Eye-tracking phenotypes reveal clinically meaningful heterogeneity of attention and executive control dysfunction after stroke

Abstract

Background

Post-stroke impairments in attention and executive control are common and clinically relevant, whereas conventional neuropsychological tests mainly capture task outcomes and provide limited information about process-level task behavior. Eye tracking can quantify fixation behavior, saccadic activity, and visual sampling, but whether integrated eye-tracking features can reveal clinically interpretable heterogeneity after stroke remains insufficiently defined.

Objective

This study examined whether integrated multi-paradigm eye-tracking features could identify exploratory eye-movement phenotypes in individuals with stroke and whether these phenotypes were associated with behavioral, clinical, lesion-related, and functional characteristics.

Methods

In this single-center exploratory case-control study, 72 subacute stroke participants (14–90 days post-onset) and 36 matched healthy controls completed five tasks of attention and executive control and six eye-tracking paradigms. Group differences in behavioral and eye-tracking measures were examined with false discovery rate correction. Within the stroke group, K-means clustering was performed using four nonredundant integrated eye-tracking features: total fixation duration, first-fixation duration, number of fixations, and total saccade distance. Cluster stability and sensitivity analyses were conducted using bootstrap resampling, alternative feature sets, and combined clustering with healthy controls. Post-clustering comparisons and exploratory regression analyses were used to evaluate the clinical relevance of the identified phenotypes.

Results

Compared with healthy controls, participants with stroke showed lower performance across attention/executive-control tasks and altered eye-tracking metrics across paradigms. The most consistent eye-movement pattern was shorter total fixation duration and a higher number of fixations across the six paradigms, whereas total saccade distance differed most clearly during the antisaccade task. K-means clustering identified two exploratory eye-movement phenotypes: a High-Fixation/Low-Saccade phenotype and a Low-Fixation/High-Saccade phenotype. The two-cluster solution remained stable across bootstrap and sensitivity analyses. The Low-Fixation/High-Saccade phenotype showed poorer cognitive performance, greater neurological and functional impairment, higher emotional symptom burden, and more frequent frontal involvement and bilateral or multifocal involvement. In exploratory hierarchical regression, eye-movement phenotype explained additional variance in Barthel Index scores beyond demographic variables, post-stroke timing, neurological severity, and conventional attention/executive-control performance.

Conclusions

Integrated multi-paradigm eye-tracking features identified two exploratory eye-movement phenotypes in individuals with subacute stroke. The Low-Fixation/High-Saccade phenotype was associated with broader clinical burden. These findings suggest that eye tracking may provide process-level information complementary to conventional assessment of post-stroke attention and executive-control dysfunction. Given the single-center and cross-sectional design, these preliminary phenotypes require validation in longitudinal and multicenter studies.

Introduction

Stroke-related cognitive impairment is a major determinant of long-term disability, reduced participation, and limited rehabilitation gains [1, 2]. Among its cognitive components, impairments in attention and executive control are particularly relevant because they affect goal maintenance, information selection, response monitoring, cognitive flexibility, and the coordination of cognition with movement during daily activities and therapy [3, 4]. In clinical practice, these abilities are commonly assessed using conventional neuropsychological tests [5]. Although such tests are useful and accessible, they mainly provide outcome-level information and are influenced by multiple task demands, including processing speed, working memory, language comprehension, auditory processing, and motor or written output [6, 7]. Therefore, conventional test scores alone may not fully characterize the process-level mechanisms underlying task performance after stroke [8].

Eye tracking provides a complementary approach by measuring gaze allocation, fixation behavior, and saccadic activity during task performance [9]. These measures can reflect how individuals visually sample information, maintain fixation, shift gaze, and regulate oculomotor responses [10]. Compared with conventional paper-and-pencil or verbally mediated assessments, eye tracking may capture aspects of visual attention and oculomotor control with relatively low requirements for limb movement or verbal output [11]. This feature is relevant in stroke populations, in whom motor impairment, fatigue, aphasia, and reduced processing speed may affect conventional test performance [12]. However, eye-tracking metrics should not be considered direct substitutes for neuropsychological tests. Rather, they provide process-level information that may help explain how behavioral performance is achieved or disrupted.

Previous stroke studies have linked eye-movement abnormalities to cognitive impairment [13], visuospatial neglect [14], aphasia-related processing [15], and oculomotor dysfunction [16]. For example, eye-tracking studies in unilateral spatial neglect have shown characteristic abnormalities in visual exploration, including reduced contralesional exploration, altered fixation distribution, and inefficient visual search patterns [17]. Eye-movement studies in aphasia have also demonstrated task-related alterations in fixation behavior, saccadic activity, and visual sampling during language-related and non-linguistic visual tasks [18]. These findings indicate that gaze behavior after stroke can be shaped by visual-spatial impairment, language-related processing difficulty, and primary or network-level oculomotor dysfunction, rather than by attention/executive-control dysfunction alone [19]. However, much of the existing work has focused on single paradigms, specific syndromes, or isolated metrics [20]. This limits the ability to determine whether eye-movement behavior reflects stable cross-task patterns or only task-specific responses. In addition, many studies emphasize group-level differences between patients and controls, while less attention has been paid to heterogeneity within the stroke population [21]. Patients with similar conventional test scores may differ substantially in fixation stability, saccadic activity, visual sampling strategy, and compensatory gaze behavior [22].

A multi-paradigm eye-tracking battery may help address this gap because different paradigms place different demands on visual engagement and oculomotor control [23]. Static picture viewing and dynamic video viewing provide relatively naturalistic viewing conditions [24]. Dynamic path tracking requires sustained following of a moving target across more complex trajectories [25]. The oculomotor task probes visually guided eye-movement control under simpler movement structures [26]. Prosaccade tasks emphasize rapid gaze shifting toward a target, whereas antisaccade tasks additionally require suppression of a reflexive response and generation of a voluntary gaze shift in the opposite direction [27]. By integrating information across these paradigms, it may be possible to characterize individual differences in post-stroke visual sampling and gaze control more comprehensively than by relying on a single task.

Accordingly, the present study focused primarily on eye-movement-based phenotyping in individuals with stroke, while using conventional tasks of attention and executive control as behavioral validators rather than as an equal parallel endpoint. We aimed to: first, characterize group differences in attention/executive-control task performance and eye-tracking metrics between participants with stroke and healthy controls; second, identify exploratory eye-movement phenotypes within the stroke group using integrated multi-paradigm features; and third, examine whether these phenotypes differed in attention/executive-control performance, global cognition, neurological severity, functional independence, motor function, emotional status, and lesion-related characteristics. We further explored whether eye-movement phenotype explained additional variance in functional independence.

Methods

Study design and ethics

This was a single-center exploratory case-control study conducted in the Department of Rehabilitation Medicine, The Third Affiliated Hospital of Soochow University. All participants completed clinical assessments, tasks of attention and executive control, and eye-tracking assessment in a quiet standardized environment. Assessments were administered by trained rehabilitation physicians according to standardized procedures. The study protocol was approved by the Ethics Committee of The Third Affiliated Hospital of Soochow University (No. 2023-14), and written informed consent was obtained from all participants before enrollment.

Participants

Participants with stroke were consecutively recruited from December 2023 to December 2025. Stroke diagnosis was confirmed by CT or MRI according to established Chinese and international stroke diagnostic criteria. Inclusion criteria were: (1) age > 18 years; (2) clinically confirmed ischemic or hemorrhagic stroke; (3) clear consciousness; (4) ability to understand and cooperate with attention/executive-control tasks and eye-tracking testing; (5) stroke onset-to-assessment interval between 14 and 90 days; and (6) provision of written informed consent by the participant.

Exclusion criteria were: (1) history of other neurological diseases or major psychiatric disorders; (2) use of medications that might substantially affect cognition, alertness, or mental status; (3) severe cognitive impairment preventing comprehension or cooperation; (4) severe aphasia that substantially interfered with understanding task instructions or providing valid behavioral responses; (5) severe visual impairment, serious ophthalmic disease, or ocular motor disorder precluding fixation or eye-tracking tasks; (6) significant unilateral spatial neglect or hemianopia that prevented valid completion of the eye-tracking paradigms; (7) inability to complete the behavioral or eye-tracking procedures for other reasons; or (8) refusal or withdrawal.

During the same period, healthy controls were recruited and matched to the stroke group by age, sex, and years of education. Controls had no history of stroke, major neurological disease, or active psychiatric illness and completed the same behavioral and eye-tracking procedures under the same testing conditions.

Demographic, clinical, and lesion characteristics

Demographic variables included age, sex, and years of education. For participants with stroke, stroke-related variables included stroke type, time from stroke onset to assessment, lesion laterality, frontal involvement, brainstem involvement, bilateral or multifocal lesions, and number of lesions. Time from stroke onset to assessment was recorded in days. Because all stroke participants were assessed between 14 and 90 days after onset, the sample represented the subacute post-stroke stage.

Lesion-related information was extracted from clinically available CT or MRI reports. Lesion involvement was coded as left hemisphere involvement, right hemisphere involvement, bilateral or multifocal involvement, frontal involvement, and brainstem involvement based on clinically available CT or MRI reports. These lesion-location variables were not mutually exclusive because some participants had bilateral or multifocal lesions. Frontal involvement and brainstem involvement were coded as present or absent based on the documented lesion location. Bilateral or multifocal lesions were defined as two or more anatomically distinct lesions documented on clinical neuroimaging reports.

