, without necessarily being on the face itself, while remaining spatially more dispersed. The persistence of these effects after residualization further supports the interpretation that they reflected AOI-centered spatial organization rather than differences driven by global scanpath structure.

5.3.3 Chest AOI in sexualized condition

For the chest AOI, clear sex-related differences emerged in fixation-based measures. Men showed higher fixation counts and longer dwell times than women, indicating greater visual engagement with this area. This pattern closely aligns with previous eye-tracking studies of sexualized stimuli, which have consistently reported increased male attention toward sexually salient upper-body areas based on fixation count and dwell time (Rupp and Wallen, 2007; Nummenmaa et al., 2012). The present results therefore replicated these findings within an immersive VR environment when using standard fixation-based metrics. However, men’s greater allocation of attention to the chest area did not translate into a more centrally focused gaze relative to the chest reference. In the present analysis, the chest AOI was defined using a single central reference point located between the two breasts, such that GRAD-based measures captured gaze distance relative to this geometric center rather than to individual anatomical features. Men’s larger mean angular deviation indicated that their gaze was organized at a greater average distance from this reference point, while lower absolute spatial dispersion (GRAD SD) suggested a more concentrated distribution around that offset distance. The absence of sex differences in relative variability (GRAD CV) indicated that variability normalized to the mean distance was comparable across groups. Together, these results describe a spatially offset but consistently organized gaze distribution relative to the chest reference. One possible explanation is that gaze alternated between the two breasts in men, although this interpretation remains speculative and other factors may also have contributed to this pattern. While higher GRAD D 2 values in the raw metrics initially suggested greater dynamic complexity of gaze trajectories in men, this effect did not persist after controlling for global scanpath organization. In contrast, effects related to attentional allocation and spatial gaze organization remained robust following residualization, indicating that they reflected AOI-centered gaze organization rather than differences driven by global exploration patterns.

5.3.4 Genital AOI in sexualized condition

For the genital AOI, women showed longer dwell time, a pattern that is consistent with previous eye-tracking studies reporting increased attention to the genital area among women (Rupp and Wallen, 2007; Lykins et al., 2008). However, this effect did not persist after controlling for global scanpath structure, suggesting that it may have reflected broader patterns of scene exploration rather than area-specific attentional engagement. Similarly, while GRAD D 2 initially showed a marginal group difference, this effect also disappeared after controlling for global scanpath organization. This suggests that the observed effects were unlikely to reflect increased temporal complexity or irregularity of gaze trajectories around the genital reference. Instead, the most robust difference that persisted after controlling for global scanpath organization was observed in GRAD CV, which was higher in women. The specificity of this effect suggests that sex-related differences in genital viewing were primarily expressed in the relative variability of gaze distance to the genital reference point, rather than in the amount of attention allocated, the average distance from the reference, or the absolute spatial dispersion of gaze. In other words, for comparable levels of visual attention and similar global exploration patterns, women showed greater variability in how gaze distance was regulated around the genital reference, a pattern that may partly reflect differences in the geometry of the genital regions displayed.

5.3.5 Out-of-AOI in sexualized condition

Out-of-AOI space does not have a defined area center and therefore does not allow the computation of AOI-centered metrics. Interpretation is thus restricted to fixation count and dwell time. In the sexualized condition, women showed longer dwell times than men prior to residualization. This observation is consistent with previous work reporting greater attention to background or contextual elements in women when using fixation-based measures (O’Kane et al., 2023). However, this difference did not persist after controlling for global scanpath organization. The disappearance of this effect after residualization suggests that out-of-AOI dwell time largely overlaps with broader patterns of scene exploration rather than reflecting area-independent differences.

5.3.6 Consistency of standard metrics and added value of dynamic measures

Overall, the present results show that standard fixation-based and time-based measures capture aspects of visual engagement in immersive virtual environments that are consistent with prior eye-tracking studies using two-dimensional images and videos. Measures such as fixation count and dwell time remain informative for characterizing how visual attention is allocated across AOIs.

At the same time, the inclusion of AOI-centered dynamic and entropy-based metrics provides additional insight into how gaze is organized and regulated over time relative to spatial references. These metrics reveal dimensions of gaze behavior that are not directly reflected in standard measures, including spatial organization and temporal variability, thereby providing a more detailed characterization of gaze behavior when gaze is treated as a continuous signal.

