Home-based eye tracking for early autism screening: a scoping review of approaches, evidence, and implementation challenges Authors: Aya A. Elkhodiry, Syed Ali Bokhari, Nejad Alagha, Dania Mohammed, Rajveen Kakkar, Afra Abdalla, Abigail Copiaco, Yassine Himeur, Christian Ritz, Amin Beheshti, Wathiq Mansoor, Ammar Albanna, Valsamma Eapen Published: 2026 ABSTRACT BackgroundEarly identification of autism spectrum disorder (ASD) is essential for improving developmental outcomes but remains challenging due to diagnostic delays, subjectivity, and resource limitations. Eye-tracking offers objective indices of social attention and could improve access to early screening when deployed on consumer devices at home. This scoping review synthesizes evidence on home-deployable eye-tracking as digital biomarkers for early ASD screening, associated machine learning methods, and feasibility of real-world implementation.MethodsFollowing the Arksey and O’Malley framework and Joanna Briggs Institute (JBI) methodology, reported per PRISMA Extension for Scoping Reviews (PRISMA-ScR), we searched PubMed and Scopus (2015-2025) for English-language human studies using terms for autism, eye-tracking, and home-based assessment. Using a Population, Concept, Context (PCC) framework (Population: infants/ children; Concept: eye-tracking as digital biomarker; Context: home/ clinical settings), two reviewers screened each record and charted data on participant characteristics, stimuli, devices, analytic methods, diagnostic performance, and implementation indicators. Extracted data, including diagnostic performance metrics (PPV, NPV, confidence intervals), study design classifications, validation methods, and biomarker readiness ratings, are provided in full in Supplementary Tables S2–S6. The protocol was preregistered on the Open Science Framework (OSF Registries osf.io/mdz2e/; 10.17605/OSF.IO/MDZ2E). Searches were last executed on 1 August 2025. A process-based estimation approach was used to quantify the EOL carbon footprint of two timber floor systems—Adhesive & Screw and Sharp Plate & Screw, following ISO 14040/44 standards. Four realistic EOL pathways (landfilling, downcycling, component reuse, and full assembly reuse) were assessed under three recovery-rate scenarios (90%, 60%, and 30%) to capture both ideal deconstruction and conventional demolition conditions. EOL impacts, biogenic carbon flows, and Stage D credits (potential environmental benefits beyond building life) were integrated to determine the EOL climate outcomes.ResultsAcross 90 studies, reported discrimination typically ranged from moderate to high. Sample sizes varied from small pilots to large datasets (n = 30–1,000+), with ages from 5 months to 18 years. Studies were categorized by system type: hardware-based, wearable, webcam-based, and home-deployed. By study purpose, studies ranged from prospective screening evaluations and internally validated case-control classification studies to exploratory group-difference analyses and feasibility evaluations; the majority fall into the latter two categories. Common tasks included faces/social scenes, biological motion, and joint-attention cues; frequently reported metrics included fixation proportion, dwell time, saccade dynamics, and interest-area contrasts (eyes/face vs. objects). Supervised learning (e.g. SVM, random forests) and deep learning pipelines were used, though external validation and calibration-robustness were inconsistently reported. Practical barriers included ambient-light variability, viewing distance, caregiver facilitation, device heterogeneity, and data privacy/governance.ConclusionThis review mapped 90 studies across four eye-tracking platform stages to assess translational readiness for home-based ASD screening. Gaze-based biomarkers carry robust discriminative signal, but the evidence derives almost entirely from laboratory-based, case-control studies. Only three studies collected data in home-deployed settings, none with externally validated accuracy. Priorities include prospective multisite home-based trials with prespecified endpoints, standardised stimulus protocols, evaluation of model robustness to distribution shift, and navigation of jurisdiction-specific regulatory pathways. Limitations include English-only sourcing, two databases, the 2015–2025 window, and lack of critical appraisal.Clinical trial numberNot Applicable.Supplementary informationThe online version contains supplementary material available at 10.1186/s12888-026-08255-y. Background Autism spectrum disorder: current challenges in early detection Autism Spectrum Disorder (ASD) is defined as a developmental disorder characterised by problems with social interaction, communication, restrictive interests and, repetitive behaviours as per the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, Text Revision [1]. According to the latest report by the Center for Disease Control and Prevention’s (CDC), the prevalence of autism in the United States (US) has been increasing steadily with the current rate of 1 in 31 [2]. Despite efforts to promote earlier detection, the average age of ASD diagnosis in the U.S. is 4 