Attention and executive-control measures

The behavioral test battery was used to characterize task performance across attention- and executive-control-related operations. Because several tasks involve overlapping cognitive processes, the measures were interpreted as tasks of attention and executive control rather than as process-pure indicators of isolated attention domains.

1. Digit Span Test [28]: Attention span and working-memory-related performance were assessed using the Digit Span Test. Digit sequences were read aloud at approximately one digit per second. Participants completed forward and backward digit span trials. Forward span started with three digits and increased progressively up to 12 digits or stopped after two consecutive errors. Backward span started with two digits and increased progressively up to 10 digits or stopped after two consecutive errors. The maximum correctly repeated span was recorded.

2. Cancellation-based tasks: Two digit cancellation conditions were administered [29, 30]. In the first condition, participants were instructed to cross out the digit “3”. In the second condition, participants were instructed to cross out the digit immediately preceding “3”. These two cancellation conditions were treated as cancellation-based attention/executive-control tasks that involve sustained task engagement, target selection, rule maintenance, and response monitoring, rather than as fully dissociable measures of sustained and selective attention.

The test comprised four pages, each containing 10 rows × 20 digits, with 800 digits in total. Outcomes included correct responses, incorrect responses, omissions, net score, and error rate. Error rate was calculated as: error rate= ( incorrect responses + omissions / 2) / correct responses ×100%.

For analysis and presentation, cancellation hit rate was calculated as 100% minus the corresponding error rate, so that higher values consistently indicated better performance.

• 3.Symbol Digit Modalities Test [31]: Attention shifting, processing speed, and cognitive set-shifting performance were assessed using the written version of the Symbol Digit Modalities Test. Participants completed symbol-digit substitutions according to a coding key within 90 s. Each correct response was scored as 1 point, and reversed symbols were scored as 0.5 points. The maximum score was 90.
• 4.Paced Auditory Serial Addition Test [32, 33]: Divided attention and online updating were assessed using the Paced Auditory Serial Addition Test. An audio recording presented one digit every 2 s. Participants were instructed to continuously add each digit to the immediately preceding digit. The total number of correct responses was recorded, with a maximum score of 60.

Eye-tracking assessment

Apparatus and setting

Eye-tracking tasks were administered using cognitive rehabilitation training and assessment software (JZ-RZ-20US, Hangzhou Jizhi Medical, Hangzhou, China), which integrates multiple eye-tracking paradigms. Stimuli were presented on a 24-inch monitor. Eye movements were recorded using a Tobii-based eye-tracking system at a sampling frequency of 120 Hz. Participants were seated approximately 60 cm from the screen, and head position was stabilized using a chin rest. Testing was conducted in a quiet room under standardized lighting conditions.

Calibration and general procedure

Before formal testing, participants completed a standardized 5-point calibration procedure according to the device instructions. If calibration quality was judged insufficient by the software or examiner, calibration was repeated before task administration. After successful calibration, participants completed the six eye-tracking paradigms in a fixed order. A central red fixation cross was presented between trials or task segments to re-establish central fixation. The total eye-tracking testing time was approximately 15 min. After testing, feature-level eye-tracking outputs were exported for analysis.

Eye-tracking paradigms

Six eye-tracking paradigms were administered.

1. Static picture task: Thirteen color photographs of children or infants, each measuring 19 cm × 15.5 cm, were presented sequentially for 5 s each. A central red fixation cross, 1 cm × 1 cm, was shown for 2 s between images. Participants were instructed to look at each picture naturally and attentively throughout the presentation period.

2. Dynamic path task: A cartoon stimulus moved along multiple trajectories, including horizontal reciprocation, diagonal movement, circular movement, S-shaped movement, and Tai Chi-like trajectories. Each segment lasted 5 s, with a 2-second central fixation cross between segments. Participants were instructed to keep their eyes on the moving target and follow its trajectory as accurately as possible. This paradigm was designed to assess sustained visual tracking under relatively complex movement paths.

3. Oculomotor task: A cartoon stimulus was either fixed centrally or moved vertically, horizontally, or circularly. Each segment lasted 5 s, with a 2-second central fixation cross between segments. Participants were instructed to look directly at the target and follow it whenever it moved while keeping their head still. Compared with the dynamic path task, this paradigm used simpler movement structures and was intended to probe visually guided eye-movement control under less complex tracking demands.

4. Dynamic video task: A full-screen video was presented for 30 s. Participants were instructed to watch the video attentively in the same way as they would watch a normal video clip. This paradigm provided a more naturalistic viewing condition.

5. Prosaccade task: A cartoon target appeared in the upper, lower, left, or right position relative to the center of the screen. Each trial lasted 5 s, with a 2-second central fixation cross between trials. Participants were instructed to shift their gaze as quickly and accurately as possible to the target when it appeared. This paradigm emphasized visually guided gaze shifting toward a target.

6. Antisaccade task: A cartoon target appeared in the upper, lower, left, or right position relative to the center of the screen. Each trial lasted 5 s, with a 2-second central fixation cross between trials. Participants were instructed not to look at the target itself but instead to look as quickly and accurately as possible to the opposite side. This paradigm emphasized suppression of a reflexive gaze response and generation of a voluntary gaze shift according to an internal rule.

For the prosaccade and antisaccade paradigms, peripheral targets were generated by the built-in software protocol at predefined upper, lower, left, and right screen positions. Vertical and horizontal target directions were administered within the same paradigm according to the device protocol rather than as investigator-defined separate horizontal and vertical blocks. The exported feature-level dataset did not retain numeric target eccentricity in centimeters or degrees of visual angle; therefore, target eccentricity was described qualitatively and was not entered as an analytical variable.

Eye-tracking metrics

For each paradigm, the following feature-level metrics were extracted and averaged: total fixation duration, fixation duration ratio, first-fixation duration, number of fixations, total saccade distance, mean saccade velocity, and proportion of fixation duration within areas of interest. Fixation duration ratio was calculated as total fixation duration divided by task duration, and mean saccade velocity was calculated as total saccade distance divided by task duration.

Eye-tracking preprocessing and data-quality control

The eye-tracking system provides standardized vendor-implemented processing procedures at the device/software level, including blink detection, temporary gaze-loss identification, invalid-sample detection, and fixation/saccade event classification. The present study used structured feature-level outputs exported by the system rather than raw gaze-coordinate trajectories. Therefore, low-level preprocessing of raw gaze data, including blink handling, missing-sample identification, tracking-loss detection, and fixation/saccade event classification, was performed automatically by the built-in algorithms of the eye-tracking system. No additional offline preprocessing, interpolation, smoothing, or reconstruction of raw gaze trajectories was performed by the investigators.

Data quality was further controlled at the participant level based on the completeness and plausibility of the exported feature-level data. After data export, the feature-level dataset was inspected across all six paradigms. For each participant, we verified whether all predefined eye-tracking metrics were successfully generated for all required tasks and trials. Participants were excluded from the final analysis if any required task- or trial-level feature output was missing, could not be exported, or contained abnormal values that precluded reliable calculation of task-level or cross-task composite eye-tracking features. No trial-level imputation or partial-case analysis was performed.

A total of 75 participants with stroke and 38 healthy controls initially completed the eye-tracking assessment. After participant-level data-quality control, 3 participants with stroke and 2 healthy controls were excluded because of incomplete or abnormal feature-level eye-tracking outputs. Therefore, the final analyses included 72 participants with stroke and 36 healthy controls, all of whom had complete analyzable eye-tracking feature data across the six paradigms.

Clinical scales

Global cognition was assessed using the Mini-Mental State Examination [34], with scores ranging from 0 to 30 and higher scores indicating better global cognitive performance. Neurological impairment was assessed using the National Institutes of Health Stroke Scale [35], with scores ranging from 0 to 42 and higher scores indicating more severe neurological deficits. Activities of daily living were assessed using the Barthel Index [36], with scores ranging from 0 to 100 and higher scores indicating greater functional independence. Motor function was evaluated using the Fugl–Meyer Assessment [37], with total scores ranging from 0 to 100. Anxiety and depressive symptoms were measured using the Self-Rating Anxiety Scale and Self-Rating Depression Scale [38, 39], respectively; standardized scores were calculated as raw score × 1.25, with higher scores indicating greater emotional symptom burden.

Study outcomes and analytical framework

The analytical framework was structured around prespecified and exploratory objectives. The primary analyses compared performance on tasks of attention and executive control and eye-tracking metrics between participants with stroke and healthy controls. Secondary analyses examined associations between task performance and eye-tracking metrics. Exploratory analyses were conducted within the stroke group using K-means clustering based on integrated eye-tracking features to identify potential eye-movement phenotypes. The resulting phenotypes were compared in terms of attention/executive-control performance, global cognition, neurological severity, functional independence, motor function, emotional status, and lesion-related characteristics. Sensitivity analyses and regression models were further used to assess the robustness and clinical relevance of the clustering solution.

