Clinically accessible eye-tracking in thyroid-associated ophthalmopathy: exploratory classification modeling and gaze allocation to disease-related periocular images Authors: Chunhui Yang, Ya Shen, Jian Song, Yuqing Chen, Chengcheng Zeng, Ruili Wei Published: 2026 ABSTRACT BackgroundThyroid-associated ophthalmopathy (TAO) is an autoimmune orbital disease that may cause ocular motility abnormalities, altered periocular appearance, and psychosocial burden. However, clinically practical tools for early recognition and objective functional assessment remain limited. This study aimed to evaluate whether clinically accessible eye-tracking could support auxiliary classification of thyroid-associated ophthalmopathy and characterize visual attention patterns toward disease-related periocular images.MethodsIn this cross-sectional study, 229 adults from a tertiary ophthalmology center were enrolled, including 181 patients with TAO and 48 healthy controls. Eye movements were recorded using a clinically accessible eye-tracking system during fixation, smooth pursuit with nine-grid spatial analysis, and image-viewing tasks. Random forest models were developed using nine-grid smooth-pursuit gaze-distribution features. A primary model included eye-tracking features plus age and sex, while a secondary model assessed eye-tracking features after demographic adjustment. Group differences in gaze metrics and region-of-interest allocation were analyzed using regression-based methods.ResultsFixation stability did not differ significantly between groups, but patients with TAO showed increased gaze dispersion. In smooth pursuit, the primary model achieved an AUC of 0.903, and the secondary model achieved an AUC of 0.778. In image-viewing tasks, gaze allocation differed between groups: compared with controls, patients with TAO paid less attention to more obvious pathological periocular features in earlier-stage images and more attention to near-recovery appearances in later-stage images. Distinct viewing patterns were also observed across preoperative and postoperative images from different surgical categories.ConclusionsClinically accessible eye-tracking may provide behavioral markers for auxiliary TAO classification and reveal altered visual attention to TAO-related periocular appearance. Further validation in larger, independent cohorts is needed.Trial registrationClinicalTrials.gov, NCT07381413. Prospectively registered. Introduction Thyroid-associated ophthalmopathy (TAO), also known as Graves’ ophthalmopathy (GO), is an autoimmune inflammatory disorder of the orbit characterized by extraocular muscle enlargement, fibrosis, and orbital adipose tissue expansion [1]. As a result of these structural and inflammatory changes, TAO may cause a wide range of ocular abnormalities, including proptosis, eyelid retraction, restrictive strabismus, ocular motility limitation, exposure-related discomfort, and, in severe cases, visual impairment [2]. Beyond its effects on ocular appearance and visual function, TAO imposes a profound psychosocial burden on affected individuals [3]. The inflammatory enlargement and subsequent fibrosis of extraocular muscles can mechanically restrict eye movements and lead to restrictive eye movement disorders [4]. Consistent with this pathophysiological basis, previous eye-movement studies in TAO have mainly focused on saccadic dynamics [5–8]. Although the findings have been heterogeneous across different disease stages, paradigms, and recording techniques, several studies have suggested that TAO may be associated with subtle abnormalities in saccadic eye movements, particularly in relation to disease severity, extraocular muscle involvement, and vertical gaze patterns [5–8]. Owing to pathological alterations of orbital tissues, patients with TAO frequently present with proptosis, eyelid retraction, and a staring appearance [9]. Even when emotionally neutral, they may be perceived by others as fearful, angry, or tense because of altered facial and periocular appearance [10]. This mismatch in facial affective signaling disrupts normal social interactions and contributes to social anxiety, depressive symptoms, and disturbances in self-image [11]. Although instruments such as the GO-QoL questionnaire are commonly used to evaluate psychosocial burden, these disturbances may also be reflected in observable visual attention patterns and gaze behavior [12]. Despite increasing understanding of TAO, clinically practical tools for its early recognition and objective assessment remain limited. First, early signs are often subtle and may be underrecognized, particularly in non-specialist settings, resulting in delayed referral and intervention [13, 14]. Second, although orbital imaging provides important structural evidence, its cost and limited accessibility reduce its practicality for broad screening and routine use in primary care [15, 16]. In addition, imaging-based evaluation focuses mainly on anatomical abnormalities and may not adequately capture dynamic ocular functional disturbances. Accordingly, the development of a practical, non-invasive, and function-oriented auxiliary assessment approach remains of substantial clinical interest. Eye-tracking