Eye tracking as a diagnostic tool in Alzheimer's disease, mild cognitive impairment, and related dementias: a systematic review

Abstract

Alzheimer's disease (AD) pathology begins years before symptoms emerge, making early detection essential. Eye tracking offers a rapid, non‐invasive means of identifying early cognitive decline through oculomotor disturbances. This Preferred Reporting Items for Systematic Reviews and Meta‐Analyses (PRISMA)‐ and Population, Intervention, Comparison, and Outcome (PICO)‐guided systematic review evaluated studies from PubMed, ACM Digital Library, and Google Scholar on eye tracking in mild cognitive impairment (MCI), AD, and related dementias. Seventy‐one studies met the inclusion criteria. Antisaccade tasks consistently distinguished AD and MCI from healthy controls, with impaired accuracy, longer latencies, and reduced gain. Non‐saccadic paradigms (e.g., visual search, free viewing) indicated diminished exploratory behavior in AD, with mixed findings in MCI. A major limitation was the lack of cohorts defined by current biological criteria, hindering clinical translation. In a subset, classical machine‐learning (ML) models and deep neural networks reported accuracies of 0.72 to 0.97. Overall, antisaccade tasks show strong promise for early AD screening; future work should adopt biologically defined cohorts and scalable, accessible eye‐tracking technologies.

Highlights

• Antisaccade tasks best distinguish AD and MCI from HCs.

• Visual search/free‐view tasks showed diminished exploratory behavior in AD.

• Most studies lack biomarker‐based AD criteria, creating a major research gap.

1. INTRODUCTION

Dementia represents a diverse group of disorders that impact memory, cognition, and daily functioning, with considerable heterogeneity in their underlying causes, risk factors, clinical manifestations, and disease trajectories. Epidemiological projections highlight the mounting global burden, with 47 million people affected in 2015 and an anticipated increase to 132 million by 2050. ¹ Yet these numbers mask important regional and demographic differences. In some high‐income countries (HICs), age‐specific incidence appears to be declining, a trend predominantly observed among individuals with higher educational attainment – a factor associated with cognitive reserve, healthier lifestyles, and improved access to healthcare from early in life, but also one that contributes to persistent disparities in dementia risk. Meanwhile, regions such as Latin America, North Africa, and the Middle East currently face some of the highest rates and sharpest increases in dementia prevalence, resulting in significant social and economic challenges, especially in settings where healthcare resources are scarce and caregiver burdens high. ²

Alzheimer's disease (AD), the predominant cause of dementia, accounts for ≈ 70% of cases. Its profound impact extends beyond health, imposing an enormous economic toll estimated at $818 billion worldwide in 2015 – equivalent to 1.1% of global GDP. ¹ Notably, the distribution of these costs varies: in HIC, formal care predominates, while in low‐ and middle‐income countries (LMIC), informal family care forms the bulk of the burden.

Biologically, AD is distinguished by the accumulation of amyloid beta (Aβ) plaques and tau‐containing neurofibrillary tangles, pathologies that precede clinical symptoms by years or even decades. Recent frameworks now define AD primarily as a biological continuum: neuropathological changes – detectable via biomarkers – emerge long before the onset of overt cognitive decline. ³ This redefinition emphasizes the importance of early detection. Mild cognitive impairment (MCI), once considered simply a prodromal stage, is now recognized as part of this AD continuum. Still, not all individuals with MCI progress to dementia, though the risk remains high: about 80% advance to dementia within 6 years.

From the earliest stages, AD pathology spreads in a predictable sequence, first affecting medial temporal lobe structures, then advancing to limbic and neocortical regions. Tau pathology, in particular, appears more closely tied to clinical decline than amyloid, further underscoring the critical need for early identification.

The search for early, accessible markers of AD has prompted growing interest in the visual system and ocular biomarkers. The retina, as an extension of the central nervous system, may reflect cerebral processes. Degenerative changes in both the retina and eye‐movement‐control pathways have been increasingly recognized as sensitive indicators of neurodegeneration – not only in AD but also in other disorders such as Parkinson's disease and frontotemporal dementia (FTD). ⁴ , ⁵ Eye movement abnormalities may provide a non‐invasive window into brain health, as their control depends on widely distributed neural circuits vulnerable to AD pathology. Consequently, eye‐tracking technology has emerged as a promising, practical, and scalable approach for early AD detection.

RESEARCH IN CONTEXT

1. Systematic review: The authors searched PubMed, ACM Digital Library, and Google Scholar for eye‐tracking literature in Alzheimer's disease, mild cognitive impairment, and related dementias. A Python‐based automated screening followed by manual review identified 71 relevant studies. These reports, covering saccadic and exploration paradigms, were critically analyzed.

2. Interpretation: Our meta‐analysis identifies the antisaccade task as the most robust oculomotor marker for distinguishing neurodegeneration from healthy aging. We provide a comprehensive framework categorizing existing paradigms to guide future clinical applications.

3. Future directions: Future research must address critical gaps: (a) the lack of studies using biological AD definitions (molecular biomarkers); (b) the high heterogeneity in eye‐tracking paradigms requiring protocol standardization; and (c) the scarcity of data from Low‐ and Middle‐Income Countries and underrepresented populations. Integrating accessible web‐technologies and machine learning will be crucial for scalable, remote disease monitoring.

A series of recent reviews underscore both the promise and complexity of this approach. Narrative and systematic reviews have shown that specific eye movement tasks – particularly antisaccade paradigms, which require the suppression of a reflexive glance toward a stimulus – are highly sensitive to early cognitive decline in AD and FTD and can help distinguish between different neurodegenerative diseases, although practical clinical implementation is still evolving. ⁵ , ⁶ Eye tracking has been highlighted as less demanding cognitively and non‐verbal, offering advantages over traditional neuropsychological tests, especially for patients with communication barriers or low literacy. ⁷ Systematic reviews and meta‐analyses ⁸ , ⁹ , ¹⁰ , ¹¹ report that eye tracking achieves moderate to good diagnostic accuracy, with pooled sensitivity and specificity values in the range of 0.73 to 0.97 and 0.72 to 0.77, respectively. Notably, these systematic reviews and meta‐analyses comprise a reduced number of studies (12 studies, ⁸ nine studies, ⁹ 39 studies, ¹⁰ and 18 studies, ¹¹ respectively), while the present review comprises 71 studies.

Also, many of these reviews often synthesize findings without a critical appraisal of the wide‐ranging and often outdated diagnostic criteria across studies. This clinical heterogeneity is a critical limitation, as varied inclusion criteria (influenced by subjective symptomatology, cultural factors, and different cognitive test batteries) make direct comparisons and generalizations challenging. Furthermore, as our review highlights, existing studies have failed to incorporate modern biological definitions of AD. Therefore, the current review adds to the existing literature by not only qualitatively synthesizing results from saccadic tasks but also by critically analyzing the methodological gap left by the lack of biologically defined cohorts.

In addition, task design has emerged as a key factor. Naturalistic eye‐tracking tasks have been found to be more ecologically valid, simulating real‐world attention and memory demands more closely, and are also better tolerated by older adults compared to more artificial or repetitive testing formats. ¹⁰ Meanwhile, visual paired comparison (VPC) paradigms, which measure preferential gaze toward novel versus familiar images, have been shown to sensitively detect early attentional and mnemonic deficits. ¹⁰ , ¹¹

Despite this progress, dementia remains widely underdiagnosed, and diagnoses are often made only at advanced stages. With the prospect of disease‐modifying therapies on the horizon, there is an urgent need for accessible, cost‐effective tools that enable screening and intervention during the earliest phase of the disease. Against this backdrop, eye‐tracking technology stands out for its potential to aid diagnosis, making early and accurate assessment feasible on a global scale.