Statistical analysis

Sample size estimation

Sample size was justified based on the main analytical objectives, including between-group comparisons of core eye-tracking outcomes, correlation analyses between attention/executive-control measures and eye-tracking metrics, and exploratory eye-movement-based clustering. For the primary between-group comparison, a medium effect size was assumed based on preliminary observations and previous eye-tracking studies in stroke populations. Using a two-sided α level of 0.05, statistical power of 80%, an allocation ratio of 2:1 between the stroke and control groups, and an expected medium effect size of Cohen’s d = 0.60, G*Power 3.1 indicated that at least 67 participants with stroke and 34 healthy controls would be required. To account for an anticipated data loss rate of approximately 10% due to incomplete eye-tracking data, poor data quality, or withdrawal, we planned to recruit approximately 75 participants with stroke and 38 healthy controls.

The planned sample size was also considered adequate for correlation analyses within the stroke group. Under the same α level and power, a sample size of approximately 62 participants would be required to detect a medium correlation coefficient of |r| ≈ 0.35. For the clustering analysis, no universally accepted formal power calculation exists for unsupervised clustering. Therefore, sample adequacy for clustering was evaluated using the number of observations relative to the number of clustering features, and the final stroke sample exceeded the commonly used minimum of approximately 10 observations per feature.

Data preparation and missing data

Statistical analyses were performed using SPSS 26.0 and Python 3.9. Eye-tracking analyses were based on complete feature-level datasets after participant-level quality control. Participants were excluded from the final analysis if any required task- or trial-level eye-tracking feature was missing, incomplete, or abnormal, because cross-task aggregation and clustering required complete data across all six paradigms. No imputation was performed for missing eye-tracking features.

Descriptive statistics and group comparisons

Continuous variables were examined for distributional normality using the Shapiro–Wilk test. Normally distributed variables are presented as mean ± standard deviation, whereas non-normally distributed variables are presented as median and interquartile range. Categorical variables are presented as frequency and percentage.

Between-group comparisons were performed using independent-samples t tests or one-way analysis of variance for normally distributed continuous variables and Mann–Whitney U tests for non-normally distributed variables. Categorical variables were compared using the χ² test or Fisher’s exact test, as appropriate. For group comparisons, full test statistics were reported where applicable, including F values with degrees of freedom, χ² values, exact p values, and corresponding effect sizes. Effect sizes were reported as Cohen’s d for group comparisons of continuous eye-tracking metrics, partial eta squared for ANOVA, and Cramer’s V or odds ratios for categorical comparisons.

Multiple-comparison control

Given the number of behavioral outcomes, eye-tracking metrics across six paradigms, correlation analyses, and post-clustering comparisons, multiple-comparison control was applied to reduce the risk of type I error. The Benjamini–Hochberg false discovery rate procedure was used within predefined families of tests, including: (1) between-group comparisons of attention/executive-control measures; (2) between-group comparisons of eye-tracking metrics; (3) correlations between attention/executive-control measures and eye-tracking metrics; and (4) post-clustering comparisons of attention, clinical, and lesion-related variables. Both uncorrected p values and FDR-adjusted q values were reported where appropriate. Results with q < 0.05 were considered statistically significant after correction.

Correlation analyses

Associations between attention/executive-control measures and eye-tracking metrics were examined separately in the stroke and control groups using Pearson correlation coefficients for normally distributed variables and Spearman rank correlation coefficients for non-normally distributed variables. Correlation coefficients were interpreted according to their magnitude and direction rather than statistical significance alone. FDR correction was applied to address multiple testing. Given the exploratory nature and large number of correlations, the main text presented representative nonredundant correlations that survived correction rather than the full correlation matrix.

Cross-task integration of eye-tracking features

To derive integrated eye-movement features and reduce task-specific scale effects, eye-tracking metrics were standardized within each task using Z-score transformation based on the stroke group. For each participant, standardized values of the same metric were then averaged across the six paradigms to obtain cross-task composite eye-tracking features.

Because task duration was fixed across participants, fixation duration ratio was linearly related to total fixation duration, and mean saccade velocity was linearly related to total saccade distance. To avoid overweighting duplicated information in clustering, the primary K-means clustering analysis used four nonredundant composite eye-tracking features: total fixation duration, first-fixation duration, number of fixations, and total saccade distance.

K-means clustering and cluster validation

K-means clustering was performed within the stroke group to identify exploratory eye-movement phenotypes. The number of clusters was evaluated using the elbow method, silhouette coefficient, Davies–Bouldin index, and Calinski–Harabasz index, together with clinical interpretability. K-means clustering was repeated with multiple random initializations to reduce dependence on starting centroids. The final cluster solution was selected based on convergence of statistical indices and interpretability of the resulting eye-movement profiles.

Cluster stability was evaluated using bootstrap resampling. The clustering procedure was repeated across 1,000 bootstrap samples, and the consistency of cluster assignment was summarized using overall agreement rates and the adjusted Rand index. Sensitivity analyses were performed to examine whether the two-cluster solution remained stable when: (1) the original six integrated eye-tracking features were used; (2) only fixation-related features were used; (3) only saccade-related features were used; and (4) healthy controls were included together with participants with stroke in the clustering procedure. These analyses were conducted to examine whether both fixation- and saccade-related features contributed to the identified phenotypes and whether the stroke-derived clusters were robust to alternative feature sets.

Post-clustering comparisons

After clustering, attention/executive-control performance, clinical characteristics, and lesion-related variables were compared between the identified eye-movement phenotypes. Continuous variables were compared using t tests, analysis of variance, or Mann–Whitney U tests, as appropriate, and categorical variables were compared using χ² tests or Fisher’s exact tests. Multiple-comparison correction was applied using the FDR procedure within the family of post-clustering comparisons.

Regression analyses

To examine the clinical relevance of eye-tracking measures beyond conventional attention/executive-control testing, exploratory regression analyses were performed. First, logistic regression was used to examine factors associated with membership in the more impaired eye-movement phenotype. Given the relatively small size of this phenotype group, predictors were first examined in univariable models, followed by parsimonious multivariable models including clinically relevant variables to avoid overfitting. Odds ratios with 95% confidence intervals were reported.

Second, hierarchical linear regression was used to examine whether eye-movement phenotype explained additional variance in functional independence, as measured by the Barthel Index. In the first model, demographic and clinical covariates were entered, including age, sex, years of education, time from stroke onset to assessment, and NIHSS score. In the second model, an attention/executive-control composite score was added. This composite score was calculated by averaging the Z-standardized scores of the Digit Span Test, Cancellation 1 hit rate, Cancellation 2 hit rate, SDMT, and PASAT, with all component scores oriented so that higher values indicated better performance. In the third model, eye-movement phenotype was added. The incremental contribution of eye-movement phenotype was evaluated using changes in R², adjusted R², F change, and p values.

Statistical significance

All statistical tests were two-sided. The nominal significance level was set at α = 0.05. For analyses involving multiple comparisons, FDR-adjusted q < 0.05 was considered statistically significant. Findings from exploratory clustering, sensitivity analyses, and regression models were interpreted cautiously and considered hypothesis-generating.

Results

Participant flow, data completeness, and baseline characteristics

The final analytical sample included 72 participants with stroke and 36 healthy controls. Demographic characteristics of the two groups are summarized in Table 1. The mean age was 64.3 ± 10.2 years in the stroke group and 64.1 ± 8.3 years in the control group, F(1,106) = 0.010, p = 0.921, partial η² = 0.000. There were 40 males in the stroke group, 55.6%, and 22 males in the control group, 61.1%, χ²(1) = 0.118, p = 0.731, Cramer’s V = 0.033. Years of education were 9.4 ± 3.5 in the stroke group and 9.6 ± 2.8 in the control group, F(1,106) = 0.084, p = 0.773, partial η² = 0.001.

Table 1.

Demographic characteristics of the final analytical sample

Variable	Stroke (n = 72)	Control (n = 36)	Test statistic	P value	Effect size
Age, years	64.3 ± 10.2	64.1 ± 8.3	F(1,106) = 0.010	0.921	0.000
Male, n (%)	40 (55.6%)	22 (61.1%)	χ²(1) = 0.118	0.731	0.033
Education, years	9.4 ± 3.5	9.6 ± 2.8	F(1,106) = 0.084	0.773	0.001

Stroke-related clinical and lesion characteristics are described in the text

Values are presented as mean ± SD, median [IQR], or n (%), as appropriate. Exact p values are shown unless p < 0.001. For continuous group comparisons, effect sizes are partial eta squared; for categorical comparisons, effect sizes are Cramer’s V. Clinical variables are presented for the stroke group only where not applicable to healthy controls. Lesion-location variables were not mutually exclusive because some participants had bilateral or multifocal lesions. Left hemisphere involvement and right hemisphere involvement indicate whether the lesion involved the corresponding hemisphere and were not mutually exclusive

The median time from stroke onset to assessment was 52 days, IQR 40–61 days. Hemorrhagic stroke was present in 9 participants, 12.5%. Left hemisphere involvement was present in 35 participants, 48.6%, right hemisphere involvement in 29 participants, 40.3%, and bilateral or multifocal involvement in 17 participants, 23.6%. Frontal involvement was recorded in 28 participants, 38.9%, and brainstem involvement in 9 participants, 12.5%.

Clinical scale scores in the stroke group were as follows: MMSE 24.5 ± 3.2, NIHSS 6.8 ± 3.1, Barthel Index 32.4 ± 15.2, Fugl–Meyer Assessment 32.6 ± 18.4, SAS 42.1 ± 10.8, and SDS 43.3 ± 12.5.