years, limiting the benefits of early interventions [2]. The process of autism diagnosis focuses on recognising deficits in social communication and interaction, along with restrictive and repetitive patterns of behaviour relying on clinical observations and caregiver reports. These methods are liable to subjective perceptions as their accuracy would depend on the clinician’s expertise, the parent/caregiver response and any inter-rater variability in observation measures resulting in variable performance accuracies [3–5]. Extensive training is usually required for both clinicians and parents with reliability checks to improve accuracy and limit biased assessments [6]. This creates a “diagnostic bottleneck” where the demand for specialized evaluations far outstrips availability, leading to inequity in access to diagnosis and long waitlists. Moreover, these ratings commonly omit data reflecting children’s behaviour in their natural environments, such as home or school. This poses challenges in accuracy and reliability for initiatives like broad population screening, implementation in low-income communities, or continuous longitudinal monitoring, all of which are indispensable for outcome evaluation in multi-site research and for symptom assessment within the general populace [7]. Even with gold-standard tools, a substantial proportion of children eventually diagnosed with ASD are not identified at the earliest ages [8]. This difficulty is compounded by the inherent challenges in reliably measuring behavioural symptoms, particularly in the 12 to 24-month age range, where differential social attention patterns are often manifested [9, 10]. Therefore, there is a critical need for biological markers (biomarkers) that enable objective detection of ASD before overt behavioural symptoms appear [11]. Ultimately, ensuring that all children receive timely, equitable identification and intervention is an ethical imperative for optimizing developmental outcomes. Early intervention is a policy priority in countries like Australia [12] and the United Arab Emirates (UAE). A recent review highlighted delays in timely ASD diagnosis and the need for culturally adapted, accessible screening solutions [13]. The Dubai Clinical Practice Guidelines for ASD emphasize early detection and standardised care pathways, laying a foundation to integrate innovative home-based eye-tracking into existing services [14]. Parents in the UAE report variable service quality and unmet needs in access and coordination, reinforcing the urgency of home-based digital screening pathways [15]. Prior studies in the UAE have explored genetic and environmental factors in ASD but their findings are limited by the subjectivity of ASD diagnoses [16, 17]. Defining biologically based phenotypes could improve diagnostic accuracy and reveal underlying mechanisms [18]. For example, a chromosomal microarray analysis of an Emirati neurodevelopmental cohort identified ASD-relevant copy number variants, demonstrating the feasibility of objective, locally validated genomic approaches. These findings underscore the need to complement policy efforts with biomarker-driven digital screening pathways, such as home-based eye-tracking, that are scalable and culturally appropriate [19]. Digital biomarkers in healthcare Digital biomarkers are measures that are objectively gathered via digital tools to capture physiological and behavioural data [20, 21]. The advancing pattern of digital biomarkers represents a significant evolution in healthcare assessment, providing new approaches to monitor diagnostic, prognostic, and therapeutic aspects [21]. Advancements in sensor technology, data analytics and computing technology enables the collection of continuous, detailed, high-quality data from individuals in their natural environments [22]. The use of digital biomarkers revolutionizes clinical as well as research practices in the field of neurodevelopmental disorders by capturing subtle changes in behaviour, cognition and physiology [23, 24]. Actigraphy, speech analysis, and eye-tracking are examples of digital markers that have demonstrated potential in identifying early signs and tracking therapeutic outcomes with greater sensitivity and reliability than conventional assessments [23–25]. Applying artificial intelligence (AI) and machine learning (ML) to the digital biomarkers data, enriches biomarker discovery, enhances predictive accuracy and enables effective and scalable solutions in the field [23]. Challenges involving data quality, validation and regulatory acceptance arise with such potential, however, technological advances are continuously tackling such challenges. This positions digital biomarkers as a cornerstone for the future of precision medicine in neurodevelopmental health [24, 26]. Given the limitations of conventional ASD screening and the promising potential of the aforementioned digital biomarkers, eye-tracking technology emerges as an intriguing objective tool for capturing early indicators of atypical social and joint attention [27]. With advances that bring eye-tracking from the laboratory into home settings [28], it shows strong potential