This systematic review, adhering to Preferred Reporting Items for Systematic Reviews and Meta‐Analyses (PRISMA) and Population, Intervention, Comparison, and Outcome (PICO) frameworks, aims to synthesize current evidence to answer the following question: In adults with AD, MCI, or dementia, which eye‐tracking tasks or metrics most effectively distinguish these conditions from healthy cognition? By clarifying the clinical value of specific paradigms and metrics, the current analysis aims to inform future research, diagnostic protocols, and clinical practice.

Notably, although the underlying motivation of the current review is to inform the development of early detection tools, its specific aim is to synthesize evidence on paradigms that effectively distinguish between diagnosed populations (e.g., AD and MCI) and healthy controls (HCs). Establishing this corpus of evidence is a necessary first step toward validating these tasks in preclinical screening.

2. MATERIALS AND METHODS

This systematic review followed a three‐step process: (1) define the research question and establish inclusion criteria using the PICO framework; (2) select relevant studies in accordance with PRISMA guidelines; and (3) extract, organize, and summarize key information from the included articles focusing on PICO elements.

2.1. Search strategy

A systematic search was performed across PubMed, ACM Digital Library, and Google Scholar databases to identify relevant studies published up to February 3, 2025. The search strategy combined psychiatric terms (AD, dementia, and MCI) with eye‐tracking keywords (eye tracking, eye movement, eye tracker, saccade, fixation, and gaze) using Boolean operators. The specific search strings used for each database are provided in Table S1.

2.2. Study selection

Search results were managed using Zotero (version 7.0.13). Duplicate records were removed initially within Zotero based on titles and subsequently using Python (version 3.11.10) based on digital object identifiers. The resulting reports then underwent an automated screening process (script available at https://github.com/gustavoveron/DADM_2025_Review). This script was designed to filter the high volume of out‐of‐scope reports retrieved by the initial broad search. Records were excluded if they were not in English or lacked an abstract. Following this, titles were required to contain at least one neurodegenerative disease term (i.e., “Alzheimer's,” “dementia,” or “mild cognitive impairment”) AND at least one eye‐tracking term (i.e., “eye tracking,” “saccade,” “fixation,” and so forth).

The reports that passed this automated filter were then manually assessed for eligibility. A primary reviewer assessed all reports, applying the following exclusion criteria: duplicates, reviews, case reports, anatomy or feasibility studies, studies focused on caregiver syndrome, illumination, writing/drawing, reading, sleep, non‐ocular fixation, or non‐ocular gaze. Studies were also excluded if they were coupled with other metrics (speech and electroencephalogram), lacked a control or patient group, lacked diagnosis information, investigated diseases other than dementia, or if the full text was unavailable. To ensure reliability, two additional reviewers independently assessed the eligibility of a random 20% subset of these reports.

2.3. Data extraction

Data extraction from the included studies was structured according to the PICO framework, including (1) Population: demographic details such as population size, age, and gender; (2) Intervention: detailed information on the eye‐tracking methodology, including the technology used, device specifications, stimuli presentation platforms, specific conducted tests, and measured metrics; (3) Comparison: diagnostic information, including methods, used cognitive tests, and diagnoses; and (4) Outcome: results focusing on eye‐tracking metrics, highlighting changes in metrics such as saccadic accuracy, latency, velocity, and fixation duration, among others.

Extracted data also included publication details from Zotero. Discrepancies or contradictory data were noted in the “Limitations” section. Extracted data were organized in a tabular format using Excel and analyzed in Python. The “Results” section focuses solely on eye‐tracking metrics and model performance, excluding data from other technologies or metrics. Eye‐tracking and diagnostic data were categorized based on frequency of mention (grouped if mentioned at least twice, otherwise categorized as “Other”). Similar terms were grouped (i.e., “duration,” “time,” and “onset” grouped as “fixation duration”). Ambiguous or unclear information was omitted (i.e., ambiguous demographics and poorly defined metrics). Group differences in eye‐tracking metrics are reported as “higher” or “lower” if statistically significant (p < 0.05), regardless of the specific statistical test used in each study.

2.4. Risk of bias and applicability concerns

The quality of the included studies was assessed using the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS‐2) tool. Given the nature of diagnostic test accuracy reviews and the specific focus on identifying effective eye‐tracking paradigms for dementia, some QUADAS‐2 domains and questions were adapted.

2.4.1. Patient selection

Risk of bias was assessed based on whether (1) appropriate inclusion/exclusion criteria were used (combining “consecutive or random sample” and “inappropriate exclusions”) and (2) significant demographic differences existed between groups (e.g., age, gender, and education) or if the control group was a biased sample (e.g., patient relatives). Applicability concerns were high if studies grouped multiple dementia pathologies or did not differentiate MCI subtypes (amnestic vs non‐amnestic). The question regarding case‐control design was not applicable, as all included studies used this design.

2.4.2. Index test

The standard QUADAS‐2 questions regarding blinding and pre‐specified thresholds were not applicable in this context. This review aimed to identify eye‐tracking paradigms that differentiate AD, MCI, and dementia groups from HCs; therefore, blinding to the reference standard is not feasible, and pre‐specified thresholds are not expected in this exploratory phase of research.

2.4.3. Reference standard

Risk of bias was assessed based on the reporting clarity of the diagnostic criteria used. High risk of bias was assigned if criteria were not reported, and unclear risk if criteria were referenced indirectly. Applicability concerns were high in both cases. Blinding to the index test was not applicable, as diagnosis is a prerequisite for inclusion.

2.4.4. Flow and timing

Risk of bias was assessed based on whether all participants received the eye‐tracking assessment and whether subgroups were appropriately considered in the analysis. The question regarding the interval between the index test and reference standard was not applicable, as no standardized minimum interval exists for dementia diagnosis.

2.5. Data synthesis

Data synthesis was conducted using a two‐tiered approach to accommodate the significant heterogeneity in reporting standards across the included literature. A narrative synthesis was performed for all studies providing directional outcomes, while a formal quantitative meta‐analysis was restricted to a subset of studies that reported sufficient quantitative data.

First, for the narrative synthesis, we extracted the qualitative outcome of comparisons between groups (e.g., AD vs HC). For each relevant metric, we recorded whether the study reported it as significantly higher, significantly lower, or not significantly different in the patient group compared to the control group. These findings were systematically tallied to identify trends and inconsistencies across the literature, and the results of this “vote‐counting” approach are visualized in summary plots. This method allowed for the inclusion of findings from studies that did not report precise statistics.

Second, for the quantitative meta‐analysis, we identified a subset of 15 studies (21%) that provided the necessary numerical data. For these studies, quantitative data were extracted, including sample size (n), mean, and standard deviation (SD), for each participant group (e.g., AD, MCI, and HC) and eye‐tracking metrics. In cases where studies reported the median and interquartile range (Q1 and Q3) instead of the mean and SD, ¹² , ¹³ these values were transformed to estimate the mean and SD using the validated methods described by Wan et al. ¹⁴

The primary outcome measure for each comparison was the standardized mean difference, calculated as Cohen's d. This was computed as the difference between two group means divided by the pooled standard deviation, which was derived from the square root of the average variance of the two groups. To synthesize findings across studies, individual effect sizes were combined using an inverse‐variance‐weighted meta‐analysis. This approach assigns greater weight to studies with more precise estimates (i.e., smaller standard errors), thereby yielding a more robust pooled effect. For each meta‐analysis, the overall pooled Cohen's d and its corresponding 95% confidence interval (CI) were calculated. A pooled effect was considered statistically significant if its 95% CI did not cross zero.

All data processing, statistical analyses, and figure generation were performed using custom scripts written in Python (version 3.11.11) within Spyder IDE (version 6.0.5). The Pandas library was utilized for data manipulation, NumPy for numerical computations, and Matplotlib for creating forest plots and other visualizations. Generative artificial intelligence models (Gemini 2.5 Pro, Claude Opus 4, and GPT‐4.1) were used to assist with script debugging and refinement and for verifying the grammatical accuracy of the manuscript.

