Anxiety disorders are among the most prevalent psychiatric conditions, typically following a chronic course and causing marked impairments in quality of life and social functioning. Reference Stein and Sareen1,Reference Xiang, Martinez, Chow, Carter, Negriff and Velasquez2 In the most recent nationwide epidemiological survey of mental disorders in China, anxiety disorders emerged as the most common diagnostic category, with a 12-month prevalence of 5.0% and a lifetime prevalence of 7.6%. Reference Huang, Wang, Wang, Liu, Yu and Yan3 Despite their significant prevalence, the neurobiological, genetic and behavioural mechanisms underlying the onset and chronic course of anxiety disorders remain incompletely mapped, reflecting substantial gaps in current pathophysiological models. Reference Ohi, Fujikane, Takai, Kuramitsu, Muto and Sugiyama4 These gaps have fuelled a growing search for objective neurocognitive markers that could elucidate the mechanisms of anxiety and inform precision diagnostics and interventions.
Increasing evidence indicates that abnormalities in cognitive and attentional processing are central features of anxiety disorders, and eye-tracking paradigms have emerged as promising tools to objectively quantify such disturbances. Reference Armstrong and Olatunji5 A key phenomenon is the attentional bias towards threat-related stimuli, often conceptualised as a hypervigilance–avoidance pattern. Reference Mogg and Bradley6 Recent meta-analytic findings from eye-tracking studies further substantiate this phenomenon, demonstrating that individuals with anxiety disorders orient rapidly to threatening stimuli and maintain prolonged gaze on these during free-viewing tasks. Reference Clauss, Gorday and Bardeen7
In addition, free-viewing tasks have also been applied in other psychiatric populations using emotionally neutral or general visual stimuli. Patients with schizophrenia typically show restricted scanning patterns, characterised by shortened scan paths and prolonged fixation durations, Reference Benson, Beedie, Shephard, Giegling, Rujescu and St Clair8 and similar abnormalities have been observed in depression, where visual exploration is likewise constrained and less efficient. Reference Zhang, Liu, Xu, Li, Xu and Xia9 Collectively, these findings highlight the broader applicability of free-viewing paradigms in delineating disorder-specific visual exploration strategies across psychiatric conditions.
In addition to free-viewing paradigms, investigations of other fundamental oculomotor paradigms in anxiety disorders remain scarce. Smooth-pursuit eye movements, which index the ability to continuously track moving targets, have been extensively investigated in schizophrenia and mood disorders, where abnormalities such as reduced pursuit gain and increased saccadic intrusions have been consistently reported. Reference Kathmann, Hochrein, Uwer and Bondy10,Reference Wolf, Ueda and Hirano11 Early work in non-clinical samples nonetheless reported that a higher state of anxiety was associated with enhanced predictive pursuit gain in young military recruits, Reference Kattoulas, Evdokimidis, Stefanis, Avramopoulos, Stefanis and Smyrnis12 suggesting that pursuit dynamics may be sensitive to anxiety-related arousal. Similarly, fixation stability tasks, which assess the capacity to maintain one’s gaze on a stationary stimulus, have been applied in schizophrenia and depression, where patients frequently exhibit fixation instability, reflected in shorter fixation durations and impaired saccadic control. Reference Zhang, Liu, Xu, Li, Xu and Xia9,Reference Liu, Li, Xu, Zhang, Cui and Wei13 However, the application of these paradigms to anxiety disorders has yet to be systematically investigated. Collectively, they may offer complementary perspectives on the oculomotor signatures of anxiety beyond threat-related scanning.
Therefore, the present study aimed to systematically delineate the eye movement profiles of patients with anxiety disorders across multiple paradigms and to evaluate their potential diagnostic utility. To enhance clinical specificity, a depression group was included as a psychiatric comparison, thereby facilitating a clearer differentiation of anxiety-specific abnormalities from those shared with affective disorders. We hypothesised that, compared with patients with depression and healthy controls, patients with anxiety would manifest hyper-scanning (e.g. increased scan paths), hyper-pursuit (e.g. higher gain) and increased fixation instability. These markers were expected to correlate with symptom severity and serve as potential diagnostic biomarkers for anxiety disorders.