to support scalable, accessible, and equitable early autism screening. This scoping review aims to explore the utility of home-based eye-tracking as a digital biomarker for early ASD screening. Specifically, it will examine the established application of eye-tracking in the clinical settings, traces the evolution toward home-based systems, assesses diagnostic performance, discusses implementation challenges, and highlights future directions. Eye-tracking technology overview Eye-tracking technology offers the capacity to measure visual attention objectively by following eye movements and fixation patterns as participants engage with visual stimuli [29, 30]. Typically, eye-tracking systems monitor eye movements using cameras or infrared sensors, usually employing corneal reflection or pupil-corneal reflection methods [31, 32]. These systems provide evidence of attentional processes, cognitive load, and data processing by measuring where, how long, and in what sequence an individual watches different elements within a visual scene [30, 33]. Traditionally, eye-tracking was confined to laboratory settings, relying on expensive, heavy equipment requiring regulated environments and participant fixture, usually employing chinrests to stabilize the head [31, 34]. Over the past two decades, major technical advancements have driven miniaturization and cost reduction of eye-tracking equipment, therefore enabling the move from research laboratory to consumer level hardware suited for usage in ecologically valid environments such as webcam attached to a portable computer or a tablet [32]. Emerging technologies, ongoing innovations in hardware, algorithms, and interfaces are promoting the potential for eye-tracking to serve as an objective metric for early neurodevelopmental screening and monitoring [31, 32]. Rationale for eye-tracking in ASD screening Numerous studies have demonstrated reliable variations in the visual focus of individuals with ASD. These measurable differences in the gaze of infants and young children closely align with the fundamental symptoms of ASD, including social communication and interaction difficulties [35, 36]. These discrepancies point to underlying neurological processing variations that influence how children with ASD perceive and comprehend important social cues in their surroundings [37]. Given that these patterns of visual attention are essential features of ASD, early, objective, and quantifiable detection of the condition becomes feasible. Beyond subjective clinical observations, eye-tracking technology provides a non-invasive and precise method to record these early variations and facilitates the development of reliable digital screening tools [38]. Reduced awareness of social information is a prominent symptom of ASD. Compared to typically developing children, individuals with ASD exhibit a diminished focus on faces, particularly the eyes, and an increased attention to objects or non-social elements within their surroundings, even when social stimuli are present [39, 40]. However, these effects are heterogeneous; some autistic toddlers do not exhibit reduced social orienting, and the magnitude of gaze differences vary by stimulus design (e.g., dynamic social scenes vs. geometric motion), task demands (free viewing, orienting, joint-attention probes), and the gaze features analysed. Collectively, these studies indicate that eye-tracking can be suited as a screening and risk-stratification aid, rather than a stand-alone diagnostic test; its outputs should always be interpreted in conjunction with clinical evaluation and developmental history. By utilising metrics such as total gaze time, the number of fixations, and eye movement patterns, researchers can effectively distinguish children with ASD from their typically developing peers as early as one year to three years of age [41, 42]. Difficulty with joint attention, the shared attention on an object or occurrence with another person, is another crucial early indication of autism [43, 44]. Eye-tracking tests that monitor how children follow another person’s gaze or pointed motions show that children with autism are less likely to look where others are looking, thereby highlighting challenges with joint attention [45, 46]. Beyond simple assessments, more complex behaviours like how varied a child’s eye movements are, how rapidly they look at social cues, the order in which they look at different things, and how they perceive emotions and speech can also indicate autism risk [46–48]. Utilising machine learning to examine these complex gaze patterns has demonstrated remarkable accuracy in diagnosing ASD, comparable to more expensive brain imaging techniques. However, prudent interpretation is essential, as the reported discriminations are often context- and sample-dependent [41, 49]. Together, these eye-tracking indicators capture the fundamental social attention and interaction challenges prevalent in individuals with ASD and transform them into unambiguous, objective data. This methodology not only facilitates the early identification of ASD but also provides a valuable, non-invasive method to monitor development