3. RESULTS

3.1. Study identification

A comprehensive search across PubMed, ACM Digital Library, and Google Scholar (Table S1) identified 1421 reports related to eye tracking and dementia, including AD, MCI, and other dementias. After removing duplicates, 936 unique reports were identified. These reports underwent an automated screening process using a Python script to filter by language, abstract presence, and title keywords (Figure 1), which yielded 171 reports for manual assessment. These reports were assessed for eligibility based on detailed exclusion criteria (e.g., reviews, case reports, and reading studies). To ensure reliability, two additional reviewers independently assessed a 20% random subset (34/171, omitting duplicates), achieving perfect or almost perfect inter‐rater agreement (Reviewers 1 and 2, Cohen's Kappa = 1.0; Reviewers 1 and 3, Cohen's Kappa = 0.9). This manual screening process yielded 77 selected studies (Figure 1). However, six were excluded due to unreported eye‐tracking metrics, missing statistical comparisons, or exclusive focus on machine‐learning (ML) models that omitted the underlying metric data required for the primary analysis of this review, resulting in a final sample of 71 studies (Table S2).

FIGURE 1.

Preferred Reporting Items for Systematic Reviews and Meta‐Analyses (PRISMA) workflow. Diagram outlining identification, screening, retrieval, and selection processes. Records were sourced from PubMed, ACM Digital Library, and Google Scholar and resulted in 77 studies being included in the review based on the PRISMA workflow.

3.2. Study characteristics and methodological quality

The 71 included studies were predominantly from the United Kingdom (14%), the United States (13%), France (13%), and China (12%), covering all continents except Africa (Europe: 39.5%; Asia: 31%; North America: 15.5%; South America: 7%; and Oceania: 7%; Figure 2A). Most studies (59%) were published between 2015 and 2025, with a peak of eight publications in 2021 (Figure 2B).

FIGURE 2.

Characteristics of included studies and participants. (A) Geographical distribution of first‐author affiliations of the included studies, with color intensity representing the number of publications per country. (B) Temporal distribution of publications from 1985 to 2025. (C) Frequency distribution of total number of participants (patients and controls) per study. (D) Age distribution of participants, showing frequency of reported mean ages for control and patient groups across studies. (E) Gender predominance in studies. “Female predominance” indicates studies where females constituted over 60% of at least one group (patients or controls); “male predominance” indicates similar for males; “balanced” suggests roughly equal representation; and “imbalanced” for other unequal distributions. (F) Distribution of studies by primary neurodegenerative conditions investigated. (G) Diagnostic methods and criteria reported for Alzheimer's disease (AD) diagnosis in studies including AD patients. “Other papers” refers to studies using criteria mentioned less frequently. (H) Diagnostic methods and criteria reported for mild cognitive impairment (MCI) diagnosis in studies including MCI patients. “Other papers” refers to studies using criteria mentioned less frequently.

Risk of bias assessment revealed that 52% of studies had an uncertain risk of bias for patient selection, while risk was low for reference standards (58% of studies) and flow/timing (96% of studies). Applicability concerns were high or uncertain for patient selection in 69% of studies and for reference standards in 42% (Figure S1).

3.3. Participant characteristics

A majority of the studies (72%) included fewer than 50 patients across all reported conditions (Figure 2C). Most studies focused on a single patient group (53%), followed by two groups (41%) and more than two (6%). Female participants were overrepresented (over 60% of the sample) in at least one group (patients or controls) in half (50%) of the studies (Figure 2D). Participants’ mean age ranged from 38 to 79 for controls and 38 to 81 years for patients (Figure 2E). Regarding the diagnostic approach for AD, 85% of studies relied on clinical diagnosis, whereas 14% incorporated a biological approach, either alone (2%) or combined with clinical criteria (12%). The most frequently cited clinical criteria for AD diagnosis were National Institute of Neurological and Communicative Disorders and Stroke (NINCDS)/ Alzheimer's Disease and Related Disorders Association (ADRDA), National Institute on Aging (NIA)/Alzheimer's Association (AA), and the Petersen criteria for MCI (Figure 2G,H). Diagnostic guidelines were not explicitly reported in only 2% and 3% of AD and MCI studies, respectively. The majority of studies focused on AD, MCI, or both, with 27% examining AD exclusively, 13% MCI, 27% both, and 13% AD/MCI alongside other diseases (Figure 2F). Among studies including AD patients, 64% included AD or probable AD patients, while 22% included mild to moderate AD, 7% AD dementia, and 7% other variations (young‐onset AD [YOAD], participants with parents with AD, and AD/MCI patients without distinction). Regarding MCI patients, only 40% specified the MCI subtype (amnestic/non‐amnestic).

To further characterize these participant cohorts, various cognitive tests were employed. The Mini‐Mental State Examination (MMSE) was the most frequently used tool (reported in 52 studies), followed by the Trail Making Test (TMT; 22 studies), the Digit Span Memory Test (19 studies), Wechsler Memory Scale (WMS; 18 studies), and the Clinical Dementia Rating (CDR; 13 studies), among others (Table S2).

3.4. Eye‐tracking techniques and tasks

Most studies employed infrared‐based eye trackers (89%), with alternative technologies such as scleral‐coil or electrooculography used only in studies conducted before 2015. Common infrared eye trackers (Figure 3A) included devices from EyeLink, Tobii, and Applied Science Laboratories, allowing for monocular (27%) or binocular recordings (35%), with 38% not specifying the type. Stimuli were primarily presented on monitors, though some were projected or displayed on LED boards and dual‐screen devices (Figure 3B).

FIGURE 3.

Eye‐tracking technologies, tasks, and metrics. (A) Distribution of common infrared eye‐tracker brands used. “Others” includes less frequently mentioned brands or unspecified infrared devices. (B) Methods used for stimuli presentation during eye‐tracking tasks. (C) Common eye‐tracking paradigms employed, including prosaccade, antisaccade, fixation, smooth pursuit, visual paired comparison, search, and free viewing tasks. (D) Overall distribution of different eye‐tracking task categories employed across all included studies. (E) Distribution of eye‐tracking tasks specifically within studies investigating Alzheimer's disease (AD). (F) Distribution of eye‐tracking tasks specifically within studies investigating mild cognitive impairment (MCI). (G) Common metrics reported for prosaccade and antisaccade tasks. (H) Common metrics reported for non‐saccadic eye‐tracking tasks. “Other saccadic metrics” refers to saccadic parameters sometimes reported within these non‐saccadic tasks.

Eye‐tracking tasks were grouped as follows (Figure 3C):

1. Prosaccade tasks: Participants are required to rapidly shift their gaze toward a newly appearing peripheral target following an initial central fixation. These tasks assess basic saccadic function, including saccadic latency, velocity, and accuracy. Variations of the task include the “gap condition,” where the central fixation point disappears briefly before the peripheral target appears, and the “overlap condition,” in which the central fixation remains visible during target onset. Additionally, some studies report simultaneous offset of the central fixation and onset of the peripheral target, while others do not provide sufficient methodological detail; both have been categorized as undefined prosaccade tasks.

2. Antisaccade tasks: Participants are instructed to look in the direction opposite to the peripheral target when it appears. These tasks probe inhibitory control and executive function, as successfully performing an antisaccade task requires suppression of the reflexive response to look at the target and instead generating a voluntary saccade away from it. Variants often include gap and overlap conditions and may be administered sequentially with prosaccade tasks.

3. Fixation tasks: Participants are required to maintain their gaze steadily on a stationary target (such as a cross or dot) for a specified period, either in the center or at lateral positions.

4. Smooth pursuit tasks: Participants are asked to follow a moving target that travels smoothly across the visual field, typically along horizontal or vertical trajectories. Unlike saccadic tasks, which measure rapid eye jumps, smooth pursuit tasks assess the eyes' ability to maintain continuous, accurate tracking of a moving object.

5. VPC: These tasks assess visual recognition memory and novelty preference. During the familiarization phase, participants view two identical images. In the subsequent test phase, one familiar image is paired with a novel one. Preference is inferred from longer fixation durations on the novel stimulus.