technology offers a non-invasive approach to capturing gaze behavior [17]. As an emerging form of digital biomarker, it enables the quantification of gaze stability, tracking errors, and other dynamic behavioral features, while also allowing objective assessment of attentional capture from a perceptual perspective. In this context, eye-tracking may provide a multidimensional assessment approach that integrates physiological function with psychosocial burden. Accordingly, the objectives of this study were twofold: first, to develop a non-invasive eye-tracking-based model for the auxiliary identification of TAO; and second, to investigate visual attention patterns in patients with TAO when viewing images depicting disease manifestations and postoperative recovery. Methodology Participants This study was registered on ClinicalTrials.gov in January 2026. Between February and March 2026, a total of 229 adults were recruited from the Department of Ophthalmology at Shanghai Changzheng Hospital. This cohort included 181 patients with thyroid-associated ophthalmopathy (TAO) and 48 healthy controls. TAO was diagnosed according to EUGOGO criteria. Participants were aged ≥ 18 years and able to complete the eye-tracking tasks. Patients with severe diplopia, marked ocular misalignment in primary gaze, or other ocular, neurological, or psychiatric conditions that could compromise eye-tracking acquisition were excluded, while those with mild diplopia were retained to reflect the real-world clinical spectrum. Healthy controls had no history of TAO or ocular motility disorders. Demographic and clinical characteristics, including age, sex, disease activity, and severity, were recorded. The study was prospectively registered on ClinicalTrials.gov (Registration number: NCT07381413) and conducted in accordance with the Declaration of Helsinki, and all participants provided written informed consent. Eye-tracking system Eye movements were recorded using a clinically accessible eye-tracking system (Airdoc, Beijing, China) with a sampling frequency of 30 Hz. Participants were seated at approximately 60 cm from the display under standardized indoor lighting. A nine-point calibration was performed before data acquisition to ensure gaze accuracy. Visual tasks A randomization protocol was implemented in this study, whereby the sequence of photographic stimuli presented to each participant was fully randomized. Throughout the visual attention tasks, participants remained entirely naive to the specific clinical diagnoses or categories of the presented images. Fixation stability task Participants were instructed to fixate a single static target at the center of the screen to evaluate gaze stability. Smooth pursuit and nine-grid task In this task, participants were asked to follow a visual target moving along a predefined trajectory on the display. Eye movements were recorded binocularly. During the pursuit, the display was divided into a 3 × 3 grid (grids 1–9) to quantify the distribution of gaze across spatial locations. Proportional dwell time within each grid was calculated and used as the primary feature for subsequent modeling. Any blinks or tracking failures were interpolated or excluded according to pre-specified thresholds. Region of interest (ROI) for social attention Stimuli for the visual attention tasks comprised standardized digital periocular photographs acquired from an independent cohort of TAO patients, documenting a broad spectrum of clinical severity, longitudinal disease progression, and post-recovery states. This included preoperative and postoperative images of three specific surgical procedures: orbital decompression, eyelid retraction correction, and strabismus surgery. To mitigate image-level confounding factors, all photographic stimuli underwent systematic preprocessing: images were standardized and digitally balanced to ensure identical brightness, contrast, crop ratio, viewing angle and color saturation. Structurally, the ROI was predefined consistently across images as a unified region that encompassed the bilateral ocular areas, explicitly including the upper and lower eyelids alongside the immediate periocular anatomy. Statistical analysis and model construction Statistical analyses were performed using R software. Continuous variables following a normal distribution were analyzed using independent-sample t tests, whereas non-normally distributed variables were analyzed using the Mann–Whitney U test. Categorical variables were compared using the chi-square test. To evaluate between-group differences in gaze-related metrics and within-group spatial distribution patterns, linear regression models adjusted for age and sex were constructed, along with linear mixed-effects models (LMMs) incorporating subject-level random effects. Post hoc pairwise comparisons were conducted using Bonferroni or Tukey corrections, as appropriate. To assess classification performance, propensity score matching (PSM) was applied to balance baseline characteristics between groups. To assess diagnostic performance, two random forest models were developed using five-fold cross-validation. The primary model incorporated nine-grid smooth-pursuit gaze distribution features together