Eye tracking and machine learning to assess cognitive impairment in post-COVID-19 patients

Abstract

This study investigates whether oculomotor dysfunction in individuals with post-COVID-19 condition can serve as a biomarker for cognitive impairment as demonstrated in other neurological conditions. Eye movement metrics, including fixation, saccades, and smooth pursuit, as well as pupil size changes, were recorded using an eye tracker. Cognitive performance was assessed through established neuropsychological tests: Digit Symbol Test, Digit Span Backward, Trail Making Test (TMT), Stroop Color and Word Test (SCWT), together with the Controlled Oral Word Association Test (COWAT) for phonological fluency and an animal-naming task for semantic fluency. A total of 103 participants with post-COVID-19 condition were included in the analysis. Statistical correlations and k-means clustering were employed to analyse the relationship between oculomotor data and cognitive test results. Correlations between eye movements metrics and cognitive tests scores were modest ranging from 0.210 to 0.371. Clustering participants into three groups based on oculomotor metrics revealed statistically significant differences in cognitive performance. These findings suggest that eye tracking, unaffected by factors like language barriers and education level, my serve as a reliable method for assessing cognitive impairment in post-COVID-19 condition. This approach may complement neuropsychological tests, offering a more accessible and objective evaluation tool. Study registration: www.ClinicalTrials.gov, identifiers NCT05307575 (01-10-2021) and NCT05846126 (01-05-2023).

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-34664-2.

Introduction

Post-COVID-19 condition (PCC) refers to a range of symptoms lasting for at least 2 months that occur usually 3 months after the onset of COVID-19 infection, and that cannot be attributed to an alternative diagnosis¹. Such symptoms may persist from the initial illness or may be completely new following recovery from the acute phase¹. The global prevalence of PCC has been estimated at 43%², with higher rates among individuals who experienced severe COVID-19 infection requiring hospitalization during the acute phase compared with those who had a milder infection and did not require hospitalization².

There is no minimum number of symptoms required for a PCC diagnosis, and affected individuals often present a wide range of manifestations involving multiple organs and systems. Common symptoms include fatigue, headache, altered smell or taste, and shortness of breath³. Cognitive and neuropsychiatric symptoms, such as brain fog, attention difficulties and confusion, are also frequently reported. Objective impairments in executive function, working memory, and processing speed have been identified through neuropsychological assessment in PCC participants, and these deficits are observed regardless of hospitalization status or subjective cognitive complaints^4–6. Notably, individuals reporting cognitive complaints tend to show higher levels of depression and fatigue⁴. Overall, these findings indicate that subjective cognitive complaints are not a reliable marker of objective cognitive impairment in PCC, highlighting the need for systematic cognitive screening.

Neuropsychological testing poses a number of challenges such as limited sensitivity to mild deficits, interaction with age and education level, and the requirement of time as well as highly trained professionals. This encourages the development of alternative methods and tools to facilitate an early and more cost-effective diagnosis of cognitive deficits. In this framework, eye movement recording is a potential diagnostic tool given that certain measures derived from instructed eye movement tasks have been shown to be associated with cognitive test performance⁷, and studies suggest that eye movements may serve as potential biomarkers of cognitive impairment⁸. For instance, parameters which characterize different aspects of saccadic function such as latency in prosaccades and antisaccades as well as error rate in antisaccades have been associated with executive function, and suggested to be an indicator of cortical and subcortical function⁹. Similarly, smooth pursuit recordings can aid in assessing or identifying cognitive disorders in individuals with neurological conditions. This is because an adequate smooth pursuit performance requires a number of complex processes, including movement prediction, error detection and subsequent correction¹⁰. Further, it has been shown that smooth pursuit measures are correlated with cognition as well as visual processing, and therefore may also correlate with other cognitive domains in Alzheimer’s disease¹¹.

From a theoretical perspective, a similar rationale may apply to PCC. Post-COVID research has reported diffuse neuroinflammatory changes¹², microglial reactivity affecting attentional and executive networks¹³, and longitudinal frontoparietal cortical thinning¹⁴. Although these findings do not specifically target oculomotor pathways, they involve large-scale cognitive systems that are functionally relevant for attentional orienting and cognitive control, thereby providing a reasonable mechanistic basis for investigating eye-movement behaviour in PCC.

Consistent with this rationale, several studies have already reported oculomotor abnormalities in PCC, including alterations in saccades, smooth pursuit, fixation stability and pupillary responses^7,15,16. Notably, Benito-León et al.¹⁶ identified a characteristic eye-movement pattern in PCC, compatible with dysfunction in frontal–subcortical circuits. In addition, recent longitudinal evidence indicates that SARS-CoV-2 infection can lead to persistent impairments in oculomotor performance¹⁷, further supporting the relevance of eye tracking metrics for characterizing cognitive dysfunction in PCC.

This study aims to evaluate the potential of eye movements as biomarkers of cognitive impairment in individuals with PCC. To achieve this, the first objective was to investigate the relationship between eye movement parameters, measured through eye tracking, and cognitive performance in a heterogeneous group of individuals with PCC. The second objective was to form groups based on eye tracking metrics and analyse how cognitive performance differs across these data-driven subgroups.

Materials and methods

Participants

Participants with PCC taking part in the Nautilus study (ClinicalTrials.gov ID: NCT05307575) and the Rehab COVID (ClinicalTrials.gov ID: NCT05846126) recruited by the Consorci Sanitari de Terrassa (Terrassa, Barcelona, Spain) between November 2021 and February 2024 were invited to participate. The protocol was approved by the Drug Research Ethics Committee (CEIm) of Consorci Sanitari de Terrassa (02-20-107-070) and designed in accordance with the Declaration of Helsinki.