6. Search tasks: Participants engage with naturalistic or multistimulus scenes, following specific instructions involving memory, categorization, reasoning, or other cognitive operations – distinct from basic saccadic paradigms. This group includes (F1) memory‐dependent targeting and recall tasks, assessing memory‐guided gaze, such as saccades to previously indicated locations or objects or detection of target objects in scenes after a brief interval or delay; (F2) visual search and feature discrimination, requiring locating a specific object or feature among distractors following a cue, such as Landolt ring searches, emotion recognition from faces or eyes, category or semantic matching across displays, detection of differences in visually similar figures, geometric/abstract forms, scene‐based searches, or sequential navigation ¹⁵ , ¹⁶ , ¹⁷ , ¹⁸ , ¹⁹ ; (F3) multiple/hybrid cognitive eye tracking, aggregating several cognitive domains in a modular battery, such as memory encoding/recall, visual attention, calculation, and working memory, or color/shape matching after familiarization ²⁰ , ²¹ ; and (F4) real‐world navigation and complex interaction, encompassing real or virtual navigation tasks, such as interacting with websites or forums or exploring virtual environments with wearable eye tracking. ²²

7. Free‐view tasks: Participants are instructed simply to observe visual stimuli without any explicit goal (search, memory, categorization, or directed motor action). Subgroups include (G1) passive viewing of faces, where participants passively view faces displaying various emotions, focus on facial features, view neutral, or artificial intelligence (AI)‐generated faces ²³ ; (G2) passive scene and object exploration, involving observation of complex scenes or objects without task demands ²⁴ , ²⁵ , ²⁶ ; and (G3) free viewing in personal/cognitive or virtual action contexts, including tasks with visual cues linked to autobiographical events, interacting with a virtual kiosk, or reproducing figures using numbered sticks. ²⁷

The most frequently used eye‐tracking tests included saccadic tasks (prosaccade or antisaccade tasks) and search tasks (Figure 3D). These varied by disease, with over 50% of tests in Alzheimer's studies involving saccadic tasks (Figure 3E), while MCI studies included both saccadic and search tasks in most cases (Figure 3F). Although there is currently a lack of consensus on measurement standards or nomenclature, saccadic test metrics (Figure 3G) were generally categorized into five domains: (1) accuracy (correct execution of a saccade toward or away from the target), (2) latency (time taken to initiate a saccade), (3) velocity (speed of the saccade), (4) amplitude (distance from the starting point to the end point), and (5) gain (ratio between saccade amplitude and target distance). Non‐saccadic test metrics (Figure 3H) included fixation duration (time spent on a specific area of interest), number of fixations, area of interest, and sometimes overlapping saccadic parameters such as latency, accuracy, velocity, and number.

3.5. Eye tracking in MCI

Although 40% of MCI studies included a specific MCI subtype (aMCI/naMCI), the remaining ones included MCI patients with memory impairment subjects or no specifications. Hence, the following analysis was carried out considering all subtypes and diagnoses. This group of 30 studies (Figure 3F) included 12 prosaccade tasks, 11 antisaccade tasks, nine search tasks, three fixation and free‐view tasks, two VPC tasks, one smooth pursuit task, and five “Others” (go/no‐go, visual short‐term memory binding, oddball, and decision tasks). The aMCI subgroup (10 studies) included six prosaccade tasks, four antisaccade tasks, and only one each for search, free‐view, fixation, and decision tasks, which will not be discussed due to their small number.

3.5.1. Saccadic tasks in MCI

In studies including MCI patients, performance on gap and overlap prosaccade tasks was largely comparable to that of controls (Figure 4, left panel; Table S3). Specifically, prosaccadic latency, accuracy, and velocity showed no significant group differences (Figure 5, Figures S2–S3). The only exception was a lower gain in gap prosaccade tasks in the MCI group ²⁸ , ²⁹ (Figure 5).

FIGURE 4.

Summary of eye‐tracking metric changes. Each panel displays the percentage of studies reporting statistically significantly (p < 0.05) higher, lower, or no significant differences (similar) in specific saccadic metrics for the first group listed compared to the second. It is important to take into account that the different metrics were not necessarily measured in the same way across tasks (rows) or compared groups (columns).

FIGURE 5.

Gap prosaccade tasks forest plot. Forest plot summarizing the meta‐analytic results for gap prosaccade task metrics. Each subplot presents the standardized mean difference (Cohen's d) for accuracy, latency, velocity, gain, and amplitude. Individual study effect sizes are represented by squares (dark purple for significant group differences and pale purple for non‐significant differences), with square size proportional to the study's weight and horizontal lines indicating the 95% confidence intervals. The overall pooled effect is shown as a purple diamond.

In contrast, in undefined prosaccade tasks, patients with MCI consistently showed lower accuracy and gain, alongside increased latency, compared to controls. ³⁰ , ³¹ , ³² Furthermore, antisaccade tasks resulted in lower accuracy and longer latencies in approximately half of the studies ¹² , ²⁸ , ²⁹ , ³⁰ , ³¹ , ³² , ³³ , ³⁴ (Figure 6).

FIGURE 6.

Antisaccade tasks forest plot. Forest plot summarizing the meta‐analytic results for antisaccade task metrics. Each subplot presents the standardized mean difference (Cohen's d) for accuracy, latency, velocity, gain, and amplitude. Individual study effect sizes are represented by squares (dark purple for significant group differences and pale purple for non‐significant differences), with square size proportional to the study's weight and horizontal lines indicating the 95% confidence intervals. The overall pooled effect is shown as a purple diamond.

3.5.2. Other tasks in MCI

In search tasks, MCI patients showed decreased accuracy ¹² and contradictory results regarding the duration and number of fixations, remaining comparable, ¹⁵ , ¹⁶ , ²⁰ , ²² decreased, ²¹ or increased ¹⁷ compared to HCs, probably due to the task design, which included websites, ²² drawings, ¹⁵ numbers and letters, ¹⁶ or several of these. ¹⁷ , ²¹ In line with this heterogeneity, when grouping by closely related tasks such as free viewing, MCI patients showed decreased fixation duration while exploring a virtual kiosk or recreating stick structures. ²⁷ Conversely, fixation duration increased while performing central (centered dot) or lateral (right, left, top, and bottom dot) fixation tasks. ¹² , ¹³

3.6. Eye tracking in AD

Our analysis revealed a diverse range of eye‐tracking tasks employed across studies involving patients with AD (Figure 3E). Specifically, the tasks included prosaccade tasks (n = 21 studies), antisaccade tasks (n = 16), visual search tasks (n = 10), free‐viewing tasks (n = 5), smooth pursuit and fixation tasks (n = 4), and a VPC task (n = 1). Additionally, five studies used other task types, such as oddball, prediction/decision, and go/no‐go tasks.

3.6.1. Saccadic tasks in AD

When compared to controls, patients with AD consistently depicted impaired saccadic performance across both pro‐ and antisaccadic tasks (Figure 4, middle column; Table S3), exhibiting higher latency and lower accuracy. ¹³ , ²⁵ , ³⁵ , ³⁶ However, other saccadic metrics varied depending on the specific task. In prosaccade tasks, saccadic velocity, amplitude, and gain were generally comparable between AD patients and controls ⁷ , ²⁵ , ²⁹ , ³⁵ , ³⁶ , ³⁷ , ³⁸ (Figure 5, Figures S2–S3). In contrast, for antisaccade tasks, AD patients exhibited significantly lower velocity and gain ⁷ , ¹² , ²⁸ , ³⁹ (Figure 6).