in at-risk children. Some studies have used eye-tracking to assess intervention outcomes, link its measures with disease severity and progression, and further investigate autism-related behaviours and underlying mechanisms through both observational and interventional approaches [50–54]. Combining multiple eye-tracking measures into a single screening tool improves accuracy, thereby capturing the wide range of ASD symptoms exhibited in different children. This integrated approach enhances the reliability and utility of screening across diverse populations. This review is positioned as a translational technology survey, tracing the trajectory of eye-tracking research for ASD screening from dedicated laboratory hardware through wearable and webcam-based platforms to the emerging generation of home-deployed studies. This breadth is intentional as laboratory research established the gaze biomarkers and validation benchmarks that any home-based system must replicate under less controlled conditions, making it a necessary foundation for evaluating translational readiness. The scope is best understood as a translational assessment of the full eye-tracking evidence base evaluated against the requirements for home-based screening. Machine learning in eye-tracking for ASD Machine learning has emerged as a powerful tool for developing predictive models for early autism screening utilising eye-tracking data. Researchers are utilising eye-tracking datasets of scan paths or heatmaps such as “Saliency4ASD” or newly collected datasets of gaze patterns, to extract features and train ML models [55]. ML algorithms learn to identify patterns in eye-tracking data that discriminate between ASD and typically developing individuals. Examples of commonly used models are Support Vector Machines (SVM), Random Forest (RF) and more recently, deep learning models such as Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM). The use of these models with eye-tracking data has shown high accuracy in identifying ASD-related gaze features, though reported performance varies substantially across datasets, model architectures, and validation strategies (reviewed in [56–59]). This workflow can be conceptualized as a pipeline where raw gaze data are transformed into quantitative metrics, pre-processed for feature extraction, and subsequently used to train machine learning models for classification or risk stratification. Figure 1 illustrates this end-to-end process, linking experimental design with clinical implementation.Fig. 1Workflow of eye-tracking and machine learning integration for autism screening. This schematic demonstrates how raw gaze data are acquired and transformed into clinically meaningful outputs. Infrared micro projectors and image sensors capture gaze points in real time, while image processing and mathematical models determine eye position and fixation. Experimental paradigms generate raw measures that are converted into derived metrics. Data undergo cleaning, preprocessing, and feature extraction before entering machine learning models for classification, regression, or clustering. The framework enables enhanced diagnosis, and personalized interventions, with clinical insights feeding back into experimental design Workflow of eye-tracking and machine learning integration for autism screening. This schematic demonstrates how raw gaze data are acquired and transformed into clinically meaningful outputs. Infrared micro projectors and image sensors capture gaze points in real time, while image processing and mathematical models determine eye position and fixation. Experimental paradigms generate raw measures that are converted into derived metrics. Data undergo cleaning, preprocessing, and feature extraction before entering machine learning models for classification, regression, or clustering. The framework enables enhanced diagnosis, and personalized interventions, with clinical insights feeding back into experimental design Methods Protocol and Registration The protocol was preregistered and publicly available on the OSF Registries osf.io/mdz2e/; 10.17605/OSF.IO/MDZ2E and follows the Joanna Briggs Institute (JBI) Manual for Evidence Synthesis and the framework of Arksey and O’Malley, refined by Levac and colleagues [60, 61]. Reporting follows the PRISMA guidelines for scoping reviews (PRISMA-ScR) [62]. Rationale and objectives Given the heterogeneity in populations, devices, tasks, and outcomes, a scoping review design was selected to map the breadth of evidence and identify gaps rather than evaluate pooled effects. Literature search strategy A comprehensive literature search of PubMed and Scopus was conducted for studies published between 1 January 2015 and 1 August 2025, restricted to English-language records. PubMed queries were limited to the Title/Abstract fields, and Scopus searches to the TITLE-ABS-KEY fields. The strategy integrated four concept domains: (i) autism spectrum disorder (autism spectrum disorder, autism, autistic, ASD); (ii) eye-tracking (eye-tracking, eye movement, gaze tracking, visual attention); (iii) context/technology (home, remote, telehealth, mobile, digital) and digital