The inclusion criteria for PCC were: (a) confirmed diagnosis of COVID-19 according to WHO criteria with signs and symptoms of the disease during the acute phase; (b) a post-infection period of at least 12 weeks; (c) presence of persistent symptoms; and (d) age between 18 and 65 years. The exclusion criteria for the participants were: (a) established diagnosis of psychiatric, neurological, neurodevelopmental disorder, or systemic pathologies known to cause cognitive deficits before the episode of COVID-19; (b) motor or sensory alterations that impeded neuropsychological examination; (c) type 1 or 2 diabetes to ensure the absence of eye movement alterations that may result from diabetic neuropathy; (d) history of intraocular or refractive surgery; (e) glaucoma or any retinal disease; (f) diagnosed or suspected strabismus; (g) stereopsis > 100”; and (h) corrected near visual acuity of > 0.2 logMAR binocularly.

Neuropsychological tests

Participants underwent an assessment of attention, processing speed, and executive functions using standardized neuropsychological tests. The assessment was conducted by qualified clinical neuropsychologists.

The digit symbol test was used to assess information processing speed and some aspects of attention. It comprises a paper-and-pencil cognitive task presented on a single sheet of paper that requires individuals to match symbols to numbers according to a key located on the top of the page. The test score is given by the number of correct symbols achieved in 90 s¹⁸.

Working memory was assessed using the digit span backward¹⁸. Participants are asked to repeat digits in the reverse order in which they are presented. The measurement collected is the span, i.e., the maximum number of digits that the participant can repeat.

The Trail Making Test (TMT)¹⁹ was used to assess processing speed (Part A) and executive functions (Part B)^20–22. In Part A, participants connect numbered circles in sequence, whereas Part B requires alternating between numbers and letters. Performance was graded by completion time for each part.

The Stroop Color and Word Test (SCWT) was used to assess processing speed and the ability to inhibit cognitive interference^23–25. It includes three conditions: word reading, colour naming, and colour–word interference, the latter requiring participants to name the ink colour while inhibiting the automatic tendency to read the printed word. Performance was indexed by the number of correct responses in each condition.

Finally, verbal fluency was assessed using the Controlled Oral Word Association Test (COWAT)^26,27. Phonological fluency required participants to recall words beginning with the letters P, M, and R within one minute. Similarly, semantic fluency was assessed using the category “animals,” where participants were asked to name as many animals as possible within one minute. In both tasks, the total number of correct words produced was recorded as the final score²⁸.

Eye movements recording

Participants attended a second visit for the eye movements recording, which was conducted in binocular viewing using a desktop-mounted eye tracker EyeLink 1000 Plus (SR-Research Ltd., Ottawa-Ontario, Canada) at 1000 Hz. Prior to eye movements and pupil response recording, participants were screened to confirm adequate near vision and binocularity, and therefore exclude any participant who did not fulfil the inclusion criteria. Both the eye-tracker calibration (mean calibration accuracy ± SD: 0.83 ± 0.71°) and the subsequent eye-movement recording were performed under dim illumination conditions.

Participants were seated 60 cm away from a 17” LCD computer monitor (1280 × 1024 pixels, frame rate 60 Hz) with their heads supported by a chin rest while wearing their habitual near refractive error correction, if any. First, the eye tracker was calibrated for each participant using the standard EyeLink 9-point calibration procedure. After a successful calibration, a battery of visual tasks previously developed by our group¹⁵ was used to elicit and record saccadic and antisaccadic, smooth pursuit and fixational eye movements, as well as changes in pupil size in response to light. All visual stimuli and the complete battery of visual tasks were generated with Matlab (MathWorks, Natick-MA, USA) using the Psychophysics Toolbox library^29–31.

To ensure the reliability of the data, both the fixation and smooth pursuit tasks were conducted twice. In brief, during the fixation task, participants’ eye movements were recorded while they were fixating a cross of 0.5° situated at the centre of the monitor, first without any other stimuli for 10 s and then, while distractors appeared at different monitor locations for 30 s. For prosaccades and antisaccades, participants were asked to fixate on a central target and change gaze towards a peripheral target located at varying amplitudes (5°, 7.5° and 10°) from the central target (prosaccades) or in the opposite direction of the peripheral target (antisaccades) following three paradigms: Posner overlap, overlap and gap^32,33. Smooth pursuit eye movements were elicited and recorded by asking participants to track a visual stimulus (0.3°) that moved following a sinusoidal and then a linear (horizontal and vertical) motion. Finally, pupil response was recorded by asking participants to fixate on a white cross on a black background located in the centre of the monitor while a white LED light was switched on for 5 s. Pupil size was measured before and after activating LED light to evaluate pupil dilation and constriction.

Data analysis

Eye movement raw data were pre-processed by custom algorithms implemented in Python and the eye tracker’s pre-loaded software. Initially, the eye tracker software identified blinks and, subsequently, data from 200-ms intervals before and after each blink were systematically removed. Additionally, the eye tracker software incorporated two different heuristic filters³⁴ to smooth data and remove signal noise. The custom algorithms performed further filtering for saccade and fixation detection; a second order Savitzky-Golay filter of 21 samples length (21 ms) was applied to the computed velocity signal. Saccades were identified using an adaptive velocity threshold, dynamically adjusted based on signal noise levels³⁵. Table 1 presents all eye movements and pupil response parameters that were computed for further analyses.

Table 1.

Eye movements and pupil response parameters computed from eye tracking data.