3.6.2. Other tasks in AD

In non‐saccadic tasks, patients with AD consistently demonstrated altered visual exploration. Across a variety of paradigms – including search tasks requiring the exploration of S‐shaped figures, geometrical shapes, Landolt rings, or scenes ¹² , ¹⁵ , ¹⁷ , ¹⁹ , ²⁰ , ²¹ , ⁴⁰ ; free‐viewing of stimuli like complex scenes, AI‐generated faces, or personal images ²³ , ²⁴ , ²⁵ , ²⁶ ; and VPC ⁴¹ – patients typically showed a decreased area of interest. In addition, they also exhibited a reduced duration and number of fixations on targets, often coupled with an increased focus on distractors. ¹⁵ , ¹⁷ , ¹⁹ , ²⁰ , ²¹ , ²⁴ , ²⁵ Despite these exploratory deficits, saccadic latency and amplitude during these tasks often remained similar to controls. A few studies reported similar fixation duration when compared to controls. ¹⁵ , ¹⁹ , ²³ , ⁴⁰ In a separate finding, one study also reported similar velocity and number of saccades in smooth pursuit tasks between controls and patients. ²⁵

3.7. Eye tracking in AD and MCI

As previously mentioned, 27% of included studies investigated both AD and MCI patient groups. Within these studies, prosaccade tasks were the most frequently employed (n = 9), followed by antisaccade tasks (n = 8), search tasks (n = 6), fixation tasks (n = 3), with a smaller number including free‐view (n = 1), smooth pursuit (n = 1), and “other” tasks (n = 2) such as decision‐making and oddball paradigms.

3.7.1. Saccadic tasks in AD and MCI

When directly comparing AD and MCI groups in prosaccade tasks, most studies found no significant differences in accuracy, velocity, and gain (Figure 4, right column; Table S3). Findings on prosaccadic latency, however, were inconsistent. Several studies reported no significant differences between groups, ¹² , ²⁸ , ²⁹ , ³³ , ⁴² while others described significantly increased latency in AD patients ¹³ , ³⁵ , ³⁶ , ⁴³ (Figure 5, Figures S2–S3).

In antisaccade tasks, a clearer pattern of impairment emerged, distinguishing AD from MCI. The majority of studies reported that, compared to the MCI group, AD patients had decreased accuracy and gain. ¹² , ¹³ , ²⁸ , ²⁹ , ⁴² , ⁴³ Saccadic velocity was generally similar between the groups, while latency in AD patients was either comparable to ¹² , ¹³ , ³³ , ³⁴ or higher than ²⁸ , ⁴² , ⁴³ that in the MCI group (Figure 6).

Only two studies directly compared both naMCI and aMCI patients. ³⁰ , ³⁴ Koçoğlu et al.’s study included both pro‐ and antisaccade tasks, reporting an increase in latency in aMCI patients when compared to controls, but comparable results when compared with naMCI patients. Wilcockson et al. also included AD patients among their cohorts, reporting increased latency and decreased accuracy in both AD or aMCI patients compared with controls and naMCI patients.

3.7.2. Other tasks in AD and MCI

AD patients depicted an increased number and duration of fixations on distractors, or a reduction on targets, ¹⁵ , ²⁰ , ²¹ decreased accuracy, and comparable latency ¹² in search tasks when compared to MCI patients. However, contradictory results were found in fixation tasks, with reports indicating similar fixation accuracy and duration (time the eyes remain focused on a target before deviating by >4°) ¹³ or decreased accuracy and increased duration. ¹² Although both tasks were similar (central and lateral fixations), discrepancies may be due to sample size, diagnostic criteria, eye tracker, or aMCI/naMCI patients ratio (not addressed in either study).

3.8. Eye tracking in other pathologies or stages

Twenty‐one percent of studies examined eye tracking in conditions other than classical AD or MCI. Half of these focused on FTD, including behavioral variant FTD (bvFTD), ¹⁸ , ⁴⁴ , ⁴⁵ FTD with motor neuron disease (FTD MND), ⁴⁶ or FTD without specified subtype (nsFTD). ⁴⁷ Most studies in bvFTD employed prosaccade tasks, reporting prosaccade latency, velocity, and amplitude similar to controls. ¹⁸ , ⁴⁴ , ⁴⁵ In contrast, FTD MND patients showed decreased prosaccadic velocity and increased amplitude, while nsFTD studies revealed higher prosaccadic latency. Additionally, decreased antisaccadic accuracy was noted in select studies ⁴⁵ , ⁴⁷

YOAD patients demonstrated more large intrusive saccades and shorter fixation durations than controls during fixation tasks; during prosaccade tasks, they exhibited lower accuracy, increased latency, and required more corrective saccades. In smooth pursuit tasks, while pursuit gain was preserved, YOAD patients spent significantly less time smoothly following the target. ⁴⁸

In familial AD, symptomatic mutation carriers (SMCs) displayed distinct eye‐tracking profiles compared to both presymptomatic mutation carriers (PMCs) and controls during visual search tasks, spending less time fixating on stimuli. Both early and late PMCs, however, exhibited fixation durations comparable to controls. ⁴⁹

Finally, among middle‐aged adults with a family history of AD (FH+), saccadic eye‐tracking tasks revealed some impairments relative to matched controls (FH−): The FH+ group made fewer uninhibited reflexive saccades and showed reduced average saccade velocity, despite similar cognitive test scores and demographics. ⁵⁰

3.9. Eye‐tracking ML models

While studies focusing solely on ML classification performance were excluded from the primary analysis, six of the selected papers also leveraged their reported eye‐tracking metrics to develop ML models. This section provides an overview of these approaches to differentiate patient groups from HCs. These studies focused on AD, ¹⁷ , ⁴¹ , ⁴⁸ MCI, ²⁷ , ³² and FTD. ⁴⁴ A variety of eye‐tracking hardware was employed, including EyeLink II, ⁴⁴ , ⁴⁸ Tobii TX300, ¹⁷ a self‐designed three‐dimensional eye‐tracker, ⁴¹ VIVE Pro Eye, ²⁷ and Tobii Pro AB. ³²

3.9.1. Eye‐tracking tasks and feature engineering

The ML models were built upon features extracted from diverse eye‐tracking tasks tailored to the specific pathologies: AD‐focused studies used features from fixation, gap prosaccade, and smooth pursuit tasks ⁴⁸ ; a visual search task involving geometric/abstract forms and scenes with familiarization and recognition phases ¹⁷ ; and a VPC task with image pairs containing missing or additional elements. ⁴¹ MCI‐focused studies drew features from a free‐view task in a virtual environment (ordering a menu item at a kiosk) ²⁷ and a battery of saccadic tasks, including prosaccade, antisaccade, and go/no‐go paradigms. ³² The FTD study employed a variation of prosaccade tasks, requiring participants to follow newly appearing circles along a straight or zigzag path. ⁴⁴

The approaches to feature selection and engineering were notably varied: Most of them extracted a comprehensive set of eye movement features from the task. ¹⁷ , ²⁷ , ³² , ⁴⁸ Additionally, Pereira et al. employed multiple feature selection algorithms, including Pearson's correlation, sequential feature selection, Las Vegas Weight, and genetic algorithm. ¹⁷ Sun et al. generated fixation heatmaps and derived features using Fourier coefficients (FOU) and Karhunen‐Loeve coefficients (KAR) for traditional ML, while also using VGG‐16 and ResNet‐18 networks for deep learning feature extraction. ⁴¹ Primativo et al., in contrast to classification, used raw x‐ and y‐gaze coordinates (250 Hz) to build a predictive model for the next gaze location rather than for disease differentiation. ⁴⁴

3.9.2. ML algorithms and reported performance

Sample sizes varied from 51 ⁴⁴ to 594, ³² with most of the studies close to the lower limit. Thus, the classification was typically based on classical ML methods like support vector machine (SVM), random forest (RF), K‐nearest neighbors (KNN), logistic regression (LR), and extreme gradient boosting (XGB). Pereira et al. also evaluated a neural network approach with different feature selection methods. Nevertheless, their best‐performing model was a RF combined with the Las Vegas Weight feature selection. Sun et al. also compared classical models with a novel deep learning model, NeAE‐Eye. This model, featuring an inner autoencoder for image‐dependent representation from fixation density maps and an outer autoencoder with a classifier, achieved the study's highest performance.