with age and sex to evaluate overall classification performance. To further assess the independent classification value of eye-tracking parameters beyond demographic confounding, a secondary model was constructed using eye-tracking features after adjustment for age and sex. To prevent information leakage, the validation fold was completely isolated during model training. Model performance and clinical utility were evaluated using the receiver operating characteristic curve (ROC-AUC), calibration curves, and decision curve analysis (DCA). A two-sided P value < 0.05 was considered statistically significant. Results Baseline characteristics section A total of 229 participants were included in this study, consisting of 181 patients with TAO and 48 healthy controls. Table 1 summarizes the demographic and clinical characteristics of the study population. Significant differences were observed in age and sex distribution between groups. Deconstructing gaze characteristics in TAO: dissociation between stability and dispersion Quantitative analysis of the fixation task revealed a dissociation between gaze stability and gaze dispersion in patients with TAO. No significant difference was observed between the TAO and control groups in fixation stability (p = 0.478), indicating comparable performance in maintaining single-point fixation during the task (Fig. 1A). This finding is consistent with previous reports [6]. In contrast, gaze dispersion metrics demonstrated altered spatial distribution patterns in the TAO group (Fig. 1B). Construction of primary and eye-tracking-independent diagnostic models based on smooth-pursuit gaze distribution A primary diagnostic model for TAO was first constructed by integrating nine-grid gaze distribution during smooth pursuit with demographic variables. This model achieved an AUC of 0.903, indicating excellent classification performance. Feature importance analysis showed that, in addition to age and sex, gaze distributions in grids 1 (upper left) and 9 (lower right) contributed substantially to discrimination. To further determine whether eye-tracking features provided diagnostic information independent of demographic factors, a secondary model was constructed after adjustment for age and sex using only the proportional gaze distribution across the nine-grid. This model achieved an AUC of 0.778 (Fig. 2A), indicating that eye-tracking parameters retained meaningful discriminatory ability even after demographic correction. Feature importance ranking (Fig. 2B) identified grids 9 (lower right), 6 (right center), and 1 (upper left) as the most informative contributors. DCA (Fig. 2C) demonstrated that, across a relatively wide range of threshold probabilities, the model yielded a higher net benefit than the strategies of classifying all subjects as TAO or classifying all subjects as normal, supporting its potential utility in auxiliary diagnosis and screening of TAO. From prominent pathology to clinical recovery: longitudinal changes in social attention in TAO After adjustment for age and sex, gaze proportions directed toward the ocular grid were compared between the control and TAO groups while viewing images representing different stages of recovery (ROI 1–6; Fig. 3). The two groups exhibited markedly divergent gaze allocation patterns with a pronounced crossover effect. During early disease stages characterized by severe pathological features, the control group allocated a significantly higher proportion of gaze to the eye region than the TAO group, with the difference reaching statistical significance at ROI 3 (p < 0.005). This pattern reflects stronger attentional capture by conspicuous pathological signs such as erythema and proptosis in healthy observers. As pathological features gradually resolved across the image sequence, the gaze allocation trends of the two groups reversed around ROI 4. In the late recovery stage, the TAO group demonstrated a substantial increase in attention to the eye region and significantly exceeded that of the control group at ROI 6 (p = 0.037). These findings reveal possible differences in visual processing strategies between groups: while gaze behavior in healthy individuals is primarily driven by abnormal pathological cues, TAO patients exhibit heightened sensitivity and evaluative scrutiny toward near-normal ocular appearances. From pathological distortion to morphological reconstruction: reorganization of gaze patterns following three surgical interventions Regarding overall attention allocation across image categories (Fig. 4A-B), TAO patients showed a significantly higher proportion of gaze directed toward eyelid retraction images compared with controls (p < 0.001), while their attention to orbital decompression and strabismus images was significantly reduced (p < 0.01). Within-group comparisons further confirmed a strong selective attentional bias toward eyelid retraction images in the TAO group (p < 0.001). In contrast, the control group exhibited relatively uniform gaze distribution across the three image categories, with no significant differences. Region-of-interest analyses comparing preoperative and postoperative images revealed distinct group-specific visual patterns (Fig. 