Definition
Fixation All metrics are calculated for the two paradigms: before distractors appear and with distractors
RMS [°]	Root Mean Square error representing the variability in gaze position computed as the distance between consecutive measurements recorded by the eye tracker at each time point. This parameter reflects the precision of fixation, with lower RMS values indicating greater stability and less fluctuation in gaze.
N° of microsaccades	Number of saccades during the fixation task with an amplitude < 1º.
Prosaccades and antisaccades All metrics are calculated for the three paradigms: Posner overlap, overlap and gap, and three stimulus amplitudes: 5°, 7.5° and 10°
Amplitude [°]	Distance travelled by the eye during the saccade.
Directional deviation [°]	Absolute deviation from horizontal, 0° being perfectly horizontal and 90° perfectly vertical. This parameter measures the accuracy of the saccade’s directional alignment, with smaller deviations reflecting greater accuracy in targeting the intended direction.
Latency [ms]	Time that elapses between the onset of the peripheral target and the onset of the saccade.
Correct saccades [%]	Percentage of saccades or antisaccades with a direction within ± 50º from the peripheral target.
Amplitude absolute error [°]	Absolute difference between the eccentricity of the peripheral target (5°, 7.5° and 10°) and the saccadic amplitude.
Main sequence’s maximum peak velocity and parameter c	The following exponential function was fitted to the peak velocity versus amplitude data: 36, where Vmax is the theoretical upper bound of eye movement speed and c describes how quickly the eye movement speed increases with amplitude. A smaller c means the peak velocity reaches its maximum with smaller saccades.
Smooth pursuit All metrics are calculated for the three trajectories: sinusoidal, horizontal and vertical
RMS [°]	Root Mean Square error denoting the distance between the target and gaze positions. For each gaze position, the RMS is computed to the position of the target at the same timestamp, with lower RMS values indicating greater tracking accuracy.
Gain	Ratio of eye velocity to stimulus velocity.
Saccadic component [%]	Percentage of time spent on saccades during the smooth pursuit task.
Pupil response
Miosis time [ms]	Time that elapses between the LED is switched on and the maximum constriction is achieved.
Area reduction [%]	Area reduction (%) during miosis. The initial value is obtained as the mean pupil area within the last second before switching on the LED and the final value corresponds to the maximum constriction measured.
Area enlargement [%]	Area growth (%) during mydriasis. The initial value is obtained as that corresponding to maximum constriction, and the final one as the mean pupil area within the last second of the testing condition (75 s).

Statistical analysis and k-means

The data distribution was assessed individually for each variable using the Shapiro-Wilk normality test. Among the neuropsychological test scores and eye movement metrics, at least one variable in each pairwise comparison was found to be non-normally distributed. Therefore, partial Spearman correlations were used to examine their relationships. All correlations were adjusted for the participants’ years of education and age as covariates. Since the years of education strongly correlate with neuropsychological test scores, controlling for this variable ensures that its influence on the correlation with oculomotor metrics is accounted for. A p-value < 0.05 was considered significant. Analyses were conducted using RStudio version 2023.03.0 Build 386 (© 2009–2023 Posit Software, PBC Boston, USA) and the ppcor package for partial and semi-partial (part) correlations.

For smooth pursuit and fixation tasks, consistency between both repetitions was confirmed; for simplicity, only the first trials are reported. Similarly, pupil variables showed consistent results for both eyes and only the values for the right eye are reported. This approach ensured the reliability of the analysis and minimized the impact of random significant correlations. Due to operational reasons, the saccade test was only applied to a subgroup of 53 participants. In this case, correlations were initially examined separately for each paradigm (gap, overlap, and Posner) across different stimulus amplitudes (5°, 7.5°, and 10°). Furthermore, correlations were analysed in three additional ways: (1) for each stimulus amplitude irrespective of the paradigm, (2) for each paradigm irrespective of the stimulus amplitude, and (3) for the overall average values across all paradigms and amplitudes.

In addition to the correlation analyses, we applied a k-means clustering approach to identify subgroups within the eye tracking data. Prior to clustering, dimensionality reduction was performed using a Principal Component Analysis (PCA)^37,38. Outliers were identified using the Local Outlier Factor (n_neighbors = 20, contamination = 0.10)³⁹. To ensure consistent variable interpretation, all variables were adjusted such that higher values reflected better performance. The PCA component scores were then submitted to k-means clustering. Although the elbow method⁴⁰ indicated k = 5, this solution yielded clusters that were too small to be clinically meaningful. A lower-k solution was therefore chosen, with k = 3 offering the best balance between cluster size and interpretability.

For each identified group, the mean and standard deviation of neuropsychological test scores were calculated. To assess differences in cognitive performance across groups, both ANOVA (for normal distributions) and Kruskal-Wallis (for non-normal distributions) tests were applied. For categorical variables such as comorbidities, chi-squared test was conducted. Bonferroni post-hoc tests were subsequently conducted to identify specific pairwise differences between groups.

Intra-correlation analyses were subsequently conducted within each group to examine the relationships between variables within these subpopulations. Identifying distinct patterns of inter-variable relationships within each group would further validate the clustering results and highlight their robustness.

Results

Eye movement data from a total of 115 participants were recorded. However, 12 participants did not meet the inclusion criteria and were excluded from the analysis. The number of valid recordings varied slightly across visual tasks due to intermittent data loss caused mainly by droopy eyelids, back reflections from glasses, etc. Fixation and smooth pursuit data were available for 103 participants, while pupil-light reflex data were available for 92 participants. Saccadic tasks were administered to a smaller subgroup (n = 53 for prosaccades; n = 51 for antisaccades) because they were introduced into the experimental protocol at a later stage. Table 2 shows demographic, clinical, visual and reported symptoms details.

Table 2.

Demographic, clinical, visual, neuropsychological and reported symptoms data of participants with PCC (n = 103). Neuropsychological test results are presented as Raw scores.