Method
Ethical standards
The authors assert that all procedures contributing to this work comply with the ethical standards of the relevant national and institutional committees on human experimentation, and with the Helsinki Declaration of 1975 as revised in 2013. All procedures involving human subjects were approved by the Research Ethics Committee of Shanghai Medical Health Centre (Ethics approval no. 2020-100; approval date 25 December 2020).
Participants
A total of 91 patients with anxiety disorders and 118 with depressive disorders were consecutively recruited from the out-patient clinic at Shanghai Mental Health Centre, with 98 healthy controls recruited from the local community via advertisements. Written informed consent was obtained from each participant or their legal guardian prior to participation.
The inclusion criteria for the anxiety group were out-patients aged 16–45 years with >6 years of formal education and who were diagnosed with an anxiety disorder according to DSM-5 criteria (generalised anxiety disorder, panic disorder, social anxiety disorder or specific phobia), as confirmed by trained psychiatrists through structured clinical interviews. Ultimately, the anxiety group comprised 91 patients, including 69 with generalised anxiety disorder (75.82%), 12 with panic disorder (13.19%) and 10 with social anxiety disorder (10.99%).
The inclusion criteria for the depression group were identical, except that participants had been diagnosed with a depressive disorder according to DSM-5 criteria (primarily major depressive disorder or depressive episode), as confirmed by trained psychiatrists through structured clinical interviews.
The inclusion criteria for the healthy controls group were as follows: age 16–45 years, >6 years of formal education, no personal history of psychiatric disorders as confirmed by clinical interview and no first-degree relatives with anxiety, depression or other major psychiatric disorders.
Exclusion criteria for all groups were (a) comorbid psychotic disorders, bipolar disorder or other major psychiatric illnesses; (b) history of neurological disease, brain injury or severe medical conditions; (c) substance or alcohol dependence within the past 6 months; and (d) contraindications for eye-tracking assessment (e.g. uncorrected severe visual impairment).
Eye-tracking paradigm
Fixation stability
Participants were instructed to maintain their gaze on a central black dot (0.5° visual angle) while ignoring peripheral interference signals (‘*’ marks, 0.5°). During the initial 5 s, only the central target was presented (simple task). Between 5 and 10 s, distractors appeared randomly to the left or right of the central dot at varying eccentricities (distracted task). Each participant completed ten trials. A total of nine oculomotor indices were derived: fixation count and duration; saccade count, amplitude and duration; blink count; and microsaccade count, amplitude and duration.
Free viewing
Participants were instructed to relax and freely view 35 still monochrome images, each presented for 10 s. The stimuli included four categories of content: patterned images, natural scenes, social scenes and static objects, and were presented in a randomised order. Thirteen indicators were extracted for analysis, including fixation measures (count, duration); saccade measures (count, amplitude, duration, mean velocity, peak velocity); scan path length; blink count; pupil dynamics (mean pupil size, dynamic range, pupil size ratio); and entropy.
In addition, indices were extracted separately for each image category. For social scenes and static-object scenes, region-of-interest (ROI) analyses were conducted, with ROIs defined as both faces in social scenes and the objects themselves (excluding background regions) in object scenes. The following ROI-based measures were derived: pupil size; dwell time (absolute and percentage); fixation count (absolute and percentage); first fixation duration; ROI run count; and entropy. Further details on the definition of ROIs are shown in Supplementary Tables 1 and 2 available at https://doi.org/10.1192/bjp.2026.10626.
Smooth pursuit
Participants were instructed to visually follow moving black dots on the screen. The target moved for 2.5 s each horizontally and vertically in a sinusoidal trajectory (horizontal-/vertical-pursuit paradigm), or followed combined horizontal and vertical Lissajous trajectories at different frequencies (slow paradigm, 20 s; fast paradigm, 10 s). Each trial was repeated once. Five pursuit-related indices were extracted: horizontal, vertical and combined velocity gain, as well as catch-up saccade count and intrusive saccade count. ROI definitions are detailed in Supplementary Tables 1 and 2.