Participants with stroke showed poorer performance on tasks of attention and executive control

Performance on tasks of attention and executive control is summarized in Table 2. Digit Span Test scores were 8.49 ± 2.12 in the stroke group and 11.89 ± 1.60 in the control group, F(1,106) = 72.279, p < 0.001, q < 0.001, partial η² = 0.405.

Table 2.

Performance on tasks of attention and executive control in participants with stroke and healthy controls

Test	Stroke (n = 72)	Control (n = 36)	Test statistic	p	q	Effect size
DST	8.49 ± 2.12	11.89 ± 1.60	F(1,106) = 72.279	< 0.001	< 0.001	0.405
Cancellation 1 hit rate (%)	98.93 ± 0.47	99.58 ± 0.12	F(1,106) = 66.647	< 0.001	< 0.001	0.386
Cancellation 2 hit rate (%)	98.70 ± 0.54	99.26 ± 0.13	F(1,106) = 38.825	< 0.001	< 0.001	0.268
SDMT	16.81 ± 5.96	38.99 ± 2.74	F(1,106) = 449.562	< 0.001	< 0.001	0.809
PASAT	19.92 ± 7.09	42.06 ± 2.71	F(1,106) = 325.781	< 0.001	< 0.001	0.755

Cancellation error rates were transformed into cancellation hit rates, calculated as 100% minus the corresponding error rate, so that higher values consistently indicated better performance. q values were obtained using the Benjamini–Hochberg FDR procedure within this family of comparisons

For the cancellation tasks, Cancellation 1 hit rate was 98.93 ± 0.47% in the stroke group and 99.58 ± 0.12% in the control group, F(1,106) = 66.647, p < 0.001, q < 0.001, partial η² = 0.386. Cancellation 2 hit rate was 98.70 ± 0.54% in the stroke group and 99.26 ± 0.13% in the control group, F(1,106) = 38.825, p < 0.001, q < 0.001, partial η² = 0.268.

SDMT scores were 16.81 ± 5.96 in the stroke group and 38.99 ± 2.74 in the control group, F(1,106) = 449.562, p < 0.001, q < 0.001, partial η² = 0.809. PASAT scores were 19.92 ± 7.09 in the stroke group and 42.06 ± 2.71 in the control group, F(1,106) = 325.781, p < 0.001, q < 0.001, partial η² = 0.755.

Group differences in eye-tracking metrics across six paradigms

Group comparisons of nonredundant eye-tracking metrics are summarized in Table 3. After FDR correction, participants with stroke showed a consistent pattern of reduced fixation-related indices and increased fixation-count activity across the six paradigms. Total fixation duration was significantly lower in the stroke group than in healthy controls across all six paradigms, with medium effect sizes (Table 3). The number of fixations was significantly higher in the stroke group across all paradigms, with the largest effects observed in the dynamic path, oculomotor, static picture, and antisaccade tasks. By contrast, first-fixation duration showed significant group differences only in the static picture and antisaccade paradigms. Total saccade distance was significantly increased only in the antisaccade task (Table 3).

Table 3.

Group differences in nonredundant eye-tracking metrics across six paradigms

Paradigm	Total fixation duration	First-fixation duration	Number of fixations	Total saccade distance
Static picture	↓ F = 12.566, p = 0.001, q = 0.002, d = − 0.72	↓ F = 8.453, p = 0.004, q = 0.010, d = − 0.59	↑ F = 19.389, p < 0.001, q = 0.001, d = 0.90	F = 1.754, p = 0.188, q = 0.221, d = 0.27, ns
Dynamic path	↓ F = 13.747, p < 0.001, q = 0.002, d = − 0.76	F = 3.630, p = 0.059, q = 0.094, d = − 0.39, ns	↑ F = 23.400, p < 0.001, q < 0.001, d = 0.99	F = 3.733, p = 0.056, q = 0.094, d = 0.39, ns
Oculomotor task	↓ F = 15.771, p < 0.001, q = 0.001, d = − 0.81	F = 2.999, p = 0.086, q = 0.122, d = − 0.35, ns	↑ F = 17.521, p < 0.001, q = 0.001, d = 0.85	F = 1.213, p = 0.273, q = 0.303, d = 0.22, ns
Dynamic video	↓ F = 6.073, p = 0.015, q = 0.029, d = − 0.50	F = 2.359, p = 0.128, q = 0.174, d = − 0.31, ns	↑ F = 8.428, p = 0.004, q = 0.010, d = 0.59	F = 0.336, p = 0.564, q = 0.592, d = 0.12, ns
Prosaccade	↓ F = 9.876, p = 0.002, q = 0.005, d = − 0.64	F = 3.228, p = 0.075, q = 0.111, d = − 0.37, ns	↑ F = 14.494, p < 0.001, q = 0.001, d = 0.78	F = 2.071, p = 0.153, q = 0.190, d = 0.29, ns
Antisaccade	↓ F = 13.338, p < 0.001, q = 0.002, d = − 0.75	↓ F = 6.410, p = 0.013, q = 0.026, d = − 0.52	↑ F = 17.280, p < 0.001, q = 0.001, d = 0.85	↑ F = 11.334, p = 0.001, q = 0.003, d = 0.69

Values represent group comparisons between participants with stroke and healthy controls. Arrows indicate the direction of the difference in the stroke group relative to healthy controls. ↓ indicates lower values in the stroke group; ↑ indicates higher values in the stroke group. q values were calculated using Benjamini–Hochberg false discovery rate correction. Cohen’s d is reported as the effect size. Only nonredundant core eye-tracking metrics are displayed. Fixation duration ratio and mean saccade velocity were not shown because they were mathematically related to total fixation duration and total saccade distance under fixed task durations. ns = not significant after FDR correction

These findings indicate that stroke was associated with shorter overall fixation time and more frequent shifts in fixation location across tasks, suggesting less stable visual engagement and a more fragmented visual sampling pattern. Differences in total saccade distance were less consistent and became most evident under inhibitory-control demands, particularly during the antisaccade paradigm.

Associations between attention/executive-control performance and eye-tracking metrics

Correlation analyses were conducted as secondary exploratory analyses to examine whether eye-movement behavior was associated with attention/executive-control performance in participants with stroke. To reduce the risk of type I error, Benjamini–Hochberg false discovery rate correction was applied. Given the large number of possible correlations, the full correlation matrix was not presented in the main text. Instead, Table 4 provides a concise representative summary by displaying, for each attention/executive-control measure, the strongest positive, an additional strong positive, the strongest negative, and an additional strong negative nonredundant correlation that remained significant after correction.

Table 4.

Representative FDR-significant correlations between attention/executive-control performance and eye-tracking metrics in participants with stroke

Attention / executive-control measure	Association type	Eye-tracking metric	r	p	q
DST	Strongest positive	Static picture | Total fixation duration	0.663	< 0.001	< 0.001
DST	Additional positive	Dynamic video | Total fixation duration	0.617	< 0.001	< 0.001
DST	Strongest negative	Dynamic video | Number of fixations	−0.592	< 0.001	< 0.001
DST	Additional negative	Static picture | Number of fixations	−0.571	< 0.001	< 0.001
Cancellation 1 hit rate (%)	Strongest positive	Static picture | Total fixation duration	0.662	< 0.001	< 0.001
Cancellation 1 hit rate (%)	Additional positive	Dynamic video | Total fixation duration	0.601	< 0.001	< 0.001
Cancellation 1 hit rate (%)	Strongest negative	Dynamic video | Number of fixations	−0.543	< 0.001	< 0.001
Cancellation 1 hit rate (%)	Additional negative	Dynamic video | Total saccade distance	−0.506	< 0.001	< 0.001
Cancellation 2 hit rate (%)	Strongest positive	Static picture | Total fixation duration	0.684	< 0.001	< 0.001
Cancellation 2 hit rate (%)	Additional positive	Dynamic video | Total fixation duration	0.628	< 0.001	< 0.001
Cancellation 2 hit rate (%)	Strongest negative	Dynamic video | Number of fixations	−0.556	< 0.001	< 0.001
Cancellation 2 hit rate (%)	Additional negative	Dynamic video | Total saccade distance	−0.545	< 0.001	< 0.001
SDMT	Strongest positive	Static picture | Total fixation duration	0.687	< 0.001	< 0.001
SDMT	Additional positive	Oculomotor task | Total fixation duration	0.577	< 0.001	< 0.001
SDMT	Strongest negative	Oculomotor task | Total saccade distance	−0.552	< 0.001	< 0.001
SDMT	Additional negative	Dynamic video | Number of fixations	−0.535	< 0.001	< 0.001
PASAT	Strongest positive	Static picture | Total fixation duration	0.673	< 0.001	< 0.001
PASAT	Additional positive	Oculomotor task | Total fixation duration	0.600	< 0.001	< 0.001
PASAT	Strongest negative	Oculomotor task | Total saccade distance	−0.560	< 0.001	< 0.001
PASAT	Additional negative	Dynamic video | Number of fixations	−0.542	< 0.001	< 0.001

Values are Pearson or Spearman correlation coefficients, as appropriate. Only representative correlations that survived Benjamini–Hochberg false discovery rate correction are shown. For each attention/executive-control measure, the strongest positive, an additional strong positive, the strongest negative, and an additional strong negative nonredundant correlation are displayed. q values indicate FDR-adjusted p values. Cancellation hit rate was calculated as 100% minus the error rate, so that higher values consistently indicate better performance. Mathematically redundant derived metrics, including fixation duration ratio and mean saccade velocity, were not displayed because they were linearly related to total fixation duration and total saccade distance under fixed task durations. DST = Digit Span Test; SDMT = Symbol Digit Modalities Test; PASAT = Paced Auditory Serial Addition Test

Overall, better attention/executive-control performance was consistently associated with more stable fixation behavior. Static-picture total fixation duration showed robust positive correlations with all task-performance measures, including DST, r = 0.663, q < 0.001; Cancellation 1 hit rate, r = 0.662, q < 0.001; Cancellation 2 hit rate, r = 0.684, q < 0.001; SDMT, r = 0.687, q < 0.001; and PASAT, r = 0.673, q < 0.001. Additional positive associations were observed for total fixation duration during dynamic video viewing or the oculomotor task, with correlation coefficients ranging from r = 0.577 to r = 0.628.