Age (years), mean ± sd	50.75 ± 7.43
Years of education, mean ± sd	14.21 ± 3.42
Sex, female, n (%)	85 (82.5)
Severity, n (%)
Hospitalized	24 (23.3)
Outpatient	79 (76.7)
Time from infection to assessment (days), mean ± sd	620.88 ± 377.19
Comorbidities, n (%)
Heart disease	4 (3.9)
Respiratory disease	14 (13.6)
High blood pressure	9 (8.7)
Dyslipidemia	14 (13.6)
Obesity	23 (22.3)
Tobacco smoking	11 (10.7)
Visual function, mean ± sd
Near-distance visual acuity (logMAR)	0.04 ± 0.07
Stereopsis (arc sec)	62.68 ± 62.54
Neuropsychology (raw score), mean ± sd
Digit symbol test	63.81 ± 17.08
Digit span backward	4.33 ± 1.11
TMT A (s)	38.54 ± 17.54
TMT B (s)	85.13 ± 41.97
SCWT color	60.66 ± 13.42
SCWT word	90.12 ± 23.68
SCWT color-word	37.29 ± 12.39
Phonological fluency	40.12 ± 11.78
Semantic fluency	20.03 ± 5.45
Symptom reported, n (%)
Fatigue	76 (75.2)
Joint Pain/Body Aches	58 (56.3)
Headaches	58 (56.3)
Dyspnea on exertion	32 (31.1)
Altered Smell	29 (28.2)
Altered Taste	3 (2.9)
Dizziness	36 (35)
Dermatologic issues	13 (12.6)
Cognitive complaints	56 (54.4)
aMemory deficits	51 (91.1)
aLack of concentration	41 (73.2)
aΒrain fog	41 (73.2)
aProblems with language	44 (78.6)
aProblems with executive functioning	34 (60.7)
Depressive symptoms	32 (31.1)
Anxiety	30 (29.1)
Post-traumatic stress	10 (9.7)
Difficulty Sleeping	13 (12.6)

^a % Calculated for people with cognitive complaints (n = 56).

Partial correlations within the full sample

In Table 3, statistically significant correlations between neuropsychological test scores and all eye movement metrics, including fixation, prosaccades, antisaccades, smooth pursuit, and pupil response, are shown. The results obtained are described below for each type of eye movement.

Table 3.

Significant correlations between neuropsychological tests and eye tracking parameters in fixation, prosaccade and antisaccades, smooth pursuit and pupil response tasks.

Neuropsychological test	Eye movement parameter	Paradigm	Correlation coefficient	p value
Fixation (n: 103)
SCWT color	RMS [°]	Distractors	− 0.252	0.010
SCWT word			− 0.259	0.007
SCWT color-word			− 0.290	0.003
Prosaccades (n: 53)
Digit span backward	Amplitude [°]	Overall averaged	0.371	0.018
TMT A			− 0.324	0.022
TMT B			− 0.303	0.033
SCWT color			0.333	0.007
SCWT word			0.321	0.023
SCWT color-word			0.365	0.009
Antisaccades (n: 51)
Digit symbol	Directional deviation [°]	10° stimulus amplitude	− 0.302	0.030
Digit span backward			− 0.299	0.031
TMT B			0.299	0.031
SCWT color			− 0.357	0.009
SCWT word			− 0.313	0.024
SCWT color-word			− 0.325	0.020
Smooth pursuit (n: 103)
SCWT color	RMS [°]	Sinusoidal	− 0.247	0.012
SCWT color-word			− 0.281	0.004
Semantic fluency			− 0.270	0.006
SCWT color-word		Vertical	− 0.289	0.003
Pupil response (n: 97)
Digit span backward	Area reduction [%]		0.225	0.031
TMT B			− 0.210	0.045
Sematic fluency			0.212	0.043

Fixation test

Throughout the period when distractors were displayed on the screen, a significant inverse correlation was observed between Stroop tests and the RMS marker. Higher performance on the Stroop tests was associated with greater fixational stability (i.e., lower RMS). Detailed numerical results are provided in Table 3.

Saccades

Paradigm-specific analyses revealed no significant associations for either prosaccades or antisaccades. However, when saccade metrics were averaged across paradigms and stimulus amplitudes, several meaningful associations emerged. Prosaccade amplitude correlated with lower performance on several neuropsychological tests (digit symbol, digit span backward, TMT B, SCWT color, and SCWT color-word tests). Also, antisaccade directional deviation (specifically at the 10º amplitude) positively correlated with performance across multiple tests (digit span backward, TMT A and TMT B, SCWT word, SCWT color, SCWT color-word tests) (Table 3).

Smooth pursuit

SCWT color, SCWT color-word, and semantic fluency tests inversely correlated with the RMS of the smooth pursuit task (Table 3). These findings indicate that lower RMS values (i.e., better tracking of the stimulus) were associated with better performance on the neuropsychological tests.

Pupil response

As a representative example of pupil size data, Fig. 1 depicts pupil area over time during an individual recording.

Fig. 1.

Representative example of a pupil response recording over time. The grey area is the timespan with the LED switched off, whereas the period with the LED on is represented in white.

In this case, borderline significant correlations were observed with various tests including the digit span backward, TMT part B, and the semantic fluency (Table 3). In particular, better performance on these tests was generally associated with larger pupil constriction.

Principal component analysis and k-means clustering

The PCA analysis and k-means clustering categorized participants into three groups (groups 0, 1, and 2) achieving a Silhouette coefficient of 0.32. This coefficient measures the quality of clustering by evaluating how well each participant fits within its assigned group compared to others⁴¹. Figure 2 illustrates the data distribution in the three-dimensional Principal Component (PC) space, where each point represents a participant.

Fig. 2.

Distribution of participants in the PC space (only the first three PCs are considered) based on eye tracking variables: a 3D overview (A) and orthogonal 2D projections (B, C and D). Each point corresponds to a participant, with color indicating group membership based on k-means clustering.

To assess internal stability of the clustering solution, k-means was repeated across 200 different random seeds. Silhouette coefficients showed limited variability across seeds (mean = 0.30, SD = 0.03), and cluster membership stability was moderate, with a mean Adjusted Rand Index (ARI) of 0.67 (SD = 0.31)⁴².