Eye-tracking devices
Eye movements were recorded using an EyeLink 1000 desktop eye tracker (SR Research, Ottawa, Ontario, Canada) in pupil–corneal reflection-tracking mode. Stimuli were presented on a 17-inch monitor with a resolution of 1024 × 768 pixels. Participants were seated in an adjustable chair positioned 90 cm from the display, with head stabilisation provided by a chin-and-jaw rest to align the eyes with the screen centre. The system tracked eye movements within ±30° horizontally and vertically, sampling from the dominant eye at 500 Hz. Eye movement data were exported using EyeLink Data Viewer (version 4.4.1 for Windows 11; SR Research Ltd, Ottawa, Ontario, Canada; https://www.sr-research.com).
Data analysis
All statistical analyses were performed using R (version 4.4.3 for Windows 11; R Foundation for Statistical Computing, Vienna, Austria; https://www.r-project.org). Prior to analysis, raw eye movement data were preprocessed to remove artefacts. Blink-related segments, fixations shorter than 20 ms, fixations outside the screen boundaries and saccades extending beyond the display were excluded. Reference Morita, Miura, Fujimoto, Yamamori, Yasuda and Iwase14 Univariate extreme outliers were identified using a distribution-based criterion (values exceeding ±3 s.d. from the mean), with additional verification by visual inspection of histograms and boxplots. To mitigate their influence on parametric analyses, extreme values were Winsorised by replacing them with the nearest valid upper or lower observation, affecting approximately 1.8% of all observations. Reference Tabachnick, Fidell and Ullman15
Demographic characteristics were compared using one-way analysis of variance for continuous variables and χ 2 tests for categorical variables. Eye movement indices were compared with analysis of covariance (ANCOVA), controlling for gender and education year, followed by post hoc comparisons with false discovery rate (FDR) correction. In addition, for the free-viewing paradigm, group (anxiety, depression and healthy controls) × image type (patterned images, natural scenes, social scenes and static objects) effects were examined using a linear mixed-effects model with subject as a random intercept, controlling for gender and education. Moreover, principal component analysis (PCA) was applied to eye movement variables that differed significantly between groups. Components were retained if they had eigenvalues >1 and together explained >70% of the variance, with variables loading ≥0.50 considered meaningful contributors. Group differences in principal component scores were examined using ANCOVA, controlling for gender and education, and associations with clinical symptoms were evaluated using Spearman correlations with FDR correction.
Machine learning analyses were conducted using principal component scores as predictors. For each pairwise group comparison (depression versus control, anxiety versus control, anxiety versus depression), three supervised algorithms were implemented: support vector machine with a radial basis function kernel, random forest and extreme gradient boosting. Models were trained on 70% of the data and tested on the remaining 30%, with feature selection based on univariate information gain performed within the training set, and hyperparameters optimised via 5-fold cross-validation using randomised search.
Results
Sample characteristics
The demographic and clinical characteristics of the study groups are presented in Table 1. Age did not differ significantly across the 3 groups (F = 1.43, P = 0.24). By contrast, gender distribution varied significantly among the groups (χ 2 = 8.01, P = 0.018), as did years of education (F = 4.00, P = 0.019). FDR-corrected post hoc analyses further revealed that years of education were significantly higher in the depression group than in the healthy controls group (P = 0.02).
Group differences in demographic, clinical and eye movement characteristics
HC, healthy controls; HAM-A, Hamilton Anxiety Rating Scale; HAM-D, Hamilton Depression Rating Scale; ROI, region of interest.
Continuous variables are presented as mean (s.d.). Eye movement group differences were assessed using analysis of covariance, controlling for gender and education (d.f. 2, 302), with effect sizes reported as partial η2. Only significant eye movement variables are shown, and non-significant results are provided in supplementary materials.