biomarkers (digital/computational biomarker, objective measure); and (iv) screening/diagnosis and paediatric age (screening, detection, diagnosis; early diagnosis/early screening; infant, toddler, child). Concepts were combined using AND across domains and OR within domains. Searches were executed in each database’s advanced interface, with results exported and de-duplicated before screening. Searches were last executed on 1 August 2025 in both PubMed and Scopus. Study selection and inclusion criteria Screening followed the PRISMA 2020 guidelines for scoping reviews (PRISMA-ScR) [62]. Studies were screened for titles and abstracts, followed by full-text review and retrieval against predefined inclusion criteria. Eligible studies were required to be English-language, full-text, empirical investigations that (i) collected experimental eye-tracking data with sufficient methodological detail reported, (ii) addressed ASD screening, early detection, or diagnostic differentiation, (iii) enrolled infants, toddlers, or children in case of hardware-based studies, and (iv) reported new participant-level data (computational-only studies based solely on publicly available datasets were excluded). Titles and abstracts were screened independently by two reviewers after a calibration exercise on a random sample of 30 records. The same procedure was applied to full-text screening. Disagreements were resolved by discussion or, when needed, third-reviewer adjudication. Reasons for full-text exclusion were recorded. The selection process is shown in the PRISMA-ScR flow diagram (Fig. 2).Fig. 2PRISMA-ScR flow diagram illustrating the identification, screening, and inclusion process for studies in this review. A total of 680 records were identified through database searches. After removing duplicates and screening for relevance to the use of eye-tracking as a diagnostic marker for autism, 90 articles were included in the final synthesis PRISMA-ScR flow diagram illustrating the identification, screening, and inclusion process for studies in this review. A total of 680 records were identified through database searches. After removing duplicates and screening for relevance to the use of eye-tracking as a diagnostic marker for autism, 90 articles were included in the final synthesis Data charting, items, and management A standardised data-charting instrument (Microsoft Excel) was collaboratively developed and piloted on five representative studies prior to full extraction. Five reviewers independently charted each included source. Extracted items comprised: (a) publication metadata; (b) methodological and sample characteristics (design, age, diagnostic subgroup, sample sizes, comparators/validation methods); (c) experimental conditions (platform type, eye-tracker specifications, stimulus type, setting); (d) diagnostic outcomes (key gaze metrics; accuracy, sensitivity, specificity, AUC; confidence intervals; performance notes); (e) feasibility and implementation indicators (calibration success, data yield, participant tolerance, caregiver or clinician burden); and (f) analytical approach (statistical methods and machine-learning models). Discrepancies were reconciled by consensus, and decision rules were version-controlled in a shared repository. Consistent with scoping-review methodology, no critical appraisal of included studies was undertaken, as our aim was to map the evidence rather than evaluate study quality. Synthesis of results No quantitative meta-analysis was planned. Data were synthesised descriptively using frequency summaries and evidence mapping across four organising dimensions: modality (hardware, wearable, webcam, home-deployed), population (infant/toddler/child; diagnostic subgroup), analytical approach (traditional vs. machine learning), and feasibility/implementation indicators. In line with scoping-review methodology, no formal risk-of-bias appraisal was undertaken. Ethics approval Ethics approval was not required because the review analysed publicly available, published data only. Terminology conventions For consistency, the following conventions apply throughout: “home-based”; “eye-tracking”; “screening” refers to population-level risk identification in unselected community samples using a defined threshold; “classification” refers to algorithmic categorisation in balanced pre-selected case-control designs, which systematically overestimates real-world screening performance; “diagnosis” refers to clinical determination by a qualified professional per DSM-5 or ADOS-2; and “feasibility evaluation” refers to assessment of technical and practical acceptability without classification outcomes. Evidence tiers in Supplementary Tables S2–S5 are graded as: validated biomarker (externally validated, prospective); candidate biomarker (internally cross-validated); exploratory marker (group-level differences only); or non-diagnostic (feasibility study). “Home-based potential” is operationalised as requiring consumer-grade hardware, no trained clinician, an unsupervised setting, minimal calibration, and functioning under variable lighting; this reflects design criteria only and does not indicate data were