4C). In the eyelid retraction region, the TAO group allocated significantly more gaze to postoperative images than to preoperative images (p < 0.001), whereas the control group showed no significant difference (p = 0.407). In the strabismus region, the two groups displayed opposite attentional tendencies: the TAO group preferentially attended to postoperative images (p = 0.028), while the control group showed significantly greater attention to preoperative images with more prominent pathological features (p < 0.001). In the orbital decompression region, only the control group demonstrated a significant preference for postoperative images (p = 0.042), whereas no significant difference was observed between preoperative and postoperative images in the TAO group (p = 0.325). Collectively, these findings indicate that patients with TAO exhibited a distinct pattern of visual attention across disease-related image categories and between preoperative and postoperative conditions. Discussion The present study suggests that clinically accessible eye-tracking may detect functional gaze abnormalities in TAO beyond conventional static fixation measures. In our cohort, fixation stability did not differ significantly between patients with TAO and healthy controls, whereas task-dependent gaze features, particularly those derived from nine-grid paradigms, showed potential discriminatory value. In addition, patients exhibited altered visual attention patterns toward disease-related and postoperative images, indicating that TAO may influence not only ocular motor behavior but also the visual processing of periocular appearance-related information. These findings should be interpreted in the context of prior studies on ocular motility in TAO, which have yielded inconsistent results across different study designs and patient populations. Earlier studies focusing on classical saccadic dynamics have produced inconsistent results, with reports of unchanged, increased, or decreased saccadic velocity. In particular, Schworm et al. initially reported that in early active Graves’ ophthalmopathy, analysis of multiple individual saccadic parameters revealed only limited statistically significant differences, with substantial inter- and intra-individual variability and marked overlap between patients and controls, leading them to conclude that conventional saccadic analysis was not a clinically useful diagnostic tool in early disease [6]. However, in a subsequent study using main-sequence analysis and more explicit disease-stage stratification, the same group detected significant abnormalities across TAO stages, including early disease, indicating that subtle ocular motor changes may be present but are more readily identified by stage-sensitive and model-based approaches than by isolated conventional saccadic metrics [8]. Similarly, Träisk et al. found no significant differences in main-sequence parameters between patients with predominantly active, non-fibrotic TAO using either magnetic scleral search coil or infrared reflection eye-tracking systems. They further suggested that neural adaptation and substantial heterogeneity in both the normal and TAO populations may obscure early disease-related differences in conventional saccadic metrics [7]. Our results are broadly consistent with these previous observations, but they also extend them. Rather than identifying a robust abnormality in classical high-precision saccadic kinematics, our study indicates that TAO-related dysfunction may become more apparent at the level of gaze behavior during task performance. In other words, TAO may not always produce a stable and reproducible saccadic velocity signature, especially in early or clinically compensated stages, but it may still alter the spatial distribution of gaze, target-following behavior, and visual attention allocation. This distinction may help explain why earlier studies based on traditional saccadic parameters often yielded negative or conflicting findings, whereas behavior-level eye-tracking features in the present study retained discriminatory information. Such an interpretation is also compatible with the hypothesis proposed in previous reports that central ocular motor adaptation may partially compensate for early orbital or extraocular muscle changes, thereby preserving some conventional saccadic metrics despite underlying disease [6]. An important finding of our study was the absence of a significant between-group difference in fixation stability. This suggests that static fixation may be relatively preserved in TAO, at least in our cohort. However, this should not be taken as evidence of normal ocular motor function overall. TAO-related abnormalities may be more readily detected during dynamic or spatially distributed tasks than during simple stationary fixation. Our results therefore suggest that fixation measures alone may have limited sensitivity for identifying clinically meaningful functional changes in TAO. The smooth-pursuit and nine-grid paradigms are of particular interest because they appear to capture broader gaze-behavior alterations than conventional static measures. The observed differences in gaze distribution may reflect the combined effects of subtle