Statistical significance is denoted as *P < 0.05, **P < 0.01, ***P < 0.001.
Eye movement characteristics
The heat map of variables illustrating inter-group differences across the three tasks is presented in Fig. 1(a) and Table 1. Detailed group comparisons of eye movement measures are provided in Supplementary Tables 4, 5, 7 and 9.
Group differences in eye movement variables. (a) Heatmap showing significant group differences in eye movement variables across paradigms. Colours indicate the direction and strength of group differences (red, higher; blue, lower in the group first listed). (b) Estimated marginal means of fixation duration (ms) during free viewing across groups. HC, healthy controls; FS (S), fixation stability (simple); FV, free viewing; ROI_FV (SS), region of interest (ROI) in free viewing (social scenes); ROI_FV (SO), ROI in free viewing (static objects); SP (V), smooth pursuit (vertical); SP (H), smooth pursuit (horizontal); SP (SL), smooth pursuit (slow Lissajous); SP (FL), smooth pursuit (fast Lissajous); FDR, false discovery rate.
Fixation stability
In the simple paradigm, the number of saccades differed across groups (F (2,302) = 5.53, P = 0.004, ηp 2 = 0.035), with post hoc analyses showing that the anxiety (P = 0.02) and depression (P = 0.04) groups exhibited more saccades than healthy controls. For the distracted paradigm, four measures, including saccade count, amplitude, duration and microsaccade count, showed significant group effects (F (2,302) range 3.30–3.47, all P < 0.05, ηp 2 range 0.021–0.022). However, none of the post hoc pairwise comparisons survived FDR correction.
Free viewing
For the full-image paradigm, significant group differences were observed across the 6 eye movement indices, including fixation count (F (2,302) = 6.42, P = 0.002, ηp 2 = 0.041), saccade count (F (2,302) = 7.94, P < 0.001, ηp 2 = 0.050), scan path length (F (2,302) = 5.28, P = 0.006, ηp 2 = 0.034), blink count (F (2,302) = 5.01, P = 0.007, ηp 2 = 0.032), pupil size ratio (F (2,302) = 4.67, P = 0.010, ηp 2 = 0.030) and pupil size (F (2,302) = 5.13, P = 0.006, ηp 2 = 0.033). Post hoc pairwise analyses indicated that the anxiety group exhibited more fixations and saccades, as well as longer scan paths, than both the depression and healthy controls groups (all P < 0.05). In addition, the anxiety group showed fewer blinks and a lower pupil size ratio than healthy controls (both P < 0.05), whereas the depression group exhibited larger pupil size compared with healthy controls (P = 0.02).
The linear mixed-effects model (group × image type) analysis showed that only the interaction effect of fixation duration reached statistical significance (Supplementary Table 6). Specifically, the interactions between social scenes (with patterned images as the reference) × anxiety-healthy controls and static objects (with patterned images as the reference) × anxiety-healthy controls were significant (β = 11.28, s.e. = 5.43, P = 0.04; β = −16.89, s.e. = 5.43, P = 0.002). FDR-corrected post hoc simple effects indicated that, on patterned images, the anxiety group had shorter fixation durations than the depression group; on natural scenes, the anxiety group had shorter durations than both the depression and healthy controls groups (all P < 0.05) (Fig. 1(b)).
ROI-based analyses revealed significant group effects in social scenes for pupil size (F (2,302) = 4.99, P = 0.007, ηp 2 = 0.032), fixation count (F (2,302) = 3.34, P = 0.037, ηp 2 = 0.022) and run count (F (2,302) = 4.41, P = 0.013, ηp 2 = 0.028). Post hoc analyses showed that the depression group exhibited larger pupil size than healthy controls, whereas the anxiety group showed more runs than healthy controls (P < 0.05). For static objects, group differences were observed for pupil size (F (2,302) = 4.83, P = 0.009, ηp 2 = 0.031) and fixation count (F (2,302) = 6.06, P = 0.003, ηp 2 = 0.039). Specifically, the depression group exhibited larger pupil size than healthy controls whereas the anxiety group showed more fixations relative to both the depression and healthy controls groups (P < 0.05) (Supplementary Table 8).