6 Limitations and future directions

An important boundary of the present study concerns the psychological interpretation of sex-related differences in gaze behavior. The sexualized VCs differed morphologically across sexes, making it difficult to attribute the observed differences in gaze organization unambiguously to underlying psychological processes. In particular, differences in anatomical structure, size, and visual detail between male and female characters may have influenced gaze behavior independently of sex-related viewing tendencies, especially for regions such as the chest and genital areas. Importantly, this constraint does not affect the primary contribution of the study. Rather than aiming to isolate psychological differences between men and women, the present work demonstrates that both standard and AOI-centered dynamic gaze metrics are sensitive to subtle, area-specific variations in visual exploration. From this perspective, variability driven by stimulus structure is not treated as noise, but as an informative component of gaze organization.

Future studies aiming to support stronger psychological inferences would benefit from experimental designs in which both men and women view and are attracted to characters of the same sex. Such designs would reduce confounds related to stimulus structure and allow a more direct assessment of sex-related differences in gaze strategies at a psychological level.

From a methodological perspective, it should also be noted that the AOIs differed in size and that no area normalization was applied. This does not affect the interpretation of Sex or Condition effects within a given AOI, because AOI size was constant within each AOI-specific analysis. However, descriptive differences across AOIs should not be interpreted as direct evidence that one region attracted more visual attention than another. A proper statistical comparison of gaze allocation across AOIs would require normalization by AOI area.

Moreover, some gaze metrics are conceptually related and may capture overlapping aspects of gaze behavior. In particular, GRAD D 2, which captures local dynamic unpredictability of gaze relative to an AOI, is conceptually related to global measures of scanpath entropy such as GTE. As a result, controlling for GTE may have attenuated the observable contribution of GRAD D 2 by removing shared variance linked to dynamic exploration. In experimental designs where global scanpath organization is comparable across groups or conditions, the contribution of GRAD D 2 may therefore become more pronounced.

Future work could also explore computing GRAD metrics exclusively during periods when gaze falls within a given AOI. Although this would reduce the continuous nature of the signal, it could improve psychological interpretability by ensuring that dynamic metrics reflect gaze behavior while actively inspecting the area, rather than including transitions outside the AOI. On the other hand, the continuous nature of the GRAD signal also opens additional methodological perspectives. In particular, temporal modeling approaches, such as recurrent neural networks or Long Short-Term Memory (LSTM) architectures, could be used to preserve the temporal structure of the signal and avoid information loss associated with summary statistics or averaged descriptors. While the present study focused on offline analyses, this property further suggests that GRAD-based metrics could be leveraged in adaptive immersive systems, where changes in gaze organization are monitored dynamically as the experience unfolds. Such perspectives align with recent conceptual work on cyberphysical and adaptive systems in forensic psychiatry, in which multimodal signals could be used to dynamically inform system behavior (Brideau-Duquette et al., 2025).

7 Conclusion

This study proposed and illustrated a methodological framework for analyzing gaze behavior in immersive virtual environments by integrating standard eye-tracking metrics, global scanpath indices, and AOI-centered dynamic measures. By distinguishing global scanpath organization from AOI-centered gaze dynamics, the framework helps clarify whether observed group differences reflect broader exploration patterns or gaze organization relative to specific spatial references. The comparison of men and women viewing sexualized virtual characters served as an empirical case study demonstrating both the complementary information provided by these different metric families and the value of their joint interpretation. Overall, this work support the use of multi-metric frameworks to study gaze behavior and provide a methodological basis for future work aiming to capture the complexity of visual exploration across contexts, including immersive environments.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. This text was translated and edited with the assistance of ChatGPT 5.2 (OpenAI), including improvements to clarity, structure, and flow. The authors reviewed and revised all AI-assisted content and take full responsibility for the accuracy and integrity of the final manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frvir.2026.1829181/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Mario Dalmaso, University of Padua, Italy

Reviewed by: Eugene Hwang, Yonsei University, Republic of Korea

Melina Boratto Urtado, University of São Paulo, Brazil