collected at home unless explicitly stated. Results Study selection The search yielded 680 records (PubMed n = 414; Scopus n = 266). After removal of 191 duplicates, 489 unique records were screened for relevance, eligibility and full text retrieval. A total of 399 studies were excluded for predefined reasons, resulting in 90 studies included in the final synthesis. The selection process is shown in the PRISMA-ScR flow diagram (Fig. 2). Study characteristics Included sources encompassed hardware-based, wearable, webcam-based, and explicitly home-deployed modalities, with paediatric samples spanning infancy through middle childhood. Task paradigms and gaze metrics were heterogeneous across studies. Studies were examined across various eye-tracking modalities. Through the years, eye-tracking methodology has evolved from stationary eye-tracking systems employed at the clinic or laboratory (approximately 82.4% of papers), to wearable devices accounting for 8.8% and, more recently, webcam-based systems accounting for 5.5% of the papers. Below, we summarize recent research for each type and the emergence of home-based eye-tracking (addressed in 3.3% of papers). These technological advances trace a clear trajectory from laboratory-based systems to wearable and webcam-based platforms. Figure 3 summarizes this evolution, highlighting the trend toward accessibility, portability, and non-invasive use. To facilitate comparison across studies, Supplementary Tables S2–S5 provide expanded data for all 90 included studies, including study design classification, validation methodology, gold-standard comparator, biomarker readiness classification, and full diagnostic performance metrics (sensitivity, specificity, PPV, NPV, AUC, and confidence intervals where reported). The heterogeneity of validation approaches and comparator standards cautions against direct metric comparisons; this limitation is addressed explicitly in the Discussion.Fig. 3Evolution of eye-tracking technologies. Four stages characterize eye-tracking development: dedicated laboratory hardware, screen-mounted remote systems, wearable trackers for real-world use, and webcam-based tracking using consumer devices with computer vision algorithms. This progression toward greater accessibility and portability enables decentralised autism screening on consumer-grade devices Evolution of eye-tracking technologies. Four stages characterize eye-tracking development: dedicated laboratory hardware, screen-mounted remote systems, wearable trackers for real-world use, and webcam-based tracking using consumer devices with computer vision algorithms. This progression toward greater accessibility and portability enables decentralised autism screening on consumer-grade devices Hardware-based eye trackers Within the time frame of the review, 74 out of 90 of the papers were based on hardware-based eye-tracking with the use of stationary hardware at the clinic or laboratory setting. Table 1 presents a condensed summary of hardware-based studies; the complete expanded dataset, including validation methods, comparator tools, PPV/NPV, and confidence intervals, is available in Supplementary Table S2.Table 1Hardware-based eye-tracking studies for autism spectrum disorder (ASD). Summary of included studies reporting sample characteristics, stimuli, gaze metrics, machine learning methods, and diagnostic performanceFirst Author (Year)Age Group (Mean)Sample Size (Total; ASD/Control)Home Based PotentialStimulus TypeKey MetricsSensitivity (%)/Specificity (%)AUC ScoreML AlgorithmRefAhmed (2023)7.88 y59 (29/30)NoDynamic/static stimuli (balloons, cartoons, facial images, objects)Visual patterns, eye contact, engagement, gaze focus97.25/98.940.98LSTM, CNN-LSTM, GRU, BiLSTM[63]Alarifi (2023)TD: 88.21 ± 4.12 y; ASD: 8 ± 3.89 y104 (63/41)NoVisual stimuli (human & animal cartoon faces)Dwell time, time to first fixationN/AN/ASVM, RF, LSTM, Logistic Regression[64]Al-Shaban (2023)7.7 y228 (144/84)NoSocial scenesDwell proportionN/A0.73N/A[65]Amestoy A (2021)8.8 ± 1.4, 14.1 ± 1.3, 29 ± 6.9 y129 (97/49)NoSaccade tasks (Gap, Step, Overlap, Antisaccade)Attention span, response time/gain, saccade accuracyN/AN/AN/A[66]Antolí A (2025)53.51 m93 (ASD:24; DLD:25/44)NoSocial & non-social image pairsVisual preferenceN/AN/ANaïve Bayes, Ripper, One Rule, Partial Decision Trees, C4.5, Logistic Model Trees[67]Bast (2021)ASD: 16.23 y, TD: 16.21 y284 (142/142)NoNaturalistic video stimuliFixation duration, fixations/sec, pupil dilation, saccade duration/amplitude/velocityN/A0.67N/A[68]Caldani (2020)12.1 y64 (32/32)NoSaccade paradigms (reflexive & voluntary)Saccade express/anticipatory/gain/error ratesN/AN/AN/A[69]Camero (2021)ASD: 21 m; TD: 20 m20 (10/10)NoFaces, pseudo-objects, pseudo-wordsFixation & gaze duration, pupil dilationN/AN/AN/A[70]Camero (2023)ASD: 27.9 m, TD: 14.4 m114 (57/57)NoCartoons, pseudo-words, pseudo-objectsGaze following & attention timeN/AN/AN/A[71]Chevallier C (2015)ASD: 12.2 y, TD: 14.9 y81 (59/22)NoStatic, dynamic & interactive visual explorationSocial prioritization scoreN/AInteractive: 0.721; Dynamic: 0.623; Static: 0.448N/A[72]Chukoskie (2018)13.9 y8 (8/0)YesVideo