PC1 reflected smooth-pursuit performance, PC2 captured pupillary light-reflex dynamics, and PC3 represented fixation stability and pursuit noise. Full PCA loadings are reported in Supplementary Tables S1–S2. Cluster centroids for the eye tracking variables (Supplementary Tables S3–S4) revealed three distinct oculomotor profiles. Cluster 2 showed intermediate pursuit values combined with the most stable fixation pattern and intermediate pupil responses. Cluster 1 was characterized by high pursuit gain, low RMS error across trajectories, and stronger pupil responses. Cluster 0 exhibited lower pursuit gain, higher RMS error, greater fixation instability, and smaller pupil responses.

Table 4 shows demographic, clinical, visual, neuropsychological and reported symptoms details for each group. No significant differences were found in demographic variables between groups. While stereopsis and dyspnea reports showed an overall difference, Bonferroni-corrected post-hoc tests revealed no significant pairwise differences. In contrast, significant differences were found in the neuropsychological tests scores, including the Digit symbol test, SCWT Color-Word, and Semantic Fluency Test.

Table 4.

Demographic, clinical, visual and neuropsychological data by PCC clusters: group 0 (n = 17), group 1 (n = 64), group 2 (n = 22). Neuropsychological test results are presented as Raw scores.

	Group 0	Group 1	Group 2	Statistical test	Statistic	p value
Age (years), mean ± sd	51.17 ± 8.15	50.63 ± 7.70	50.78 ± 6.29	Kruskal-Wallis	0.15	0.928
Years of education, mean ± sd	14.00 ± 3.30	14.14 ± 3.51	14.59 ± 3.36	Kruskal-Wallis	0.78	0.678
Sex, female, n (%)	16 (94.2)	54 (84.4)	15 (68.2)	χ2	3.93	0.140
Severity, n (%)				χ2	0.27	0.875
Hospitalized	4 (23.5)	14 (21.9)	6 (27.3)
Outpatient	13 (76.5)	50 (78.1)	16 (72.7)
Time from infection to assessment (days), mean ± sd	715.53 ± 349.14	613.95 ± 389.38	567.91 ± 364.65	Kruskal-Wallis	1.59	0.452
Comorbidities, n (%)
Heart disease	0 (0)	4 (6.3)	0 (0)	χ2	2.53	0.281
Respiratory disease	3 (17.7)	7 (10.9)	4 (18.2)	χ2	1.02	0.602
High blood pressure	2 (11.8)	5 (7.8)	2 (9.1)	χ2	0.27	0.875
Dyslipidemia	1 (5.9)	10 (15.6)	3 (13.6)	χ2	1.09	0.581
Obesity	3 (17.7)	16 (25.0)	4 (18.2)	χ2	0.70	0.706
Tobacco smoking	3 (17.7)	7 (10.9)	1 (4.6)	χ2	1.74	0.419
Visual function, mean ± sd
Near-distance visual acuity (logMAR)	0.07 ± 0.01	0.03 ± 0.07	0.01 ± 0.03	Kruskal-Wallis	6.01	0.050
Stereopsis (arc sec)	84.94 ± 95.06	64.35 ± 58.45	40.10 ± 27.76	Kruskal-Wallis	6.90	0.032*
Neuropsychology (raw score), mean ± sd
Digit symbol test	54.82 ± 14.84	67.05 ± 15.68	61.32 ± 20.21	Kruskal-Wallis	6.59	0.037*
Digit span backward	3.94 ± 0.83	4.41 ± 1.23	4.41 ± 0.85	Kruskal-Wallis	3.25	0.197
TMT A (s)	42.53 ± 18.47	38.95 ± 18.57	34.27 ± 12.96	Kruskal-Wallis	3.23	0.199
TMT B (s)	110.59 ± 6.51	82.39 ± 39.10	73.41 ± 29.07	Kruskal-Wallis	5.39	0.068
SCWT color	53.77 ± 15.78	61.84 ± 12.22	62.55 ± 13.77	ANOVA	2.81	0.065
SCWT word	81.77 ± 28.69	90.95 ± 21.55	94.14 ± 25.07	Kruskal-Wallis	4.29	0.117
SCWT color-word	29.88 ± 10.55	37.16 ± 10.62	43.41 ± 15.45	Kruskal-Wallis	11.30	0.004*
Phonological fluency	37.94 ± 14.09	39.89 ± 11.69	42.46 ± 10.16	ANOVA	0.73	0.484
Semantic fluency	17.35 ± 6.08	19.98 ± 4.87	22.23 ± 5.83	ANOVA	4.07	0.020*
Symptom reported, n (%)
Fatigue	13 (76.5)	45 (70.3)	18 (81.8)	χ2	1.313	0.519
Joint Pain/Body Aches	11 (64.7)	36 (56.3)	11 (50.0)	χ2	1.337	0.512
Headaches	10 (58.8)	37 (57.8)	11 (50.0)	χ2	0.709	0.702
Dyspnea on exertion	7 (41.2)	23 (35.9)	2 (9.1)	χ2	6.942	0.031*
Altered Smell	4 (23.5)	19 (29.7)	6 (27.3)	χ2	0.194	0.907
Altered Taste	0 (0)	3 (4.8)	0 (0.0)	χ2	1.865	0.394
Dizziness	6 (35.3)	22 (34.4)	8 (36.4)	χ2	0.043	0.979
Dermatologic issues	2 (11.8)	9 (14.1)	2 (9.1)	χ2	0.395	0.821
Cognitive complaints	7 (41.2)	34 (53.1)	15 (68.2)	χ2	2.474	0.290
aMemory deficits	6 (85.71)	32 (94.11)	13 (86.7)	χ2	2.188	0.335
aLack of concentration	4 (57.14)	26 (76.47)	11 (73.3)	χ2	2.852	0.240
aΒrain fog	5 (71.43)	27 (79.41)	9 (60.0)	χ2	0.929	0.628
aProblems with language	6 (85.71)	25 (73.53)	13 (86.7)	χ2	3.143	0.208
aProblems with executive functioning	5 (71.43)	20 (58.82)	9 (40.9)	χ2	0.810	0.667
Depressive symptoms	7 (41.2)	17 (26.6)	8 (36.4)	χ2	1.942	0.379
Anxiety	6 (35.3)	16 (25.0)	8 (36.4)	χ2	1.493	0.474
Post-traumatic stress	2 (11.8)	5 (7.8)	3 (13.6)	χ2	0.738	0.691
Difficulty Sleeping	1 (5.9)	7 (10.9)	6 (27.3)	χ2	4.490	0.106

^a% Calculated for people with complains in each group. *p < 0.05. Statistically significant differences across groups.