Smooth pursuit
For vertical movements, vertical velocity gain differed across groups (F (2,302) = 8.83, P < 0.001, ηp 2 = 0.055), with both the anxiety and depression groups exhibiting higher values than healthy controls (P < 0.01). For horizontal movements, significant group effects were observed for horizontal velocity gain (F (2,302) = 5.05, P = 0.007, ηp 2 = 0.032) and intrusive saccade count (F (2,302) = 4.24, P = 0.015, ηp 2 = 0.027). Specifically, the anxiety and depression groups showed higher velocity gain than healthy controls whereas the anxiety group exhibited fewer intrusive saccades relative to healthy controls (P < 0.05).
For slow Lissajous movement, significant main effects of group emerged for four indices: horizontal velocity gain (F (2,302) = 18.72, P < 0.001, ηp 2 = 0.11), vertical velocity gain (F (2,302) = 15.92, P < 0.001, ηp 2 = 0.095), combined gain (F (2,302) = 6.75, P = 0.001, ηp 2 = 0.043) and intrusive saccade count (F (2,302) = 6.23, P = 0.002, ηp 2 = 0.04). Post hoc tests showed higher velocity gains in anxiety and depression groups than in healthy controls, and fewer intrusive saccades in the anxiety group than in healthy controls (P < 0.05). For fast Lissajous movement, both horizontal and vertical velocity gains differed by group (horizontal, F (2,302) = 7.04, P = 0.001, ηp 2 = 0.045; vertical, F (2,302) = 6.34, P = 0.002, ηp 2 = 0.04). Post hoc tests indicated higher gains in the anxiety group than in healthy controls for both indices, and higher horizontal gain in the depression group than in healthy controls (P < 0.05) (Supplementary Table 9).
PCA
Principal component identification
PCA was performed on 50 eye movement variables, showing significant group differences. Based on the eigenvalue criterion (eigenvalue >1) and cumulative variance explained (>70%), 10 components were initially retained. Examination of the loading structure indicated that 6 components were defined by high-loading variables (|loading| > 0.50), allowing for meaningful interpretation. These components were labelled as active visual exploration (principal component 1 (PC1)), pupillary arousal (PC2), smooth-pursuit control (PC3), pupillary variability (PC4), blink activity (PC5) and microsaccade instability (PC6). Components 7–10 did not yield interpretable loadings and were excluded from subsequent analyses. Supplementary Fig. 1 presents the scree plot, and Supplementary Tables 10 and 11 summarise the loading matrix and component interpretations.
Group differences in principal components
Group differences were observed across four principal components (Supplementary Table 12), including PC1 (F (2,302) = 6.84, P = 0.001, ηp 2 = 0.043), PC2 (F (2,302) = 9.51, P < 0.001, ηp 2 = 0.059), PC3 (F (2,302) = 10.04, P < 0.001, ηp 2 = 0.062) and PC5 (F (2,302) = 3.41, P = 0.034, ηp 2 = 0.022). Post hoc analyses indicated that the anxiety group showed higher scores than healthy controls across PC1–3 (all P < 0.05), and exceeded the depression group on PC1 (P = 0.003). The depression group also scored higher than healthy controls on PC2 and 3 (both P < 0.01). For PC5, despite a significant overall effect, no pairwise comparisons survived FDR correction (Fig. 2(a)).
Principal component (PC) differences and associations with clinical symptoms. (a) Group differences in PC1–3 scores. (b) Associations between principal components and symptom severity (Hamilton Anxiety Rating Scale (HAM-A) in the anxiety group, Hamilton Depression Rating Scale (HAM-D) in the depression group). Lines represent Spearman correlations. HC, healthy controls. *P < 0.05, **P < 0.01, ***P < 0.001.