gamesFixation duration, saccade patterns, spatial attentionN/AN/AN/A[73]Cilia (2021)ASD: 7.58 y; non-ASD: 8 y59 (29/30)NoSocial vs. non-social videos/photos (joint attention)Visual patterns via scan paths83/800.9CNN[74]Congiu (2024)ASD: 56.63 m; TD: 60.42 m38 (19/19)NoHuman faces and objectsPrioritization, preference, sustained attentionN/AN/AN/A[75]de Belen (2023)ASD: 4.57 y; TD: 4.61 y77 (60/17)NoJoint attention task with actor & toysGaze shift scores, response timeN/AN/AN/A[27]de Belen (2024)ASD: 4.63 y; TD: 4.61 y74 (57/17)NoGeoPref Test dynamic stimulusVisual attention100/76.470.91CNN-LSTM[25]Fernandez-Lanvin (2024)14 m122NoSocial stimuli (video scenes)Social engagement, gaze exchanges, gaze following94.3/94.70.9845 (RF)RF, Naive Bayes, KNN, AdaBoost, ANN, SVM, J48, PART[76]Fletcher-Watson (2018)ASD: 23.06 m, TD: 23.25 m39 (19/20)YesApp screenshots, cartoons, photographsFixation duration, AOI, TFFN/AN/AN/A[77]Fong (2019)ASD: 68 m, TD: 61 m40 (20/20)NoColoured overlaysOcular performance (fixation duration/count, visit duration), reading speed/errorsN/AN/AN/A[78]Franco (2023)TD: 6.7 y, ASD: 6.8 y77 (33/44)NoSocial scenes, geometric scenesFixation points69/930.82RF, SVM, ANN[79]Frazier (2016)ASD: 4.9 y, TD: 5.8 y79 (40/39)YesStill pictures, dynamic videosFixation duration, fixation count, gaze distribution80/820.86N/A[6]Frazier (2018)ASD: 5.7 y, TD: 6.8 y201 (91/110)YesFacial affect, gaze following, joint attention, social images, naturalistic scenesGaze patterns, saccades, blinksN/A0.86N/A[80]Griffin (2020)TD: 13.9 y, ASD: 13.5 y70 (35/35)NoGaze following/perception tasksVisual attention, gaze shifts-N/AN/A[81]Harrop (2019)ASD (M): 9.53, ASD (F): 8.33, TD (M): 7.68, TD (F): 7.93 y74 (42/32)NoDyadic & parallel play scenesTotal fixation durationN/AN/AN/A[82]He (2021)HFA: 5.08 y, LFA: 4.98 y, TD: 5.24 y74 (26 HFA/24 LFA/24 TD)YesGaze stimuli (human, cartoon, finger, arrow)Reaction time to AOI, fixation length79.2/900.938KNN[83]Hudry (2025)TD: 44 m, ASD: 40 m196 (96/100)NoSocial/non-social scenes, faces, referential attentionPeople vs. geometry, eyes vs. mouth, PLD, RJA82/700.82Classification Algorithm[84]Ioannou (2020)N/A73 (12/25)NoStatic social images (low vs. high complexity)Fixation duration on faces, TFF90.0/78.0N/AN/A[85]Jones (2023)Discovery: 22.4 m, Replication: 25.4 m1089 (519/570)NoSocial interaction videosSocial visual engagementN/A0.89, 0.90N/A[86]Jones (2023)TD: 23.4 m, ASD: 24.9 m335 (150/185)NoSocial interaction videos (Eye-tracking index test)Social visual engagement78/85.40.9N/A[87]Kang (2020)ASD: 4.52 y, TD: 4.64 y157 (77/80)NoFace photos (own-race, other-race, familiar)Corneal reflections, gaze direction, fixation duration, saccade featuresN/A0.89K-means, linear SVM[88]Kang (2020)ASD: 4.29 y, TD: 4.26 y97 (49/48)NoFace photos (own-race young girl), dynamic kitten videoFixation time in AOIN/A0.93SVM[89]Keehn (2024)ASD: 2.6 m; TD: 2.4 m146 (102/44)NoGeoPref test, gap-overlap paradigm, PLR, animated cartoonNon-social preference, no-shift %, PLR latency/amplitude, fixation durations77.5/77.30.93NA[38]Kong (2022)Toddler ASD: 2.4 y; Toddler Control: 2.2 y; Preschool ASD: 4.6 y; Preschool control: 4.8 y173 (92/81)NoDynamic stimuli (rabbit, faces, biological motion, objects, infant face)Gaze patterns, fixation timeToddlers: 80/82; Pre-schoolers: 81/61N/ASVM[41]Kou (2019)ASD: 3.72 y; TD: 4.10 y66 (32/34)NoDynamic videos (dancing adult) & geometric shapesSocial attention via visual preferenceN/A0.764 (fixation counts for DSI)N/A[90]Krol (2019)ASD: 16.27 m; TD: 16.31 m44 (21/23)NoFacial features, emotional recognitionFace scanning patternsN/AN/AEmotion classification model[91]Król (2019)ASD: 16.27 m, TD: 16.31 m44 (21/23)YesFace images (FACES database, various expressions)Eye fixations, Scan Match similarityN/AN/ALDA[92]Lencastre (2024)7.74 y57 (28/29)NoNon-organic distractor (balloon), handwaving cartoonGaze analysis95.1/93.1N/ATime neural network[93]Lian (2023)ASD: 11.17 y, TD: 8.96 y49 (23/26)YesApp-based UI tasks (icons, buttons, labels)Number of fixations, TFF, total fixation durationN/AN/AN/A[94]Liu (2016)ASD: 7.90 y, TD: 7.86 y87 (29/29; 29)YesFacial images for recognitionFixation duration93.1/86.20.8963SVM (RBF kernel)[95]Lockwood (2024)ASD: 4.50 y; TD: 4.20 y78 (46/32) + 26 IDNoProsaccade/antisaccade stimuli (circles, animal animations)Prosaccades, antisaccadesN/ATask1: 0.82, Task 2: 0.74, Task 3: 0.74N/A[96]Lohan (2018)ASD: 128.63 m, TD: 126.38 m53 (24/29)YesFaces, houses, abstract patterns, avatar interactionPupil diameter, gaze location, AOIN/AN/AN/A[97]Meng (2023)39.70 m161 (117/44)NoVisual stimulusAttention patterns, engagement75/N/A0.81CNN, RF[98]Moore (2018)ASD: 30.0 m, TD: 29.7 m277 (76/51)NoDynamic social vs. geometric patternsFixation duration, saccade patterns35.00/94.000.75N/A[99]Muratori (2019)ASD: 25.1 m; TD: 31.7 m27 (12/15)NoJoint attention video clipsJoint AttentionN/AN/AN/A[100]Oliveira (2021)3–18 y106 (76/30)NoVideos of biological vs. geometric movementsFixation points69/930.822 (ANN with GA); 0.775 (SVM)SVM, ANN/VAMs[101]Ozdemir (2022)ASD: 34.85 m; TD: 32.90 m133 (61/72)NoSocial interaction vs. non-social toy/animation videosDifferential visual attention