In contrast, poorer task performance was associated with greater saccade-related activity and a more fragmented visual sampling pattern. Dynamic-video number of fixations was negatively correlated with DST, r = − 0.592, q < 0.001; Cancellation 1 hit rate, r = − 0.543, q < 0.001; Cancellation 2 hit rate, r = − 0.556, q < 0.001; and PASAT, r = − 0.542, q < 0.001. Total saccade distance during the oculomotor task was negatively associated with SDMT, r = − 0.552, q < 0.001, and PASAT, r = − 0.560, q < 0.001.

In the control group, correlation coefficients were generally smaller and less consistent across paradigms. After FDR correction, fewer associations remained significant, and no coherent cross-task pattern comparable to that observed in the stroke group was identified. Therefore, the representative table focuses on the stroke group, where eye-movement behavior showed clearer associations with task performance.

These representative corrected correlations suggest that better attention/executive-control performance was associated with longer and more stable fixation, whereas poorer performance was associated with more frequent shifts in fixation location and greater saccade-related activity. Because these analyses were cross-sectional and exploratory, the findings were interpreted as association patterns rather than evidence of causal relationships.

K-means clustering identified two exploratory eye-movement phenotypes in the stroke group

To characterize heterogeneity in post-stroke eye-movement behavior, K-means clustering was performed within the stroke group using four nonredundant integrated eye-tracking features: total fixation duration, first-fixation duration, number of fixations, and total saccade distance. These features were selected to avoid overweighting mathematically redundant variables derived from fixed task durations.

Model-selection indices supported a two-cluster solution. The primary nonredundant-feature model showed a silhouette coefficient of 0.573, Davies–Bouldin index of 0.635, and Calinski–Harabasz index of 134.941. The resulting clusters showed clearly contrasting eye-movement profiles. One cluster was characterized by relatively higher fixation-related composite scores and lower saccade-related composite scores and was labeled the High-Fixation/Low-Saccade phenotype, HF-LS, n = 48. The other cluster was characterized by lower fixation-related scores and higher saccade-related scores and was labeled the Low-Fixation/High-Saccade phenotype, LF-HS, n = 24.

The two phenotypes differed substantially across all four clustering features (Table 5). Compared with the HF-LS phenotype, the LF-HS phenotype showed lower total fixation duration Z scores, − 1.116 ± 0.690 versus 0.558 ± 0.329, F(1,70) = 195.632, p < 0.001, q < 0.001, partial η² = 0.736, and lower first-fixation duration Z scores, − 0.849 ± 0.226 versus 0.424 ± 0.601, F(1,70) = 99.913, p < 0.001, q < 0.001, partial η² = 0.588. In contrast, the LF-HS phenotype showed higher number-of-fixations Z scores, 1.116 ± 0.362 versus − 0.558 ± 0.417, F(1,70) = 280.317, p < 0.001, q < 0.001, partial η² = 0.800, and higher total saccade distance Z scores, 0.751 ± 0.896 versus − 0.376 ± 0.376, F(1,70) = 56.710, p < 0.001, q < 0.001, partial η² = 0.448.

Table 5.

Integrated nonredundant eye-tracking features by eye-movement phenotype in the stroke group

Feature	HF-LS n = 48	LF-HS n = 24	F(1,70)	p	q	Partial η²
Total fixation duration Z	0.558 ± 0.329	-1.116 ± 0.690	F(1,70) = 195.632	< 0.001	< 0.001	0.736
First-fixation duration Z	0.424 ± 0.601	-0.849 ± 0.226	F(1,70) = 99.913	< 0.001	< 0.001	0.588
Number of fixations Z	-0.558 ± 0.417	1.116 ± 0.362	F(1,70) = 280.317	< 0.001	< 0.001	0.800
Total saccade distance Z	-0.376 ± 0.376	0.751 ± 0.896	F(1,70) = 56.710	< 0.001	< 0.001	0.448

HF-LS=High-Fixation/Low-Saccade phenotype; LF-HS = Low-Fixation/High-Saccade phenotype. Features were Z-standardized within each task and then averaged across the six paradigms. q values were obtained using FDR correction

Thus, the HF-LS phenotype represented a relatively fixation-stable and lower-saccade profile, whereas the LF-HS phenotype represented a more fragmented, saccade-dominant eye-movement profile.

Cluster stability and sensitivity analyses

Cluster stability and sensitivity analyses are summarized in Table 6. The primary nonredundant-feature solution identified two clusters with acceptable internal separation, silhouette = 0.573, Davies–Bouldin index = 0.635, and Calinski–Harabasz index = 134.941.

Table 6.

Cluster model selection, stability, and sensitivity analyses

Analysis	K	Silhouette	DBI	CH index	Agreement with primary (%)	ARI / stability metric	Interpretation
Primary nonredundant features	2	0.573	0.635	134.941	100	1.000	Main solution
Six-feature model	2	0.586	0.631	134.959	95.833	0.835	Consistent
Fixation-only model	2	0.590	0.572	148.161	98.611	0.944	Partial/consistent
Saccade-only model	2	0.591	0.630	133.269	95.833	0.836	Partial/consistent
Stroke + control model	2	0.592	0.633	174.843	100	1.000	Sensitivity analysis
Bootstrap resampling	—	—	—	—	—	Mean ARI = 0.931 Median ARI = 1.000 IQR 0.835-1.000	Stability check

For the stroke + control model, agreement and ARI were calculated only for stroke participants relative to the primary stroke-only solution

Sensitivity analyses yielded highly consistent results. When the original six-feature model was used, including fixation duration ratio and mean saccade velocity, the two-cluster structure was largely preserved, with 95.833% agreement with the primary nonredundant-feature solution and adjusted Rand index = 0.835. The fixation-only model showed 98.611% agreement with the primary solution and adjusted Rand index = 0.944, while the saccade-only model showed 95.833% agreement and adjusted Rand index = 0.836.

As an additional sensitivity analysis, K-means clustering was repeated after including healthy controls together with participants with stroke. The combined-sample analysis again supported a two-cluster solution, with a silhouette coefficient of 0.592, Davies–Bouldin index of 0.633, and Calinski–Harabasz index of 174.843. When restricted to the stroke participants, the cluster assignments from this combined-sample model were identical to those from the primary stroke-only model, with 100% agreement and an adjusted Rand index of 1.000. This finding suggests that the stroke-derived phenotype structure was not materially altered by including healthy controls in the sensitivity analysis.

Bootstrap resampling further supported the stability of the two-cluster solution. Across 1,000 bootstrap resamples, the mean adjusted Rand index was 0.931, with a median of 1.000 and an interquartile range of 0.835–1.000. Cluster-wise Jaccard similarity was 0.975 for the HF-LS phenotype and 0.948 for the LF-HS phenotype. These results indicate good stability of the clustering solution.

The primary two-cluster solution remained stable across bootstrap resampling, alternative feature sets, and combined clustering of stroke participants and healthy controls. These analyses support the robustness of the HF-LS and LF-HS phenotypes, although the clustering results should still be interpreted as exploratory and require validation in independent cohorts.

Eye-movement phenotypes differed in attention/executive-control performance and clinical characteristics

The two eye-movement phenotypes differed substantially in attention/executive-control performance, global cognition, neurological severity, functional independence, motor function, emotional symptoms, and lesion characteristics (Table 7). Demographic variables were comparable between the two phenotypes. Age, education, sex distribution, and time from stroke onset to assessment did not differ significantly after FDR correction, suggesting that the observed phenotype differences were not primarily attributable to basic demographic imbalance or post-stroke assessment timing.

Table 7.