Figure 3 presents bar charts illustrating the mean scores of the neuropsychological tests for each of the three groups. Group 0 consistently demonstrated the lowest mean performance across all tests, followed by Group 1, which showed intermediate performance (except for the digit symbol test). Group 2, in contrast, exhibited the highest mean performance (except for the digit symbol test). It is important to note that in the TMT A and TMT B, higher scores indicate worse performance, as they represent the time required to complete the task. Statistically significant differences were observed between Group 0 and Group 1 in the digit symbol test (p = 0.032) and SCWT color-word (p = 0.028). Additionally, significant differences were found between Group 0 and Group 2 in the SCWT color-word (p = 0.008) and semantic fluency (p = 0.049).

Fig. 3.

Bar charts of the mean scores of the neuropsychological tests within each group identified with the k-means clustering. Standard deviation is shown as error bars. Significant differences between groups are indicated by asterisks: between Group 0 and Group 1 in the digit symbol (p = 0.032) and SCWT color-word (p = 0.027), and between Group 0 and Group 2 in the SCWT color-word (p = 0.008) and semantic fluency (p = 0.049).

Partial correlations within each group

Across clusters, significant associations occurred almost exclusively in saccade metrics and involved most cognitive domains. Specifically, Group 0 presented 13 significant correlations between saccade metrics and neuropsychological tests scores and a significant correlation between pupil dilation and the digit span backward test. Group 1 showed 8 significant correlations of three saccade metrics. Group 2 exhibited 19 significant correlations. Within-group partial correlation results are detailed in the Supplementary Table S5.

Discussion

The first aim of this study was to assess the relationship between eye tracking parameters and cognitive performance in a heterogeneous group of participants with PCC. Significant associations between specific eye movement metrics and cognitive performance were found in the studied sample.

In the fixation task, participants were instructed to maintain fixation on a stationary cross and consisted of two parts: with and without distractors. Less stable fixation when distractors were present was linked to worse performance on the three parts of SCWT, which aligns with the attentional control required in both tasks.

Fixation stability, measured as RMS, was significantly associated with the performance on both the word and color parts of the SCWT, which primarily reflect processing speed. This result is in agreement with a previous study in stroke patients where fixation measures were related to processing speed, specifically through the digit symbol test⁴³. While processing speed is not exclusive to attention, its link to fixation stability may stem from the attentional demands required for gaze control. Fixation stability also showed a correlation with the color-word part of the SCWT, which primarily assesses inhibitory control. This association is consistent with the attentional demands of the distractor condition, where participants had to suppress reflexive eye movements to maintain fixation. These results support the idea that fixation stability also depends on inhibitory control in tasks requiring resistance to distractors.

Given that saccade data were available for only 53 participants, the strength of the corresponding associations is necessarily more limited. Performance on saccadic eye movement tasks showed significant correlations with cognitive measures in the full sample. However, paradigm-specific analyses revealed no significant effects for either prosaccades or antisaccades. When saccade amplitude was averaged across all stimulus amplitudes (5°, 7.5°, and 10°) and paradigms, greater prosaccade amplitude correlated with better performance on working memory, processing speed, and executive function tests. It should be noted that the prosaccade association was present only for the averaged amplitude measure; neither amplitude error (hypometria or hypermetria) nor directional deviation correlated with cognitive scores. This indicates that the effect is unlikely to reflect accuracy. Averaging across paradigms may instead capture a broader saccadic response tendency, although the cognitive relevance of this pattern remains unclear and warrants further investigation.

In the antisaccade task, significant correlations with cognitive performance emerged only at the largest stimulus amplitude. Specifically, greater antisaccade directional deviation was associated with poorer performance on working memory, processing speed, and inhibitory control tests. Antisaccades are cognitively demanding, and their performance is linked to inhibitory control as they require participants to suppress the reflex impulse to look toward a stimulus and instead execute a voluntary movement in the opposite direction^44,45. In this sense, individuals with greater antisaccade directional deviation may exhibit less efficient suppression mechanisms. Previous studies have also found associations between antisaccade performance and cognitive functions across different populations. Stroop interference has been linked to antisaccade errors in people with multiple sclerosis⁴⁶ and mild traumatic brain injury⁴⁷. Deficits in attention and working memory have been significantly correlated with increased antisaccade error rates in subjects with schizophrenia⁴⁸. Additionally, working memory capacity has been associated with antisaccade performance in healthy young adults⁴⁹. These findings, together with our results, reinforce the robust relationship between antisaccade performance and cognitive function.

In this study, saccade parameters grouped by stimuli amplitude only showed significant associations at the largest amplitude (10°) for both the full sample tests and for intra-group correlations. This pattern may reflect amplitude-dependent properties of the saccadic system. Previous work has shown that small saccades tend to overshoot and larger ones undershoot their targets, producing the range effect. As a consequence, mid-range amplitudes (~ 4°–6°) show greater accuracy and therefore less variability⁵⁰. Such limited variability may constrain the detection of individual cognitive differences. From a measurement perspective, small saccades are also more susceptible to noise, whereas larger trajectories show a clearer signal-to-noise ratio. At the neural level, larger saccades require stronger interaction between the superior colliculus, which generates the movement, and frontal and parietal regions that support planning, spatial attention, and inhibitory control. Because these areas are also involved in cognition, performance on larger saccades may be more affected by cognitive status⁵¹. The present findings tentatively suggest that larger-amplitude saccades may be more informative for detecting oculomotor–cognitive associations in PCC.