Correlational analyses of principal components
In the anxiety group, Hamilton Anxiety Rating Scale (HAM-A) scores were negatively associated with PC4 (r = –0.331, P = 0.008) and positively associated with PC5 (r = 0.298, P = 0.012) and PC6 (r = 0.245, P = 0.038), whereas a marginal association was noted with PC2 (r = 0.223, P = 0.05). In the depression group, Hamilton Depression Rating Scale (HAM-D) scores were negatively correlated with PC1 (r = –0.286, P = 0.010) and PC4 (r = –0.234, P = 0.032), whereas other components did not reach significance (Fig. 2(b) and Supplementary Table 13).
Machine learning
Machine learning models based on the six principal component scores were applied to pairwise binary classifications among the three groups. As shown in Fig. 3, random forest models yielded the highest test area under the receiver operating characteristic curve (AUC) for anxiety versus healthy controls (AUC = 0.82, accuracy 0.74, sensitivity 0.77, specificity 0.71), and for depression versus healthy controls (AUC = 0.83, accuracy 0.65, sensitivity 0.90, specificity 0.44). For depression versus anxiety, the extreme gradient-boosting model performed best (AUC = 0.61, accuracy 0.62, sensitivity 0.42, specificity 0.74). Detailed classifier results are available in Supplementary Table 14.
Receiver operating characteristic (ROC) curves of optimal classifiers across pairwise groups. ROC curves for the best-performing classifiers in pairwise group comparisons. AUC, area under the ROC curve; HC, healthy controls.
Discussion
The present study provides novel evidence that anxiety disorders are characterised by distinctive eye movement alterations that extend beyond the widely described threat-related attentional biases. Specifically, compared with both depression and healthy controls groups, individuals with anxiety exhibited a hyper-scanning pattern during free viewing marked by more saccades and enlarged scan path length, as well as a hyper-pursuit pattern during smooth pursuit, reflected in greater pursuit gain, reduced intrusive saccades and increased catch-up saccades. Importantly, these alterations were closely linked to clinical symptom severity and contributed to disorder-level differentiation. Collectively, these findings indicate that aberrant eye movement control represents a distinctive signature of anxiety-related dysfunction and may constitute a promising candidate behavioural marker in clinical psychiatry.
The present findings demonstrate that the anxiety group exhibited hyper-scanning behaviour, characterised by longer scan paths and more frequent saccades, even during neutral viewing. This pattern suggests heightened environmental monitoring rather than optimised information sampling. In line with previous findings, anxiety disorders are characterised by robust attentional biases towards threat, involving both preferential orienting and impaired disengagement. Reference Pergamin-Hight, Naim, Bakermans-Kranenburg, van and Bar-Haim16 Correspondingly, eye-tracking evidence reveals that individuals with anxiety exhibit fragmented scanning and unstable fixation patterns in response to threatening stimuli, manifesting a core hypervigilant attentional style. Reference Lisk, Vaswani, Linetzky, Bar-Haim and Lau17
This neutral-context hyper-scanning aligns with Attentional Control Theory, suggesting that anxiety-related impairments in processing efficiency lead to compensatory but inefficient attentional deployment. Reference Eysenck, Derakshan, Santos and Calvo18 At the neural level, this manifests as a shift from top-down, goal-directed control – mediated by prefrontal executive systems – towards bottom-up, stimulus-driven processing. Reference Bishop19 Electrophysiological evidence confirms that anxiety is associated with enhanced early visual responses, suggesting a hyper-responsive perceptual system that generalises beyond explicit threat to fuel exaggerated monitoring of neutral stimuli. Reference Gupta, Kujawa and Vago20 Furthermore, neuroimaging studies demonstrate disrupted medial prefrontal–amygdala connectivity, consistent with impaired top-down regulation of limbic salience processing. Reference Kim, Gee, Loucks, Davis and Whalen21 Together, these mechanisms facilitate a generalised expectancy of threat Reference Aslanidou, Andreatta, Wong and Wieser22 in which hyper-scanning persists due to deficient suppression of stimulus-driven responses, suggesting that attentional control processes may represent a key mechanism linking aberrant visual exploration to the underlying pathophysiology of anxiety.