Most of the studies report cross‐validation results, without assuring a held‐out set. Although not ideal, this is understandable due to the low sample size and the difficulty of stratifying the data based on several variables – age, sex, years of education, and diagnosis – because many studies include more than one pathology. In addition, hyperparameter tuning is often performed on the same dataset used for both training and validation, which can lead to overly optimistic performance estimates. Best practices recommend employing nested cross‐validation to avoid this bias. Most of the studies report several performance measures (area under the receiver operating characteristic [ROC] curve [AUROC], accuracy, precision, recall, and F1‐score), but there were studies with poor details on the performance measures and, for instance, how they summarize the results across validation sets.

Overall, the studies presented high performance with accuracies between 0.72 and 0.97, AUROC over 0.79, and precision, recall, and F1‐score over 0.70. Notably, some studies report accuracies or AUROC over 95%. Nevertheless, it is difficult to compare between approaches because of the differences in pathologies (and diagnostic criteria within pathologies), train‐test data segmentation, and reporting criteria. Additionally, some studies include participants with visual impairments, which can introduce further variability and confound comparisons across models. These factors, combined with small sample sizes or lack of external validation in some cases, can elicit overly optimistic results and limit the generalizability of the findings.

4. DISCUSSION

4.1. Updates on AD diagnosis

The paradigm for diagnosing AD is undergoing a significant transformation, driven by the development of highly accurate blood‐based markers and the regulatory approval of treatments targeting its core neuropathology. This shift has reframed AD diagnostic and staging criteria, establishing that AD is currently defined by its underlying biology; consequently, the detection of AD neuropathologic changes through biomarkers is now considered equivalent to diagnosing the disease itself. ³ This biological definition has also consolidated the conception of AD as a continuum, where the disease first becomes evident with the appearance of disease‐specific biomarkers, often while individuals are still asymptomatic. Although the precise pathophysiologic mechanisms involved in the early processing and clearance of pathogenic protein fragments remain under investigation, core AD biomarkers (Aβ proteinopathy and phosphorylated and secreted AD tau) are central to this new framework. Both fluid and imaging biomarkers – that is, amyloid positron emission tomography (PET), validated cerebrospinal fluid (CSF), or plasma assays – can fulfill diagnostic criteria; thus, abnormality on specific core biomarkers are sufficient to diagnose AD.

This biologically grounded definition seamlessly informs the current approach to AD staging. Six clinically defined stages now span from (1) biomarker evidence of AD in asymptomatic individuals, through (2) transitional decline (earliest detectable clinical symptoms in cognitively unimpaired individuals) and (3) objective cognitive impairment without significant functional loss, to stages 4 to 6, which represent progressively worsening functional loss corresponding to mild, moderate, and severe dementia, respectively. Notably, the term “prodromal AD/MCI,” previously used for individuals with abnormal AD biomarkers and impairment short of dementia, is now encompassed within “AD clinical stage 3,” emphasizing that such individuals have the disease, not merely a prodrome.

4.2. Eye tracking and AD screening

The imperative for early intervention in AD is widely recognized, as therapeutic strategies are often ineffectual in later stages due to irreversible neuronal damage, whereas early action might decelerate disease progression. Although current methods like CSF biomarkers and PET imaging allow for early detection of neuropathology, their utility for widespread screening is hampered by invasiveness, high cost, and limited accessibility, ⁴ which underscores the need for more accessible early detection tools.

Emerging evidence suggests that attention is among the first non‐memory domains affected in AD. Given that attention and oculomotor control are thought to recruit overlapping brain regions, saccadic eye movements are likely to be disturbed by the reductions in inhibitory control and executive function characteristic of neurodegenerative disorders. Consequently, the utility of relatively low‐cost eye‐tracking technologies for distinguishing various neurodegenerative conditions from healthy aging has raised considerable interest. ¹⁰ Eye movement assessment offers several advantages in neurological practice: It is a non‐invasive bedside test, comparatively quicker than standard cognitive assessments, modern eye‐tracking technology provides precise quantitative metrics of oculomotor disturbance and enhances diagnostic value. ⁵

4.3. Methodological considerations in reviewed eye‐tracking literature

The current body of literature on eye tracking in dementia, as represented by the studies included in this review, presents several methodological considerations that may influence the interpretation of findings. The majority of studies (86%) originated from Europe, Asia, or North America, indicating a potential cultural and ethnic bias in the described results. This geographical skew is particularly concerning given that the global burden of AD and other dementias is rising sharply, with over two‐thirds of affected individuals residing in LMICs. ⁵¹

Beyond geographical distribution, the risk of bias assessment revealed that approximately half of the included studies exhibited an uncertain risk concerning patient selection. This often stemmed from factors such as gender overrepresentation in study groups, significant age differences between patient and control groups, and, in some instances, the recruitment of patients' relatives or spouses as controls. While the latter approach might facilitate recruitment and matching on certain demographic variables, it introduces potential confounding due to shared environmental exposures, caregiving burden, or genetic predispositions. Consequently, this methodological choice raises concerns about selection bias and limits the comparability between case and control groups.

An additional methodological consideration is the lack of assessment for early childhood trauma. This factor should be considered in future research to help clarify the underlying source of neurocognitive dysregulation, as child maltreatment is known to alter visual attention patterns, ⁵² and adverse childhood experiences are associated with an increased risk for dementia. ⁵³

Compounding these issues, the diagnostic approaches for AD and MCI varied widely across the reviewed studies. While many employed well‐established criteria such as NINCDS/ADRDA, NIA/AA, or Petersen criteria, others relied on more subjective, less standardized, or undisclosed benchmarks, including references to other papers, specific combinations of parameters, or non‐explicit methodologies. This heterogeneity in diagnostic rigor is critical because the specificity and sensitivity of these differing criteria can be influenced by reliance on subjective symptoms, cultural factors inherent in diverse study populations, and the wide variety of cognitive tests used for assessment. Furthermore, it is critical to note that none of the included studies fully aligned their diagnostic criteria with the most recent biological definition of AD. This lack of adherence to current biological definitions, coupled with the varied clinical diagnostic approaches, likely impacted the homogeneity of patient cohorts and limit the direct applicability of many findings to currently defined disease stages. Consequently, future meta‐analyses or systematic comparisons should carefully consider the impact of differing diagnostic criteria, potentially by excluding studies with less rigorous methodologies or by performing subgroup analyses to avoid diluting the strength of findings from more robustly characterized cohorts.

4.4. Saccadic task performance in AD and MCI

Saccadic eye movements, fundamental for rapidly shifting gaze, have been extensively studied in dementia using prosaccade and antisaccade tasks. ⁵ , ¹¹ Prosaccade tasks, which assess basic saccadic function through gap, overlap, or simultaneous stimuli presentation, yielded varied findings, particularly for AD patients. While most studies reported similar prosaccadic velocity, gain, and amplitude values between AD patients and older controls, increased latency was noted in ∼60% of studies using gap and overlap paradigms (and 100% for undefined tasks).

Prosaccadic accuracy showed conflicting results: Approximately half of the studies reported lower results than older controls, ¹² , ³³ , ³⁸ , ⁴² while the other half found similar values. ¹³ , ³⁵ , ³⁶ , ³⁷ A closer examination revealed that these discrepancies might stem from varied definitions of “accuracy.” For instance, some studies defined it as the percentage of successful saccades toward the target, ¹² , ¹³ , ³⁷ , ³⁸ others as the percentage of uncorrected saccades or direction errors, ³³ , ⁴² and two as the primary saccade amplitude relative to target eccentricity. ³⁵ , ³⁶ This definitional heterogeneity likely contributed to the divergent findings, suggesting that studies reporting similar values to controls might not fully capture all aspects of performance accuracy.