Unlike the rapid and discrete saccadic movements, smooth pursuit enables the eyes to steadily and accurately track a moving object. We found that the accuracy of tracking a sinusoidal trajectory correlated with better inhibitory control, cognitive flexibility, and semantic retrieval, whereas vertical smooth pursuit accuracy was only associated with better inhibitory control.

Sinusoidal pursuit involves predicting and continuously adjusting to a changing motion, relying on anticipatory and predictive processes⁵². This type of predictive tracking involves frontal areas that support planning and anticipation^53,54, parietal regions that help integrate motion information⁵⁵, and cerebellar areas that fine-tune eye velocity⁵⁶. This may explain why it correlated not only with SCWT tests but also with semantic fluency, which similarly depends on flexible, strategic retrieval. In contrast, linear vertical pursuit represents a simpler tracking task that primarily requires sustaining a constant velocity. A similar consideration is the anisotropy between horizontal and vertical linear pursuit. Although both tasks used constant speed motion, vertical pursuit is generally less efficient, showing lower gain and greater variability than horizontal pursuit⁵⁷. Vertical pursuit may increase reliance on inhibitory control and processing speed, thus, helping explain why only this task showed associations with cognitive function. Vertical movements are also more susceptible to eyelid interference, especially at downward gaze positions, and higher noise in video-based eye tracking⁵⁸. Therefore, the observed anisotropy likely reflects a combination of greater task demands for vertical tracking and modest differences in measurement reliability.

Previous studies have identified impairments in smooth pursuit eye movements in individuals with various pathological conditions. For example, deficits in schizophrenia patients have been associated with executive function impairments⁵⁹. Similarly, in individuals with mild traumatic brain injury, deficits in predictive smooth pursuit have been correlated with difficulties in attention and cognitive flexibility⁶⁰. Previously, our group¹⁵ found that PCC participants exhibited greater corrective saccades during the horizontal smooth pursuit compared to healthy controls, regardless of COVID-19 severity. The present findings build upon these results, showing that smooth pursuit accuracy correlates with executive function and semantic retrieval in our PCC sample.

Finally, greater pupillary constriction was associated with better performance in working memory, cognitive flexibility, and semantic fluency across the entire sample. Previous research has linked pupillometry to cognitive processes such as forward digit retrieval in healthy children and adults⁶¹ and inhibitory control in healthy adults, as measured by the SCWT⁶². However, these studies examined task-evoked pupillary responses during cognitive engagement, whereas our measure reflects a light-evoked pupil response (PLR) obtained during passive fixation. This distinction limits direct comparability and constrains the interpretations that can be drawn. The PLR primarily reflects autonomic and brainstem mechanisms, including parasympathetic pathways, although it may still be modulated indirectly by central arousal systems such as the LC–NE pathway⁶³. To avoid overstating the role of cognitive control, we interpret the observed correlations cautiously: the observed associations align with the idea that pupillary responses reflect cognitive processes beyond basic autonomic function but do not imply a direct causal relationship. Additional methodological factors may also influence the PLR and should be considered as limitations. These include time of day, fatigue, and substances such as caffeine or medication, all of which can alter baseline pupil size and reactivity. Although ambient light was standardized, residual variability from these sources cannot be completely excluded.

From the perspective of effect-size interpretation, most of the observed correlations fall within the moderate range typical of psychological and neuropsychological research. According to empirical benchmarks⁶⁴, most correlation coefficients in our study fall within the middle and upper thirds of the distribution reported across 380 meta-analyses (|ρ| ~0.20–0.37). Using field-specific guidelines for Spearman correlations⁶⁵, fixation, smooth pursuit, and pupil associations fall within the moderate range (|ρ| ~ 0.21–0.29), whereas correlations with saccade parameters reach the ‘large’ range for individual-differences research (|ρ| ≥0.30). However, these strongest saccadic effects were accompanied by wide confidence intervals (CI) due to the smaller sample size (e.g., |ρ| = 0.37, 95% CI 0.11–0.58), indicating substantial uncertainty in their true magnitude. In contrast, weaker effects from fixation and pursuit tasks showed narrower CIs due to larger sample sizes. Although the analyses were hypothesis-driven, the results are subject to an increased risk of Type I error because multiple correlations were examined. Thus, these findings should be interpreted cautiously and viewed as preliminary trends rather than definitive clinical markers. Replication and longitudinal studies will be necessary to determine whether these associations have prognostic value.

Our second aim was to establish groups based on k-means clustering with eye tracking measurements and compare cognitive performance between these subgroups. Furthermore, we explored how eye tracking parameters were related to neuropsychological variables within each group to identify patterns of performance. Machine learning has previously been used with eye tracking data to classify clinical populations. Yang et al.⁶⁶ used fixation and visual scanning patterns to distinguish individuals with depression from healthy controls. Moreover, they identified distinct groups within the clinical population based on differences in eye tracking metrics. Benito-León et al.¹⁶ analysed eye tracking metrics to classify COVID-19 patients with cognitive impairment and healthy controls.