Comparisons of attention/executive-control performance, clinical characteristics, and lesion features between eye-movement phenotypes in the stroke group

Variable	HF-LS n = 48	LF-HS n = 24	Test statistic	p	q	Effect size
Age, years	63.41 ± 10.92	66.08 ± 8.53	F(1,70) = 1.100	0.298	0.386	η²p = 0.015
Education, years	9.55 ± 3.37	9.11 ± 3.81	F(1,70) = 0.245	0.622	0.685	η²p = 0.003
Time from stroke onset, days	50.12 ± 16.42	53.50 ± 19.87	F(1,70) = 0.587	0.446	0.546	η²p = 0.008
MMSE (0–30)	25.25 ± 3.90	21.67 ± 3.52	F(1,70) = 14.376	< 0.001	< 0.001	η²p = 0.170
NIHSS (0–42)	5.81 ± 2.50	9.20 ± 3.01	F(1,70) = 25.756	< 0.001	< 0.001	η²p = 0.269
Barthel Index (0–100)	38.50 ± 13.20	18.50 ± 10.10	F(1,70) = 42.563	< 0.001	< 0.001	η²p = 0.378
Fugl-Meyer Assessment (0–100)	38.89 ± 16.10	18.40 ± 12.50	F(1,70) = 29.834	< 0.001	< 0.001	η²p = 0.299
SAS (standard score)	38.90 ± 8.51	51.50 ± 9.00	F(1,70) = 33.796	< 0.001	< 0.001	η²p = 0.326
SDS (standard score)	39.80 ± 9.80	53.80 ± 10.50	F(1,70) = 31.153	< 0.001	< 0.001	η²p = 0.308
DST	9.17 ± 1.92	7.12 ± 1.85	F(1,70) = 18.578	< 0.001	< 0.001	η²p = 0.210
Cancellation 1 hit rate (%)	99.06 ± 0.41	98.65 ± 0.46	F(1,70) = 14.427	< 0.001	< 0.001	η²p = 0.171
Cancellation 2 hit rate (%)	98.86 ± 0.44	98.37 ± 0.57	F(1,70) = 15.971	< 0.001	< 0.001	η²p = 0.186
SDMT	18.46 ± 5.60	13.50 ± 5.34	F(1,70) = 12.940	< 0.001	< 0.001	η²p = 0.156
PASAT	22.02 ± 6.19	15.71 ± 7.02	F(1,70) = 15.222	< 0.001	< 0.001	η²p = 0.179
Male, n (%)	24 (50.0)	16 (66.7)	χ²(1) = 1.188	0.276	0.379	Cramer’s V = 0.128
Hemorrhagic stroke, n (%)	6 (12.5)	3 (12.5)	χ²(1) = 0.000	1	1	Cramer’s V = 0.000
Left hemisphere involvement, n (%)	22 (45.8)	13 (54.2)	χ²(1) = 0.174	0.677	0.709	Cramer’s V = 0.049
Right hemisphere involvement, n (%)	22 (45.8)	7 (29.2)	χ²(1) = 1.220	0.269	0.379	Cramer’s V = 0.130
Frontal involvement, n (%)	13 (27.1)	15 (62.5)	χ²(1) = 7.020	0.008	0.014	Cramer’s V = 0.312
Brainstem involvement, n (%)	4 (8.3)	5 (20.8)	χ²(1) = 1.286	0.257	0.379	Cramer’s V = 0.134
Bilateral or multifocal involvement, n (%)	6 (12.5)	11 (45.8)	χ²(1) = 8.095	0.004	0.008	Cramer’s V = 0.335

Higher cancellation hit rate indicates better performance. q values were calculated using FDR correction within the post-clustering comparison family. HF-LS = High-Fixation/Low-Saccade phenotype; LF-HS = Low-Fixation/High-Saccade phenotype

Compared with the HF-LS phenotype, the LF-HS phenotype showed poorer performance across all attention/executive-control tasks. DST scores were lower in the LF-HS group, 7.12 ± 1.85 versus 9.17 ± 1.92, F(1,70) = 18.578, p < 0.001, q < 0.001, partial η² = 0.210. Cancellation 1 hit rate was also lower, 98.65 ± 0.46% versus 99.06 ± 0.41%, F(1,70) = 14.427, p < 0.001, q < 0.001, partial η² = 0.171, as was Cancellation 2 hit rate, 98.37 ± 0.57% versus 98.86 ± 0.44%, F(1,70) = 15.971, p < 0.001, q < 0.001, partial η² = 0.186. The LF-HS phenotype also showed lower SDMT scores, 13.50 ± 5.34 versus 18.46 ± 5.60, F(1,70) = 12.940, p < 0.001, q < 0.001, partial η² = 0.156, and lower PASAT scores, 15.71 ± 7.02 versus 22.02 ± 6.19, F(1,70) = 15.222, p < 0.001, q < 0.001, partial η² = 0.179.

The LF-HS phenotype also showed a more severe clinical profile. MMSE scores were lower in LF-HS than in HF-LS, 21.67 ± 3.52 versus 25.25 ± 3.90, F(1,70) = 14.376, p < 0.001, q < 0.001, partial η² = 0.170. NIHSS scores were higher, 9.20 ± 3.01 versus 5.81 ± 2.50, F(1,70) = 25.756, p < 0.001, q < 0.001, partial η² = 0.269. Functional independence and motor function were lower in LF-HS, as reflected by lower Barthel Index scores, 18.50 ± 10.10 versus 38.50 ± 13.20, F(1,70) = 42.563, p < 0.001, q < 0.001, partial η² = 0.378, and lower Fugl–Meyer Assessment scores, 18.40 ± 12.50 versus 38.89 ± 16.10, F(1,70) = 29.834, p < 0.001, q < 0.001, partial η² = 0.299.

Emotional symptom burden was also greater in the LF-HS phenotype. SAS scores were higher in LF-HS than in HF-LS, 51.50 ± 9.00 versus 38.90 ± 8.51, F(1,70) = 33.796, p < 0.001, q < 0.001, partial η² = 0.326. SDS scores were similarly higher, 53.80 ± 10.50 versus 39.80 ± 9.80, F(1,70) = 31.153, p < 0.001, q < 0.001, partial η² = 0.308.

Lesion-related comparisons showed that frontal involvement and Bilateral or multifocal involvement were more frequent in the LF-HS phenotype. Frontal involvement was present in 62.5% of LF-HS participants compared with 27.1% of HF-LS participants, χ²(1) = 7.020, p = 0.008, q = 0.014, Cramer’s V = 0.312. Multiple involvement was also more common in LF-HS, 45.8% versus 12.5%, χ²(1) = 8.095, p = 0.004, q = 0.008, Cramer’s V = 0.335. In contrast, stroke type, left/right hemispheric involvement, and brainstem involvement did not differ significantly between phenotypes after correction.

The LF-HS phenotype was characterized by a more impaired eye-movement pattern, poorer attention/executive-control performance, worse global cognition, greater neurological and motor impairment, lower functional independence, higher emotional symptom burden, and more frequent frontal involvement and multiple lesions. These findings support the clinical interpretability of the eye-movement phenotypes, while the exploratory and cross-sectional nature of the clustering analysis warrants cautious interpretation.

Adjusted and regression analyses supported the clinical relevance of eye-tracking phenotypes

To examine whether eye-movement phenotypes were associated with functional independence beyond conventional clinical and behavioral measures, exploratory regression analyses were performed (Table 8). In hierarchical linear regression predicting Barthel Index scores, the base model including age, sex, education, time from stroke onset, and NIHSS explained 11.4% of the variance, R² = 0.114, adjusted R² = 0.047. Adding the attention/executive-control composite score improved model fit, increasing explained variance by 13.5%, ΔR² = 0.135, F change(1,65) = 11.658, p = 0.001, with total R² increasing to 0.249 and adjusted R² to 0.180.

Table 8.

Hierarchical regression models explaining variance in Barthel Index in participants with stroke

Model	Predictors	R²	Adjusted R²	ΔR²	F change	p for ΔR²
Model 1	Age, sex, education, time from stroke onset, NIHSS	0.114	0.047	—	—	—
Model 2	Model 1 + attention/executive-control composite score	0.249	0.180	0.135	F(1,65) = 11.658	0.001
Model 3	Model 2 + eye-movement phenotype (LF-HS vs. HF-LS)	0.446	0.385	0.197	F(1,64) = 22.714	< 0.001

The attention/executive-control composite score was defined as the mean of the Z-standardized scores of the Digit Span Test, Cancellation 1 hit rate, Cancellation 2 hit rate, SDMT, and PASAT. All component scores were oriented so that higher values indicated better performance

When eye-movement phenotype was added in the final model, explanatory power increased further, ΔR² = 0.197, F change(1,64) = 22.714, p < 0.001. The final model explained 44.6% of the variance in Barthel Index scores, R² = 0.446, adjusted R² = 0.385. This indicates that eye-movement phenotype provided additional explanatory information for functional independence beyond demographic variables, neurological severity, post-stroke timing, and conventional attention/executive-control performance.

Logistic regression analyses further examined variables associated with membership in the more impaired LF-HS phenotype. In univariable analyses, higher NIHSS scores, lower MMSE scores, lower attention/executive-control composite performance, frontal involvement, and multiple lesions were associated with LF-HS phenotype membership. A parsimonious exploratory multivariable model was then constructed to reduce overfitting while retaining clinically relevant variables. After mutual adjustment, higher NIHSS scores, lower attention/executive-control composite performance, frontal involvement, and multiple lesions remained associated with LF-HS membership. Specifically, NIHSS was associated with higher odds of LF-HS membership, OR = 1.977, 95% CI 1.243–3.147, p = 0.004, whereas better attention/executive-control composite performance was associated with lower odds, OR = 0.212, 95% CI 0.073–0.620, p = 0.005. Frontal involvement, OR = 5.394, 95% CI 1.078–26.982, p = 0.040, and multiple lesions, OR = 7.617, 95% CI 1.214–47.769, p = 0.030, also remained associated with LF-HS membership (Table 9).

Table 9.