PCA reveals the multidimensional structure underlying eye movement patterns. PC1 (active visual exploration) captured a characteristic hyper-scanning pattern and exhibited robust group-level differences, suggesting that fragmented and inefficient scanning may serve as a stable, trait-like signature. In contrast, PC2, reflecting tonic-dominated pupillary activity, was interpreted as a putative indirect index of sustained locus coeruleus–noradrenergic engagement, with elevations in anxiety indicating heightened baseline autonomic arousal. Reference Murphy, O’Connell, O’Sullivan, Robertson and Balsters23 Consistent with this interpretation, PC2 scores correlated positively with HAM-A severity, in line with evidence that pupil-linked locus coeruleus–noradrenergic engagement tracks stress-related arousal in anxiety. Reference Grueschow, Stenz, Thörn, Ehlert, Breckwoldt and Brodmann Maeder24 Furthermore, additional components, including pupillary variability (PC4), blink activity (PC5) and microsaccade instability (PC6), were correlated with HAM-A scores. These associations align with established findings that pupil dynamics reflect sustained autonomic arousal, Reference Martín-Rodríguez, Martín-Sánchez, del Pozo Vegas, Lopez-Izquierdo, Martín-Conty and Silva Alvarado25 blink-related activity is linked to dopaminergic modulation and vigilance states, Reference Maffei and Angrilli26 and that microsaccade instability indexes impaired attentional control under affective load. Reference Krejtz, Żurawska, Duchowski and Wichary27 Collectively, these components appear to capture state-related physiological dynamics that track symptom severity, suggesting potential utility for monitoring clinical fluctuations. This integration of relatively stable cognitive features with dynamic state indicators supports a dual-dimensional perspective for the objective characterisation of anxiety.
Linear mixed-effects models revealed shorter fixation durations in anxiety patients across patterned and natural stimuli, indicating rapid yet inefficient visual sampling consistent with avoidance-related processing biases. Reference Shechner, Jarcho, Wong, Leibenluft, Pine and Nelson28 ROI-based analyses revealed pronounced effects in social scenes, where anxiety patients showed increased ROI entry frequency consistent with hypervigilant monitoring of socially salient cues. Reference Hauffe, Rauschenbach, Fassot, Schmitz and Tuschen-Caffier29 Although such effects may partially overlap with mechanisms implicated in social anxiety, the presence of similarly increased fixation frequency in non-social object scenes indicates that this redundant scanning pattern reflects a broader, transdiagnostic feature of anxiety in which social contexts amplify but do not uniquely drive this bias. By contrast, depression was primarily characterised by larger pupil responses across conditions, reflecting heightened autonomic responsivity rather than altered exploration strategies. Reference Brendler, Schneider, Elbau, Sun, Nantawisarakul and Pöhlchen30 Together, these results suggest that the anxious phenotype is marked by unstable and fragmented visual exploration, distinguishing it from the more physiologically driven arousal profile characteristic of depression.
Smooth-pursuit analysis revealed that, compared with controls, the anxiety group exhibited increased velocity gain, more catch-up saccades and fewer intrusive saccades, a profile suggestive of over-compensatory pursuit control during continuous tracking. A similar but attenuated pattern was also observed in the depression group, suggesting partially shared mechanisms across anxiety and depression. Pursuit dynamics have been shown to rely on predictive processing and attentional control, Reference Fukushima, Fukushima, Warabi and Barnes31 and the exaggerated gain in anxiety may reflect excessive monitoring and anticipatory adjustment, consistent with hypervigilant predictive processing. At the neural level, smooth pursuit is thought to rely on cortical–cerebellar loops, with the frontal eye fields and cerebellum playing central roles in adaptive gain control and predictive tracking. Reference Lisberger32 The elevation of PC3 (smooth-pursuit control) in both anxiety and depression may therefore reflect heightened but inefficient engagement of pursuit control mechanisms, with the alterations appearing more pronounced in anxiety. This pattern suggests that these smooth-pursuit abnormalities may be transdiagnostic, yet anxiety shows stronger over-compensatory control, consistent with its hypervigilant and fragmented exploration style.