In MCI patients, prosaccade metrics generally showed results similar to controls in terms of accuracy, velocity, and gain in standard gap/overlap tasks. However, undefined prosaccade tasks (encompassing simultaneous presentation or undisclosed methods) more consistently reported decreased accuracy ¹² , ³³ , ⁴² in MCI, highlighting that specific paradigm variations might be more sensitive to subtle deficits in this population and warrant further exploration beyond conventional gap/overlap designs.

Prosaccadic latency in MCI patients, however, presented a more complex and inconsistent pattern across studies. Such discrepancies in a temporally sensitive measure underscore the need for careful consideration of methodological factors. ⁵⁴ , ⁵⁵ These include the diversity of stimuli presentation media – ranging from regular computer screens ²⁸ , ³⁵ , ³⁶ , ⁴² to specialized screens ³³ and dual‐screen displays ¹³ – and the variety of eye trackers employed (e.g., IRIS, EyeKnow, SMI RED, Eyeseecam, Grass Technologies P18). Noteworthy, saccadic latency determination is inherently susceptible to hardware and software parameters like sampling rate, temporal resolution, and proprietary saccade‐detection algorithms.

When comparing AD and MCI patients directly, prosaccadic tasks (gap or overlap) appeared suboptimal for differentiation, as both groups generally depicted similar results across most metrics. Only latency showed conflicting results, likely influenced by the aforementioned metric sensitivity, variations in patient diagnostic criteria (e.g., mild to moderate AD vs aMCI ³³ , ³⁵ , ³⁶ or unspecified AD severity ¹³ , ²⁸ , ⁴² ), and the range of eye‐tracking technologies used, including infrared systems ¹³ , ²⁸ , ³⁵ , ³⁶ , ⁴² and electrooculography. ³³

The observed increase in prosaccadic latency in AD aligns with literature associating it with impaired function of cortical oculomotor areas, particularly parietal and frontal regions. ³⁵ , ³⁶ Yang further suggested that increased latency, even in healthy aging, is attributable more to cortical degeneration than peripheral factors, positioning saccade‐latency increase as an indicator of cortical dysfunction. ³⁵ The relative lack of consistent latency abnormality in the reviewed MCI studies might reflect less advanced cortical pathology compared to AD, consistent with typically milder biological findings in MCI. ³⁶

In contrast to the more mixed findings with prosaccades, our review suggests that antisaccade tasks, which demand complex attention shifting, decision‐making, and robust inhibitory control, ⁹ , ⁴² hold considerable promise for identifying oculomotor dysfunction in dementia. These tasks appear sensitive for differentiating AD patients, including those across various stages of the disease continuum, from HCs, and potentially for tracking disease progression. Across studies, individuals with AD consistently demonstrated impaired antisaccade performance, with most reporting decreased accuracy and gain, accompanied by increased latency. Similar patterns, though often less pronounced, were observed in MCI patients, who also frequently showed decreased accuracy and increased latency compared to controls. Importantly, when AD patients were directly compared to MCI patients, the former typically exhibited even lower antisaccadic metrics, particularly for accuracy, highlighting the task's potential to reflect the degenerative course of the disease.

As with prosaccades, while most studies involving AD patients depicted an increase in antisaccadic latency, findings in MCI patients were more contradictory. This again suggests that latency, as a temporally sensitive measure, may be significantly influenced by the selected eye‐tracking technology and specific experimental parameters, making direct comparisons challenging without strict methodological standardization. ⁵⁴ , ⁵⁵ Antisaccadic velocity was consistently found to be decreased in AD patients when compared to normotypical older controls and, to a lesser extent, when AD was compared with MCI patients; this latter observation is consistent with MCI patients often depicting velocities similar to controls in antisaccade tasks. ¹² , ¹³

Structurally, the error rate in the antisaccade task has been associated with the integrity of the frontal eye fields and frontoparietal cortex. ⁴³ The superior frontal gyrus, encompassing key oculomotor network nodes, has shown correlations between gray matter volume and antisaccadic performance in normal elderly individuals, as well as in patients with AD dementia and frontotemporal degeneration. This suggests that alterations in antisaccade performance likely reflect damage to the frontoparietal network crucial for voluntary saccade control and executive functions.

4.5. Non‐saccadic tasks in AD and MCI

Beyond basic saccadic paradigms, non‐saccadic tasks encompass a wide array of approaches, from goal‐driven activities following specific instructions (e.g., visual search and VPC) to passive, instruction‐less observation of various visual stimuli (e.g., free viewing). These tasks reveal distinct patterns in dementia: Most studies indicate that AD patients exhibit a diminished exploratory drive when encountering novel visual stimuli, manifested as a decreased exploration area and duration (e.g., reduced area of interest, shorter fixation duration, or fewer fixations on targets, sometimes with increased fixation on distractors). In contrast, findings in MCI patients were largely inconsistent across non‐saccadic tasks, likely reflecting the wide variety of paradigms employed and the inherent heterogeneity of the MCI population itself.

While non‐saccadic eye‐tracking tasks offer the advantage of assessing an individual's cognitive function during more naturalistic activities – which inherently require more complex cognitive interactions than traditional, constrained eye‐tracking paradigms – their diversity poses significant challenges for standardization and cross‐study comparison. As Readman noted in a review of naturalistic eye‐tracking paradigms, while diagnostic tests for AD and MCI could occur in naturalistic environments (e.g., an individual's home), they are more likely to be conducted in clinical settings. ¹⁰ A critical distinction should be made in the literature between tasks that use naturalistic stimuli or goal‐directed paradigms within a lab‐based setting and tasks that are truly naturalistic, occurring in real‐world settings. Recognizing this distinction, it becomes clear that further research analyzing eye movements during genuinely naturalistic tasks in individuals with AD and MCI is required.

Compared with eye movements under structured tasks with simple artificial stimuli, spontaneous or voluntary eye movement in complex environments is thought to involve a wider range of brain regions. These include not only primary visual and oculomotor areas but also higher cortical areas such as the temporal cortex for processing stimulus structures like objects or scenes and the frontal cortex for top‐down attentional guidance. From this perspective, as Yamada suggests, a multifaceted characterization of spontaneous gaze allocations to naturalistic complex scenes may provide critical insights into how these allocations relate to the pathophysiology of neurodegenerative dementias. ²⁶ Such tasks inherently test a combination of executive function, memory, and attention, all of which can be early indicators of AD. For example, Tamaru noted that considerable attention is required to accurately recognize complex shapes, and a higher number of visual distractors needed longer target fixation times, implying that individuals with AD might spend more time on gaze fixation due to diminished attentional function. ⁵⁶

Altered patterns of eye movements during the exploration of complex, real‐world scenes could offer a critical basis for understanding and potentially mitigating the widespread impact of altered visual attention on activities of daily living. This understanding could inform environmental modifications aimed at enhancing attentional processing and improving daily task performance, such as the provision of visually salient cues in object identification and wayfinding for individuals with dementia. ²⁶

4.6. Research gaps and methodological imperatives for future studies

This comprehensive systematic review, while scrutinizing past and recent research on eye tracking in neurodegenerative disease assessment, also highlights several critical limitations and research gaps that hinder direct comparability across studies and the broader applicability of findings. First, the evolving diagnostic landscape for AD and MCI over recent decades has inherently affected patient group homogeneity and the consistency of inclusion/exclusion criteria across the reviewed literature. The adoption of the most recent AD diagnostic guidelines, grounded in biological markers, ³ is anticipated to pave the way for more rigorous, comparable studies that can more accurately distinguish individuals across the AD continuum. In this context, eye tracking emerges as a promising, non‐invasive addition. It could potentially serve as an accessible screening mechanism to raise awareness, facilitate earlier engagement with care, and help identify individuals who might benefit most from definitive biomarker assessment. However, realizing this potential requires addressing important concerns regarding the standardization, equitable access, and democratization of these eye‐tracking tools, especially for vulnerable or historically excluded populations.