motility restriction, gaze asymmetry, and compensatory viewing strategies. Clinically, this is relevant because such features can be obtained using relatively accessible and non-invasive eye-tracking methods. Thus, eye-tracking-derived functional markers may serve as an auxiliary tool to complement, rather than replace, standard structural and clinical assessments in TAO. Another notable finding was the altered visual attention pattern toward disease-related and postoperative images in patients with TAO. This may indicate that, compared with healthy controls, patients differ in their visual sensitivity or attentional bias toward periocular appearance-related information. However, these interpretations should remain cautious, as eye-tracking data alone cannot determine the underlying psychological mechanism. A more comprehensive understanding would require integration with additional measures, such as fMRI, psychological questionnaires, and other behavioral or neurophysiological assessments. Therefore, our results are better understood as indicating altered visual processing of TAO-relevant facial information, rather than providing direct evidence of a specific emotional or cognitive response. Importantly, our results do not contradict earlier studies that questioned the usefulness of classical saccadic analysis in early TAO. Rather, they highlight a difference in analytical level. Previous work primarily examined whether traditional ocular motor parameters, such as peak saccadic velocity or main-sequence constants, could distinguish patients from controls [7, 8]. In contrast, the present study focused on clinically accessible gaze behavior, including task-related gaze distribution and disease-related visual attention. It is therefore plausible that even if classical saccadic metrics are not sufficiently robust for early discrimination in routine clinical practice, broader gaze-behavior features may still carry clinically relevant information. This broader framework may be particularly appropriate for TAO, given the heterogeneity of disease presentation and the likely interaction between orbital mechanics and central compensatory mechanisms. Several limitations should be acknowledged. First, although pivotal clinical dimensions, including CAS, severity, and ocular motility restrictions, have been successfully integrated into our baseline characterization, the impact of inherent clinical heterogeneity cannot be entirely disregarded. Specifically, owing to the pragmatic constraints of routine clinical documentation, granular metrics such as disease durations and fragmented external treatment histories could not be systematically standardized. Given that these unmeasured baseline variations might subtly modulate gaze kinetics, the lack of comprehensive subtype stratification remains a limitation that warrants further investigation. Second, a baseline numerical imbalance existed between the patient and control groups during initial recruitment. Although we mitigated the potential for majority-class bias and statistical confounding by executing a one-to-one propensity-score matching protocol prior to model training, subsequent validation in naturally balanced, independent cohorts would be beneficial to further confirm the performance stability of our approach. Third, this study was conducted within a single-center cohort and lacks an independent external validation dataset. Although the classification framework demonstrated promising performance within the current pool, the absence of external validation limits the immediate generalizability of our findings to broader populations. Fourth, if a lower-frequency clinical eye-tracker was used, the present metrics should not be interpreted as equivalent to the high-resolution saccadic parameters reported in search coil or laboratory infrared studies. Our findings are more appropriately understood as behavioral eye-tracking markers than as precise measurements of ocular motor physiology. Fifth, the interpretation of disease-related and postoperative viewing patterns is limited by the absence of concurrent psychosocial measures such as quality-of-life scales, appearance-related distress measures, or anxiety and depression assessments. ly, the cross-sectional design does not allow conclusions regarding causality or the predictive value of these eye-tracking features for disease progression, treatment response, or postoperative outcome. Conclusion In conclusion, clinically accessible eye-tracking may provide useful behavioral markers for the auxiliary classification of thyroid-associated ophthalmopathy. In this cross-sectional study, dynamic gaze features, particularly those derived from smooth pursuit and nine-grid spatial distribution, showed potential screening-oriented value beyond conventional static fixation measures. In addition, exploratory analyses suggested that patients with TAO exhibit altered visual attention to disease-related and postoperative periocular appearance. Together, these findings support the potential of eye-tracking as a non-invasive adjunct to conventional clinical assessment in TAO. Further validation in larger, independent, and clinically stratified cohorts is needed to confirm