Univariable and exploratory multivariable logistic regression analyses of variables associated with LF-HS phenotype membership

Model	Predictor	OR	95% CI	p
Univariable	NIHSS	1.760	1.284–2.414	< 0.001
Univariable	MMSE	0.785	0.674–0.914	0.002
Univariable	Attention/executive-control composite score	0.272	0.129–0.572	< 0.001
Univariable	Frontal involvement	4.487	1.581–12.735	0.005
Univariable	Bilateral or multifocal involvement	5.923	1.832–19.146	0.003
Multivariable	NIHSS	1.977	1.243–3.147	0.004
Multivariable	Attention/executive-control composite score	0.212	0.073–0.620	0.005
Multivariable	Frontal involvement	5.394	1.078–26.982	0.040
Multivariable	Bilateral or multifocal involvement	7.617	1.214–47.769	0.030

Odds ratios in the univariable analysis were obtained from separate logistic regression models. The exploratory multivariable model included NIHSS, attention/executive-control composite score, frontal involvement, and bilateral or multifocal involvement to reduce overfitting given the limited number of LF-HS cases. LF-HS = Low-Fixation/High-Saccade phenotype; OR = odds ratio; CI = confidence interval

These regression results suggest that the LF-HS phenotype was linked to clinically more severe impairment and may provide additional information relevant to functional independence. However, because the analyses were exploratory, the LF-HS group was relatively small, and the lesion-related confidence intervals were wide, these findings should be interpreted as hypothesis-generating rather than as evidence of predictive validity.

Discussion

Overview of principal findings

This single-center exploratory case-control study combined tasks of attention and executive control with a six-paradigm eye-tracking battery in participants with stroke and healthy controls. The main objective was to determine whether integrated eye-tracking features could identify clinically interpretable eye-movement heterogeneity after stroke and whether these phenotypes were associated with attention/executive-control performance and clinical characteristics. Several findings emerged.

First, participants with stroke showed lower performance across tasks involving attention span, cancellation-based task engagement and target selection, processing speed, cognitive flexibility, working memory, and divided attention. Second, group comparisons of nonredundant eye-tracking metrics showed a consistent pattern of shorter total fixation duration and increased number of fixations across the six paradigms, whereas total saccade distance differed most clearly during the antisaccade task. Third, corrected exploratory correlation analyses showed that, in the stroke group, better task performance was generally associated with longer fixation duration, while lower task performance was associated with increased fixation counts or greater saccade-related activity. Fourth, K-means clustering based on four nonredundant integrated eye-tracking features identified two eye-movement phenotypes: a High-Fixation/Low-Saccade phenotype and a Low-Fixation/High-Saccade phenotype. The LF-HS phenotype showed poorer attention/executive-control performance, lower global cognition, greater neurological impairment, lower functional independence, poorer motor function, higher emotional symptom burden, and more frequent frontal involvement and bilateral or multifocal involvement. Finally, exploratory regression analyses suggested that eye-movement phenotype explained additional variance in Barthel Index scores beyond demographic variables, post-stroke timing, neurological severity, and conventional attention/executive-control performance.

These findings suggest that integrated eye-tracking features may capture clinically relevant heterogeneity in post-stroke cognitive and oculomotor function. However, because the study was cross-sectional and exploratory, the identified phenotypes should be interpreted as preliminary eye-movement profiles rather than definitive clinical categories.

Attention and executive-control performance after stroke

The behavioral results showed lower performance in the stroke group across all five tasks of attention and executive control. This pattern is consistent with the clinical observation that post-stroke cognitive impairment often affects multiple interacting cognitive operations [40]. Attention and executive control are not isolated processes; they support goal maintenance, information selection, response monitoring, working memory, processing speed, and the flexible coordination of cognitive and motor behavior during rehabilitation [41].

The present task battery should therefore be interpreted as assessing task performance across attention- and executive-control-related operations rather than as providing pure measures of separate cognitive domains. Digit Span involves attention span and working memory [42]; SDMT depends on processing speed, visual scanning, associative learning, and cognitive flexibility [43]; PASAT requires divided attention, auditory processing, working memory updating, and continuous monitoring [44]; and the cancellation tasks involve sustained task engagement, target selection, rule maintenance, and response control [45]. This conceptual framing is important because it avoids overassigning each test to a single cognitive construct and better reflects the overlapping cognitive demands of clinical neuropsychological tasks.

The behavioral findings also provide an external context for interpreting eye-tracking results. Conventional tests capture outcome-level performance, whereas eye tracking captures process-level aspects of visual sampling and gaze control. The relationship between these two levels is clinically meaningful, but they should not be treated as identical. Eye tracking does not simply replace neuropsychological testing; rather, it may help explain how patients visually engage with tasks, maintain fixation, shift gaze, and regulate saccadic activity during performance.

Eye-movement abnormalities across paradigms

Across the six eye-tracking paradigms, participants with stroke showed shorter total fixation duration and a higher number of fixations. This pattern indicates altered visual sampling after stroke. Shorter fixation duration may reflect reduced stability of visual engagement, insufficient time for information extraction, or difficulty sustaining attention to task-relevant stimuli. A higher number of fixations, in this context, does not necessarily indicate more efficient exploration. When combined with shorter fixation duration and poorer behavioral performance, it is more plausibly interpreted as a fragmented sampling pattern.

The consistency of total fixation-duration differences across paradigms suggests that fixation-related measures may be sensitive to generalized alterations in visual engagement after stroke. By contrast, total saccade distance showed a more task-specific pattern and differed most clearly during the antisaccade task. This is clinically plausible because antisaccade performance requires suppression of the reflexive response toward a target and generation of a voluntary gaze shift in the opposite direction. Such tasks place greater demands on inhibitory control, goal maintenance, and voluntary oculomotor regulation. Therefore, saccade-related abnormalities may become more evident when patients are required to inhibit automatic gaze responses and implement an internally guided response rule.

These findings should not be interpreted as reflecting attention impairment alone. Eye movements depend on distributed oculomotor and cognitive-control systems, including frontal eye fields, supplementary eye fields, dorsolateral prefrontal cortex, posterior parietal regions, basal ganglia, superior colliculus, thalamus, cerebellar circuits, and brainstem ocular motor pathways [46–48]. Stroke affecting these regions or their connections may directly impair fixation stability, saccade initiation, saccade suppression, or visually guided tracking [49]. Thus, the observed eye-movement abnormalities likely reflect the combined influence of impaired attention/executive control and disruption of oculomotor networks.

Task demands and interpretation of correlation patterns

The six eye-tracking paradigms differed in their cognitive and oculomotor demands. Static picture viewing and dynamic video viewing provided relatively naturalistic viewing conditions. Dynamic path tracking and the oculomotor task required sustained target following. Prosaccade and antisaccade tasks involved explicit gaze shifting, and the antisaccade task added inhibitory-control demands. This range of task demands helped reveal that fixation-related differences were broadly present across paradigms, whereas saccade-distance differences were more apparent under inhibitory-control conditions.

The exploratory correlation analyses provided converging evidence. In the stroke group, better attention/executive-control performance was associated with longer total fixation duration, whereas lower task performance was associated with increased fixation counts or greater saccade-related activity. These associations suggest that patients with better behavioral performance tended to show more stable fixation patterns, while those with poorer performance tended to show more fragmented gaze behavior. In the control group, correlations were generally smaller and less consistent after correction, suggesting that gaze behavior and task performance may be more tightly coupled in the post-stroke state.

These correlations should be interpreted cautiously. They do not establish causal relationships between eye movements and cognitive performance. Poorer task performance may contribute to unstable gaze behavior, but oculomotor impairment, lesion burden, visual deficits, fatigue, mood symptoms, and general neurological severity may also influence both behavioral and eye-tracking measures. The correlation findings are therefore best understood as describing association patterns that support the broader phenotyping analysis rather than as standalone evidence for mechanism.

Eye-movement phenotypes and post-stroke heterogeneity

A central finding of this study was the identification of two eye-movement phenotypes based on integrated nonredundant features. The HF-LS phenotype was characterized by higher fixation-related composite scores and lower saccade-related composite scores, whereas the LF-HS phenotype showed lower fixation-related scores and higher saccade-related scores. This pattern provides a compact representation of individual differences in gaze behavior across the six paradigms.

The use of nonredundant features was methodologically important. Fixation duration ratio and mean saccade velocity were mathematically related to total fixation duration and total saccade distance under fixed task durations. Excluding these derived variables from the primary clustering analysis reduced the risk of overweighting duplicated information. Sensitivity analyses showed that the two-cluster solution was preserved when the original six-feature model was used, when fixation-only or saccade-only features were analyzed, and when healthy controls were included in the clustering procedure. These results suggest that the phenotype structure was not driven solely by a single metric type or by redundant feature construction.

The fixation-only and saccade-only sensitivity analyses both showed high agreement with the primary solution, indicating that fixation and saccade measures each contributed to the phenotype structure. This is clinically relevant because fixation stability and saccadic activity may reflect partially distinct but interacting aspects of visual sampling and oculomotor control [50]. A patient with reduced fixation duration and increased saccadic activity may have difficulty maintaining task-relevant visual engagement, suppressing irrelevant gaze shifts, or efficiently sampling visual information [51]. The combined fixation-saccade profile may therefore be more informative than either dimension alone.

Including healthy controls in the clustering procedure did not change the subtype assignment of stroke participants. This supports the internal stability of the stroke-derived phenotype structure. Nevertheless, this analysis should be considered a sensitivity test rather than external validation. Independent cohorts are needed to determine whether these phenotypes are reproducible across settings, devices, task batteries, and recovery stages.