Second, the wide variety of eye‐tracking tasks and paradigms presented in this review underscores a pressing need for greater standardization. Future efforts should aim to develop robust, standardized tasks that perhaps combine the well‐characterized metrics from established paradigms (like saccadic tasks) with the ecological validity of more naturalistic approaches that better reflect daily activities and engage higher cortical areas. Such hybrid paradigms might be more sensitive in capturing subtle gaze‐movement differences between patient groups and neurotypical individuals. Moreover, the positive participant feedback noted in some studies using naturalistic tasks, with participants expressing enjoyment and interest in repetition, ²⁷ hints at the potential for eye‐tracking tasks to be adapted into digital therapeutic tools for cognitive rehabilitation. Given the consistent reports of impaired antisaccade performance in AD patients, which underscores the robustness of certain paradigms for neurodegenerative screening, these tasks could form a foundational component of an initial assessment battery, leveraging information retrieved from diverse technologies.

Third, a significant limitation evident from this review is the lack of global representation in eye‐tracking research. The selection algorithm did not retrieve studies involving African populations, and reports from Latin American and many Asian countries were scarce. This is a critical oversight, as the majority of new dementia cases annually occur in LMICs. ¹ Dementia is frequently underrecognized and undermanaged in these regions, where stigma and lack of awareness contribute to low diagnosis rates and inadequate care. Limited healthcare capacity, insufficient policy attention, and low public awareness in LMICs compound this issue, posing a substantial public health and socioeconomic challenge that future eye‐tracking research must strive to address through more inclusive and globally representative studies. ⁵¹

Finally, this review identified the wide variety of eye‐tracking hardware employed across studies as a potential source of variability. These included not only different infrared brands but also older technologies like scleral coil and electrooculography, each potentially possessing different sensitivities in gaze detection. While modern infrared eye trackers allow for highly precise measurement of gaze, as the pupil–iris contrast is clearer in the infrared spectrum, the field must also consider emerging approaches aimed at enhancing accessibility and wider use.

4.7. The potential of web‐based eye tracking and integration with ML

While none of the studies included in the primary analysis of this review employed remote, web‐based eye tracking, its growing potential warrants discussion. Two studies by Greenaway et al., which investigated remote eye tracking in AD patients and older controls using built‐in laptop webcams and the Gorilla platform, were initially identified but excluded from our main analysis due to missing statistical comparisons or unreported eye‐tracking metrics. The first of these focused on feasibility, reporting metrics like pixel error during calibration/validation and success rates. ⁵⁷ While generally feasible, calibration failures were noted, particularly in older adults, with eye‐watering identified as a potential confounder. The authors recommended ample setup time, large video feeds, and off‐screen assistants to improve data quality. However, this initial feasibility study did not include direct comparisons between AD and control populations.

The second study by Greenaway et al. used a modified dot‐probe task, which involved the presentation of a central fixation cross followed by two faces from the same actor to the either side of the cross, presented in sad‐angry, sad‐happy, sad‐neutral, angry‐happy, angry‐neutral, and happy‐neutral facial emotion pairings. Once the faces had disappeared, a black dot appeared in the center of one of the faces’ previous locations. Participants were instructed to look at the cross and the dot quickly and fixate on them and to naturally view the facial stimuli when presented. By measuring dwell time on different expressions, they observed a general trend away from sad faces and toward angry and happy expressions. ⁵⁸

Notably, a positive bias toward happy faces (compared to sad) was primarily driven by participants concurrently experiencing anxiety and/or depression, underscoring the importance of considering mood and individual differences in such assessments. This work highlights the potential of web‐based eye tracking for assessing emotional processing in AD using relatively simple tasks that only require differentiation of gaze to the left or right screen half. Future research could expand on these paradigms, for instance, by adapting robust tasks like the antisaccade task, which has shown promise in web‐based formats with neurotypical young adults, ⁵⁹ in parallel with ongoing efforts to improve the precision of web eye‐tracking algorithms and the development of novel hardware.

The advent of web eye trackers opens possibilities for remote disease assessment. While eye tracking in isolation is unlikely to become a primary determinant of clinical decision‐making, its integration with ML approaches may add significant value. This could aid in detecting subtle changes in at‐risk and presymptomatic individuals, monitoring disease progression over time (especially within clinical trials), and improving the discrimination and characterization of disease and syndromic phenotypes. ⁴⁸ Indeed, foundational work has begun to address these needs. For example, studies have demonstrated the feasibility and validity of web‐based eye tracking for cognitive paradigms like the antisaccade task, replicating laboratory findings remotely and highlighting its promise for scalable assessments. ⁵⁹ Furthermore, research integrating multimodal AI approaches, including speech, acoustic, and facial markers, has shown high accuracy in differentiating MCI subjects within an Argentine cohort, emphasizing the importance of local data calibration and the potential of these tools in Latin American populations. ⁶⁰ Other developments include enhancing existing cognitive assessment tools through digital innovation, such as by integrating hand and eye tracking in computerized versions of the TMT to reveal subtle differences in MCI, ¹⁶ and developing sophisticated computational models to predict and understand human eye movements in complex hybrid visual search and free‐viewing tasks, which is crucial for designing robust and ecologically valid eye‐tracking assessments. ⁶¹ , ⁶²

A key challenge for this integrated approach is the need for interpretable results. Unlike high‐cost or invasive techniques, non‐invasive tools like eye trackers can yield data that, ideally, would not require extensive expert interpretation in clinical settings and pose no risk to patients. The ability of non‐experts to use eye‐tracking metrics will depend on establishing consistent performance via simple or automated classification algorithms that are robust to potential confounding factors such as age and educational level. ³² In this sense, this review included studies that have already begun developing ML models for disease‐status classification based on eye tracking. These models have successfully identified AD patients using exploration metrics from search tasks, VPC heatmaps, and smooth pursuit metrics for YOAD. ¹⁷ , ⁴¹ , ⁴⁸ Other studies have focused on saccadic task metrics and search task exploration metrics for differentiating MCI patients from healthy older controls. ²⁷ , ³² The high‐performance metrics reported across these studies underscore the significant opportunity for designing algorithms capable not only of assisting in the identification of individuals at high risk of developing neurodegenerative conditions like AD but also in tracking disease progression over time.

5. CONCLUSIONS

The recent paradigm shift defining AD by its underlying biology points to the need for early detection, as timely intervention may slow irreversible neuronal damage. Eye tracking has emerged as a particularly promising tool, offering non‐invasive, rapid, and quantitative assessment of oculomotor disturbances. This review highlights that antisaccade tasks consistently differentiate individuals with AD and MCI from healthy older controls and can reflect disease severity, indicating their robust potential as screening tools. However, there is a current gap in this area, as no reviewed studies used the latest biological AD diagnostic criteria, presenting fertile ground for research to build upon current findings – such as relevant saccadic metrics and the diminished exploratory drive seen in non‐saccadic tasks – with more homogeneously defined cohorts. Building on our group's foundational work in modeling eye movements during complex visual tasks, developing web‐based eye‐tracking platforms, and leveraging AI for cognitive assessment in Latin American populations, we envision a clear path forward. The integration of eye tracking with accessible web technologies and ML algorithms offers a powerful approach to detecting subtle cognitive changes, monitoring disease progression, and promoting earlier clinical engagement. Our research group aims to directly address the identified gaps by developing standardized web‐based eye‐tracking paradigms in collaboration with neuroscientists. With an initial focus on Latin American populations, we aspire to create accessible and effective tools that can be adapted and extended to underrepresented regions, ultimately contributing to more equitable and timely AD detection worldwide.

CONFLICT OF INTEREST STATEMENT

The authors declare no competing interests. Any author disclosures are available in the Supporting Information.

Supporting information

Supporting Information

Supporting Information

Supporting Information

Supporting Information

Supporting Information

Supporting Information

Supporting Information

Verón GL, Juantorena GE, Keller G, Crivelli L, Kamienkowski JE. Eye tracking as a diagnostic tool in Alzheimer's disease, mild cognitive impairment, and related dementias: a systematic review. Alzheimer's Dement. 2026;18:e70238. 10.1002/dad2.70238