its robustness, clarify its incremental value, and determine its applicability in real-world screening and follow-up settings. Fig. 1Gaze characteristics in TAO patients and healthy controls. (A) Fixation stability (Score %); (B) Gaze dispersion (Divergence 0–1). Box plots represent the median and interquartile range (IQR) with individual data points overlaid Gaze characteristics in TAO patients and healthy controls. (A) Fixation stability (Score %); (B) Gaze dispersion (Divergence 0–1). Box plots represent the median and interquartile range (IQR) with individual data points overlaid Fig. 2Performance and clinical utility of the predictive model: (A) ROC showing sensitivity as a function of specificity; (B) Variable importance plot illustrating the relative contribution of different predictors (Grid 1–9) to the model; and (C) DCA demonstrating the net benefit of the model (pink line) across a relatively wide range of high-risk threshold probabilities, compared with the strategies of classifying all subjects as TAO (light grey line) or classifying all subjects as normal (dark grey line) Performance and clinical utility of the predictive model: (A) ROC showing sensitivity as a function of specificity; (B) Variable importance plot illustrating the relative contribution of different predictors (Grid 1–9) to the model; and (C) DCA demonstrating the net benefit of the model (pink line) across a relatively wide range of high-risk threshold probabilities, compared with the strategies of classifying all subjects as TAO (light grey line) or classifying all subjects as normal (dark grey line) Fig. 3Gaze allocation toward the ocular region across stages of clinical recovery. Six representative facial images, ordered from the upper left to the lower right, depict the progression from acute severe TAO to an almost healthy appearance after recovery. For each image, the bilateral eye area was defined as one image-specific region of interest (ROI), resulting in ROI 1–ROI 6 corresponding sequentially to images 1–6. Gaze proportions directed to these ocular ROIs were compared between TAO patients and controls Gaze allocation toward the ocular region across stages of clinical recovery. Six representative facial images, ordered from the upper left to the lower right, depict the progression from acute severe TAO to an almost healthy appearance after recovery. For each image, the bilateral eye area was defined as one image-specific region of interest (ROI), resulting in ROI 1–ROI 6 corresponding sequentially to images 1–6. Gaze proportions directed to these ocular ROIs were compared between TAO patients and controls Fig. 4Gaze allocation across three surgical intervention categories. Representative images are arranged by surgical type and perioperative stage: the upper-left and lower-left panels show preoperative and postoperative orbital decompression, respectively; the upper-middle and lower-middle panels show preoperative and postoperative eyelid retraction surgery, respectively; and the upper-right and lower-right panels show preoperative and postoperative strabismus surgery, respectively. The bilateral ocular regions in each image were defined as ROIs for gaze analysis. (A) Distribution of ROI ratios across the three surgical categories. (B) Mean ROI ratios in the TAO and CON groups within each category. (C) Comparison of ROI ratios between preoperative (Pre) and postoperative (Post) images for each surgical category Gaze allocation across three surgical intervention categories. Representative images are arranged by surgical type and perioperative stage: the upper-left and lower-left panels show preoperative and postoperative orbital decompression, respectively; the upper-middle and lower-middle panels show preoperative and postoperative eyelid retraction surgery, respectively; and the upper-right and lower-right panels show preoperative and postoperative strabismus surgery, respectively. The bilateral ocular regions in each image were defined as ROIs for gaze analysis. (A) Distribution of ROI ratios across the three surgical categories. (B) Mean ROI ratios in the TAO and CON groups within each category. (C) Comparison of ROI ratios between preoperative (Pre) and postoperative (Post) images for each surgical category Table 1Demographic and clinical characteristicsCharacteristicsNormalN = 481TAON = 1811p-value2Age (years)33.96 ± 14.8743.89 ± 12.06< 0.001Gender< 0.001 Female8 (17%)124 (69%) Male40 (83%)56 (31%)CAS < 387 (48.1%) ≥ 394 (51.9%)Disease severity Mild34 (19%) Moderate-to-severe144 (80%) Very severe3 (1.7%)Ocular motility restriction Yes101 (56%) No80 (44%)1Mean (SD); n (%)2Welch Two Sample t-test; Pearson's Chi-squared test Table 1Demographic and clinical characteristicsCharacteristicsNormalN = 481TAON = 1811p-value2Age (years)33.96 ± 14.8743.89 ± 12.06< 0.001Gender< 0.001 Female8 (17%)124 (69%) Male40 (83%)56 (31%)CAS < 387 (48.1%) ≥ 394 (51.9%)Disease severity Mild34 (19%) Moderate-to-severe144 (80%) Very severe3 (1.7%)Ocular motility restriction Yes101 (56%) No80 (44%)1Mean (SD); n (%)2Welch Two Sample t-test; Pearson's Chi-squared test Demographic and clinical characteristics 1Mean (SD); n (%) 2Welch Two Sample t-test; Pearson's Chi-squared test