Clinical interpretation of the LF-HS phenotype

The LF-HS phenotype showed a more impaired clinical profile than the HF-LS phenotype. Participants in the LF-HS group had poorer attention/executive-control task performance, lower MMSE scores, higher NIHSS scores, lower Barthel Index scores, lower Fugl–Meyer Assessment scores, and higher SAS and SDS scores. Frontal involvement and bilateral or multifocal involvement were also more frequent in this phenotype. These findings indicate that the LF-HS pattern was associated with broader cognitive, neurological, functional, emotional, and lesion-related burden.

One possible interpretation is that the LF-HS phenotype reflects impaired top-down control of visual sampling. Reduced fixation-related scores may indicate difficulty sustaining visual engagement, while more frequent shifts in fixation location and increased saccade-related activity may reflect fragmented sampling, inefficient search, or impaired suppression of irrelevant gaze shifts. The higher frequency of frontal involvement in the LF-HS group is compatible with this interpretation, because frontal and frontoparietal systems support voluntary gaze control, inhibitory control, working memory, and executive regulation.

However, frontal involvement should not be interpreted only as evidence of attention or executive-control impairment. Frontal lesions may also directly affect oculomotor control through frontal eye fields, supplementary eye fields, and fronto-basal ganglia-superior colliculus circuits. Therefore, the association between frontal involvement and the LF-HS phenotype may reflect both cognitive-control impairment and direct disruption of saccade generation or suppression. Similarly, bilateral or multifocal involvement may indicate broader network disruption affecting both cognition and eye-movement control. Future studies combining eye tracking with lesion-symptom mapping and structural or functional connectivity analysis will be needed to clarify these mechanisms.

The lower MMSE scores in the LF-HS group also require careful interpretation. Although all participants were able to understand and complete the study procedures, MMSE performance can be influenced by language comprehension, auditory processing, education, and lesion location. In the present study, lesion laterality did not differ significantly between phenotypes after correction, but formal language and comprehension measures were not included. Therefore, language-related or comprehension-related influences on some behavioral tasks cannot be fully excluded, especially for tasks with auditory-verbal demands such as Digit Span and PASAT. This further supports interpreting the behavioral measures as broad indicators of attention/executive-control task performance rather than as isolated cognitive constructs.

Functional relevance of eye-movement phenotypes

The exploratory regression analyses suggested that eye-movement phenotype explained additional variance in Barthel Index scores beyond demographic variables, time from stroke onset, NIHSS, and an attention/executive-control composite score. This finding may be clinically relevant because it suggests that integrated eye-tracking features captured aspects of functional status that were not fully represented by standard clinical severity or conventional cognitive task performance in this sample. Several mechanisms may explain this association. Functional independence after stroke depends not only on motor ability but also on attention, visual exploration, response monitoring, error correction, emotional state, and the ability to coordinate cognition with action [52]. A fragmented eye-movement profile may reflect inefficiencies in visual sampling and cognitive control that interfere with daily activities and rehabilitation participation [53]. The LF-HS phenotype may therefore be interpreted as a preliminary process-level indicator of broader functional vulnerability.

This interpretation should remain cautious. The regression analyses were exploratory and cross-sectional. Barthel Index is influenced by multiple factors, including motor impairment, neurological severity, cognition, mood, comorbidities, environmental support, and rehabilitation exposure. The model was developed within a single dataset and was not prospectively validated. Therefore, the present results suggest potential clinical relevance but do not establish predictive utility. Longitudinal studies are needed to test whether eye-movement phenotypes predict functional recovery, treatment response, or long-term participation outcomes.

Implications for rehabilitation assessment and research

The present findings support the potential role of eye tracking as a process-level adjunct to conventional assessment. Traditional neuropsychological tests provide important information about task outcomes, but they provide limited insight into how patients allocate gaze, maintain fixation, suppress inappropriate saccades, or adapt visual sampling during task execution. Eye-tracking metrics can provide complementary information about these processes.

Eye-movement phenotyping may also help refine research stratification in post-stroke rehabilitation. Patients with similar scores on conventional tests may differ in fixation stability, saccadic activity, and visual sampling strategy. These differences may reflect distinct patterns of network disruption or compensation. Future intervention studies could examine whether patients with different eye-movement phenotypes respond differently to visual scanning training, attention stabilization training, inhibitory-control exercises, or dual-task rehabilitation.

At this stage, the clinical implications should be considered preliminary. The study does not establish treatment algorithms or clinically actionable cutoffs. It also does not test whether eye-tracking-guided rehabilitation improves outcomes. Before clinical deployment, the phenotypes need to be validated across centers, recovery stages, task protocols, and eye-tracking systems. Their longitudinal stability and responsiveness to intervention also require further study.

Limitations and future directions

Several limitations should be acknowledged. First, this was a single-center study with a modest sample size. Although sensitivity analyses supported the two-cluster solution, the LF-HS group remained relatively small. Replication in larger independent cohorts is needed to confirm the reproducibility and generalizability of the identified phenotypes.

Second, the study was cross-sectional. Although time from stroke onset to assessment was recorded and did not differ significantly between phenotypes, the design cannot determine whether the eye-movement profiles are stable traits, stage-dependent phenomena, or markers of recovery trajectory. Longitudinal follow-up is needed to examine whether patients transition between phenotypes and whether baseline phenotype predicts recovery or treatment response.

Third, eye-tracking analyses were based on structured feature-level outputs rather than raw gaze-coordinate trajectories. This approach reflects the clinical workflow of the device and supports standardized feature extraction, but it limits the ability to examine sample-level artifacts, saccade latency, antisaccade error direction, directional saccade asymmetry, pursuit gain, scan-path topology, or spatially lateralized gaze bias. Numeric target eccentricity for the pro- and antisaccade paradigms was also not retained in the exported feature-level dataset. Future studies should retain raw gaze trajectories and direction-specific indices to clarify whether lateralized oculomotor abnormalities contribute to the identified phenotypes.

Fourth, lesion characterization was based on clinically available imaging information rather than detailed lesion-symptom mapping. Frontal involvement, brainstem involvement, hemispheric involvement, and bilateral or multifocal involvement were coded from clinical imaging reports, but arterial territory, lesion volume, precise lesion coordinates, and tract-level involvement were not available. Therefore, the analysis could not determine whether the association between the LF-HS phenotype and frontal or multifocal lesions reflected attention/executive-control network disruption, direct oculomotor pathway involvement, overall lesion burden, or a combination of these mechanisms. Future work should combine eye tracking with standardized lesion segmentation and vascular-territory or network-based lesion mapping.

Fifth, several potential confounders related to visual, language, and oculomotor function should be considered. Participants with obvious aphasia, severe visual impairment, significant unilateral spatial neglect, hemianopia, or ocular motor disorders that would prevent valid task completion were excluded. Nevertheless, subtle visual-field deficits, mild spatial neglect, subclinical oculomotor abnormalities, or mild language-comprehension difficulties were not systematically quantified using dedicated standardized tests. Future studies should incorporate formal visual-field testing, standardized neglect and aphasia assessments, and direction-specific oculomotor measures to clarify the relative contributions of attentional/executive-control dysfunction, visual-spatial deficits, language comprehension, and primary oculomotor impairment to the observed eye-movement phenotypes.

Sixth, several attention/executive-control tasks used in this study are influenced by processing speed, auditory-verbal processing, working memory, and motor or written output. Therefore, these behavioral measures should be interpreted as task-performance validators rather than pure measures of individual cognitive domains. Future studies may benefit from including control tasks that separate processing speed, motor output, and sensory modality from attentional and executive-control demands.

Finally, although regression analyses suggested that eye-movement phenotype explained additional variance in Barthel Index scores, these findings do not establish predictive validity. Prospective multicenter studies are needed to determine whether eye-movement phenotypes predict functional recovery, rehabilitation response, or long-term participation outcomes.

Conclusion

This study showed that participants with subacute stroke had altered performance on tasks of attention and executive control and a consistent eye-movement pattern characterized by shorter total fixation duration and more frequent shifts in fixation location across paradigms. Integrated nonredundant eye-tracking features identified two exploratory eye-movement phenotypes, HF-LS and LF-HS. The LF-HS phenotype was associated with poorer task performance, greater neurological and functional impairment, higher emotional symptom burden, and more frequent frontal involvement and bilateral or multifocal involvement. Sensitivity analyses supported the stability of the two-cluster solution, and exploratory regression suggested that eye-movement phenotype explained additional variance in functional independence. These findings support eye tracking as a promising process-level adjunct for characterizing post-stroke cognitive and oculomotor heterogeneity, while emphasizing the need for longitudinal, multicenter, and mechanistically informed validation before clinical application.

Author contributions

X.F. designed the study. X.F., Y.T., and Y.S. conducted the experiments and collected the data. X.F. performed the analyses and drafted the manuscript. All authors reviewed and approved the final manuscript.

Funding

This work was supported by the Jiangsu Province Capability Improvement Project through Science, Technology, and Education (ZDXYS202210), Open Subjects of Jiangsu Health Development Research Center (JSHD2022037), the Third Batch of Changzhou Science and Technology Program Projects in 2023 (CJ20230060).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Ethics approval and consent to participate

The study protocol was approved by the Ethics Committee of the Third Affiliated Hospital of Soochow University (No. 2023-14) and all participants in the present study provided informed consent.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.