The PCA-derived components provided a low-dimensional representation of eye movement features, demonstrating that these indices carry sufficient signal to differentiate patients from healthy controls with clinically relevant discriminative performance. Specifically, anxiety and depression could be reliably distinguished from healthy controls (AUC > 0.8), supporting the view that eye movement alterations constitute disorder-relevant biomarkers. Reference Zhang, Liu, Xu, Li, Xu and Xia9,Reference Gao, Xin, Wang, Gao, Wang and Yu33 By contrast, discrimination between anxiety and depression was modest (AUC = 0.61), probably reflecting shared general distress and clinical comorbidity, and consistent with transdiagnostic perspectives emphasising overlapping alterations in arousal regulation and attentional control. Reference Richter, Stahi, Mirovsky, Hel-Or and Okon-Singer34 Nonetheless, the contrasting eye movement profiles observed across the two conditions (e.g. hyper-scanning in anxiety versus reduced exploration in depression) suggest that eye-tracking may capture distinct underlying pathological dimensions. This supports the utility of eye movement measures for refining dimension-specific characterisation within comorbid affective presentations.
These findings highlight eye-tracking as a scalable, mechanism-informed tool for anxiety assessment. By bridging stable eye movement traits with symptom-sensitive physiological dynamics, this approach provides an objective framework to complement traditional clinical characterisation. The ongoing shift toward portable, low-cost technology Reference Morita, Miura, Toyomaki, Makinodan, Ohi and Hashimoto35 further enhances the feasibility of integrating these measures into routine clinical or community-based settings. Such accessibility supports the development of objective, data-driven approaches to mental health assessment in real-world environments.
Several limitations should be noted. First, gender and years of education were not matched across groups and, although included as covariates, residual confounding cannot be fully excluded. Second, HAM-A and HAM-D were not administered to healthy controls, limiting the assessment of subclinical symptom variation. Third, the imbalanced distribution of anxiety subtypes constrained subtype-specific analyses. Finally, given the substantial comorbidity between anxiety and depression, findings should be interpreted with consideration of overlapping mechanisms.
In conclusion, anxiety disorders exhibited distinct eye movement profiles, characterised by hyper-scanning and altered pursuit dynamics, reflecting altered attentional and sensorimotor processing. These findings position eye tracking as a scalable, low-burden modality capable of underpinning clinically actionable biomarker frameworks for psychiatric stratification and longitudinal monitoring.
Supplementary material
The supplementary material is available online at https://doi.org/10.1192/bjp.2026.10626
Data availability
The data that support the findings of this study are available from the corresponding author upon reasonable request. Some data may not be made available because of privacy or ethical restrictions.
Acknowledgement
The authors thank all participants for their time and valuable contribution to this study.
Author contributions
D.Z. and Y.L. contributed equally to this work and share first authorship. D.Z., Y.L. and T.Z. conceptualised and designed the study. D.Z., L.X., Y.X., X.L., W.Z., Y.H., J.Z., Y.W. and H.C. were responsible for data collection and preprocessing. D.Z. and Y.L. performed data analysis and drafted the manuscript. H.L. provided methodological support and supervision. T.Z. and J.W. secured funding, provided critical revisions and supervised all stages of the research. All authors reviewed and approved the final version of the manuscript.
Funding
This work was supported by the National Key R&D Programme of China (Ministry of Science and Technology of China, grant no. 2023YFC2506800); the National Natural Science Foundation of China (grant nos 82171544, 82371505, 82151314, 82101623, 82101582); the Clinical Research Special Project of Shanghai Municipal Health Commission of China (grant no. 202440203); and the Clinical Research Plan of Shanghai Hospital Development Centre (grant no. SHDC12022113).
Declaration of interest
None.
eLetters
No eLetters have been published for this article.