Our machine learning-based clustering analysis identified three distinct eye movement profiles. The silhouette coefficient (0.32) obtained falls in the range that Kaufman and Rousseeuw⁶⁷ describe as weak internal structure. Repeating the clustering across 200 random seeds yielded consistent silhouette values (mean = 0.30, SD = 0.03) and moderate membership stability (mean ARI = 0.67, SD = 0.31), indicating that while the overall partition is reproducible, individual participant assignments show some variability. Importantly, physiological measures, such as eye movement metrics, tend to be inherently noisy and variable across individuals, which can blur cluster boundaries. Accordingly, the clusters should be interpreted with caution. Examination of cluster centroids (Supplementary Tables S4–S5) revealed three characteristic eye-movement profiles. Cluster 2 showed the most stable fixation pattern with intermediate pursuit performance. Cluster 1 showed the opposite pattern and stronger pupil light-reflex responses. In contrast, Cluster 0 exhibited the least efficient oculomotor profile. Although clusters partly overlapped, these centroid patterns suggest that the clustering captured a gradient of oculomotor efficiency, with Cluster 0 representing the weakest profile.

Overlap to some extent is expected in this context between intermediate and higher-performing groups. Unlike studies that classify patients versus healthy controls where group differences are larger^16,66, all participants in this study had the same condition marked by mild and heterogeneous cognitive changes. This results in broader distributions and naturally blurs cluster boundaries, especially between intermediate profiles. In this framework, eye tracking metrics appear most effective for identifying individuals with markedly impaired profiles (Cluster 0).

Despite blur boundaries, the subgroups showed consistent differences in cognitive performance, providing external validation of the clustering solution. These differences were most apparent in tasks involving working memory and executive control. Taken together, these results suggest that eye tracking metrics might have captured meaningful variability in cognitive function within our PCC sample, even though the internal cluster structure was modest.

Intra-group correlations did not reveal any distinct pattern between groups. Although the number of significant correlations varied (Group 0: 13; Group 1: 8; Group 2: 19), the types of associations were mostly with saccadic metrics and maximum amplitude stimuli with nearly all cognitive domains evaluated. This suggests that while eye tracking data can distinguish individuals based on cognitive performance, it may not be sensitive enough to detect potential differences in the mechanisms linking eye movements and cognition across groups. Instead, clustering may have primarily grouped individuals by the severity of their impairment rather than by distinct cognitive processes. Another possible explanation is that clustering reduced variability within each group, making these relationships less evident. Additionally, although no differences were found between groups in emotional state, fatigue, or symptom severity and duration, fluctuations in these factors at the time of testing may have influenced cognitive performance, potentially introducing noise into the observed correlations.

It is crucial to evaluate the study’s strengths and limitations in order to interpret its findings. The cross-sectional design is a limitation, as it only offers a snapshot of the relationships between cognitive performance and eye movements. This underscores the necessity of longitudinal studies to monitor the evolution of PCC cognitive impairment over time. Additionally, the study lacks a large control group, limiting the ability to analyse correlations within controls. At the cluster analysis level, the moderate silhouette score suggests that there is some overlap between groups. Although future studies could apply other clustering approaches, such as hierarchical or density-based algorithms, to potentially refine cluster separability, the gradient-like organization of the data suggests that substantial improvements in cluster separation are unlikely. However, the study’s strengths include a comprehensively evaluated cohort of PCC participants and a large sample size, which were assessed using an extensive array of cognitive tests. This approach ensured a cognitively pure sample by excluding individuals with other deficits. In addition, it provides valuable insights into the relationship between oculomotor function and cognitive performance by examining a distinctive set of oculomotor parameters across a variety of isolated eye movement tasks, despite the aforementioned limitations.

Conclusions

To conclude, our findings reveal significant associations between oculomotor performance and cognitive abilities, particularly in tasks requiring executive function, attention and processing speed. Additionally, the cluster analysis highlights distinct oculomotor profiles linked to varying levels of cognitive function, particularly for the group with the lowest cognitive scores. Several design and data limitations must be acknowledged. The correlation effect sizes were generally moderate with wide confidence intervals for saccadic metrics (e.g., |ρ| = 0.37, 95% CI = 0.11–0.58), due to the reduced sample size. The clustering solution showed only modest cohesion (silhouette = 0.32) and imbalanced cluster sizes. Future studies should further investigate whether eye movement metrics offer predictive information about cognitive status, beyond their observed associations. Nonetheless, its clinical utility will depend on prospective validation and diagnostic-accuracy studies that determine its predictive value and feasibility as part of real clinical routines. Forgiven the time-intensive nature of neuropsychological testing and barriers like language, social disparities, and the need for skilled professionals, eye tracking may serve as a promising adjunct to existing cognitive approaches.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Joan Goset and Mar Ariza contributed equally to the study. Mikel Aldaba, Meritxell Vilaseca, Mar Ariza and Maite Garolera contributed to the study conception design and funding. Material preparation and data collection was conducted by Valldeflors Vinuela-Navarro, Luis Perez-Mañá, Joan Goset, Clara Mestre, Neus Cano, Mar Ariza and Bárbara Delàs. Data analysis was performed by Joan Goset and Clara Mestre. The first draft of the manuscript was written by Joan Goset under the supervision of Mar Ariza, and all authors reviewed and provided feedback on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

Grant PID2023-146101OB-I00 funded by MICIU/AEI/ 10.13039/501100011033 and by ERDF, EU.This publication is part of the project TED2021-130409B-C54 and TED2021-130409B-C51, funded by Ministerio de Ciencia, Innovación y Universidades,Agencia Estatal de Investigación (MCIN/AEI/10.13039/501100011033) and the European Union “NextGenerationEU”/PRTR.

Data availability

The dataset generated during this study has been deposited in the CORA-RDR repository. It will be made publicly available upon acceptance and publication of the article under a CC BY-NC-SA 4.0 license at the following DOI: [https://doi.org/10.34810/data2224](https:/doi.org/10.34810/data2224) .

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval

The protocol was approved by the Drug Research Ethics Committee (CEIm) of Consorci Sanitari de Terrassa (02-20-107-070) and designed in accordance with the Declaration of Helsinki.

Consent to participate

Written informed consent was obtained from all individual participants included in the study.