patternN/A0.708DT, NB, RF, SVM[102]Pierce (2016)ASD: 28 m, TD: 23.6 m444 (152/115)NoSide-by-side DGI & DSI imagesFixation duration/patterns, saccade numbers, AOI21/99N/AN/A[103]Robain (2022)ASD: 3.19 y, TD: 2.95 y110 (90/20)YesCartoon & real-life videosLooking time, social orienting, fixation duration/count, TFFN/AN/AN/A[104]Rudling (2024)10 m/14 m/18 m169 (EL-ASD: 35/EL-non-ASD: 94; TD: 40)NoDirect gaze interactive taskPercentage fixation, TFF latencyN/AN/AN/A[105]Samad (2018)ASD: 13 y, TD: 16 y16 (8/8)Yes3D facial expressions, facial recognitionGaze fixation duration, eye-hand coordination correlationN/AN/ASVM (RBF kernel)[106]Scherf (2018)10–18 y84 (34/50)NoVideo clips of actors expressing emotionsJoint attention, gaze trajectories, fixation durationN/AN/AN/A[107]Shic F (2022)ASD: 8.55 y; TD: 8.51 y399 (280/119)NoActivity, social, static, biomotion, PLR tasksOculomotor Index of Gaze to Human Faces (OMI)/Biomotion/PLRN/AN/AN/A[108]Sosnowski (2022)8.56 y54 (54/0)YesFacial emotion recognition tasksEmotion recognition scores pre/post-interventionN/AN/AN/A[109]Sun (2023)ASD: 3.1 y, TD: 2.9 y59 (32/27)NoNon-social objects; electronic & transport imagesGaze pattern & analysisADOS: 91/78.7, RRB: 90/77.40.812N/A[110]Sun (2024)ASD: 3.1 y; TD: 2.9 y59 (32/27)NoVisual stimulusGaze patterns66.7/70.60.75Multimodal Logistic Regression[111]Tenenbaum (2020)ASD: 65.65 m; TD: 49.95 m176 (93/40)YesSocial interaction (dyadic, joint attention, sandwich-making, moving toy)Frame-by-frame fixations, scan path distanceN/AN/AN/A[112]Thompson (2018)8.52 y16 (16/0)YesSinging (song) and speech (story)Fixation count, dwell timeN/AN/AN/A[113]Thorup (2016)HR: 10.25 m, LR: 10.27 m64 (47/17)YesLive interaction (experimenter, puppets)Gaze shifts, gaze fixationN/AN/AN/A[114]Tsuchiya (2021)ASD: 10.3 y; TD: 9.5 y141 (39/102)YesSocial vs non-social scenes, preferential paradigmsGaze fixation %, AOI74/800.84Best fit diagnostic algorithm[115]Vacas (2021)ASD: 55.89 m, TD: 53.53 m38 (19/19)NoStatic paired images (faces vs. CIO/non-CIO objects)Prioritization, preference, duration (sustained attention)N/AN/AN/A[116]Wadhera (2021)ASD: 16.32 m; TD: 25.43 m60 (30/30)YesMaze-based cognitive learning taskFixation duration, saccade amplitude, EEG metricsN/AN/AN/A[117]Wan (2019)ASD: 4.6 y, TD: 4.8 y74 (37/37)YesDynamic rabbit, video of a female mouthing alphabetGaze patterns, fixation time86.5/83.8N/AN/A[118]Wang (2018)ASD: 3.73 y, TD: 3.98 y39 (19/20)YesShort cartoon videosScreen dwelling, gaze positions, fixationsN/AN/AN/A[119]Wang (2024)ASD: 3.80 y; non-ASD: 3.30 y39 (22/17)NoDynamic videos (eyes vs mouth)AOI Switch Counts (ASC), Favoured AOI Shifts (FAS), AOI Vacancy Counts (AVC), Total Gaze Count (TGC), Total Fixation Time (TFT)88–100/80–88N/AN/A[42]Wang (2025)ASD: 4.3 y, TD: 4.46 y45 (22/23)YesSocial stimuliAttention & response100/1001ResNet18 + Bi-LSTM[120]Wei (2024)Training: 4.1 y; Validation: 3.1 y529 (290/239)YesSocial stimuliGaze movements & pattern82.2/69.40.849Random Forest[121]Wen (2022)24.37 m1863 (725/1138)NoDynamic social vs. geometric side-by-side videoFixation to geometric AOIs, ASDGeo subtype17.00/98.00 (69 fixations); 23.00/95.00 (CV); 33.00/95.20 (fixations + saccades)N/AN/A[122]Xiong (2024)18–24 m76 (41/35)NoVisual stimulusGaze movements & pattern87.8/45.70.684Binary Logistic Regression[123]Zacharov (2022)ASD: 53.54 m, TD: 49.50 m37 (13/24)NoAnimated cartoon train (A-not-B WM task)Looking duration at AOIs, first-gaze accuracyN/AN/AN/A[124]Zhang (2019)ASD: 59.40 m, TD: 62.30 m34 (15/19)NoTarget/distractor stimuli (squares, circles)Saccade amplitude, peak velocity, durationN/AN/AN/A[125]Zhang (2021)ASD: 71.67 m, TD: 70.21 m34 (15/19)NoEmotional face distractors (angry, happy, neutral)Saccadic errors, saccade latencies (RDP)N/AN/AN/A[126]Zhang (2021)3–6 y42 (21/21)NoET tasks (e.g., “NativeFace1/2”), static & dynamicET metrics (fixation duration, gaze entropy) + EEG featuresN/AN/ARF (ET feature selection) + GCN (fusion classifier)[127]Zhang (2022)ASD: 5.37 ± 0.74 y; TD: 5.54 ± 0.75 y; TD adults: 25.44 ± 2.62 y299 (143/113 children + 43 adults)NoDynamic video clips (actors pointing/looking)RJA events, valid sampling rateN/A (not reported)0.82Logistic Regression[128]Zhang (2025)ASD: 3.89 y; TD: 4.56 y59 (33/26)NoInteractive, visualEye movements97/88.50.978ResNet-50+ LSTM[129]Zhou (2018)ASD: 5.7 y; TD (5y): 5.6 y; TD (4y): 4.6 y75 (26/49)NoAuditory sentences + visual imagesAnticipatory/competitor looks, fixation latency, gaze proportionN/AN/AN/A[130]Note. Age units: m, months; y, years. Participant groups: ASD, autism spectrum disorder; TD, typically developing; HR, high risk; LR, low risk; M/F, male/female; Disc, discovery cohort; Repl, replication cohort; HFA, high-functioning autism; LFA, low-functioning autism; DLD, developmental language disorder; ID, intellectual disability. Stimulus: App UI, application user-interface tasks. Key metrics: AOI, area of interest; TFF, time to first fixation; EEG, electroencephalography RJA, responsive joint attention; PLR, pupillary light reflex; WM, working memory; DGI, dynamic geometric images; DSI, dynamic social images; Dwell time, total time the eyes fixate on a defined AOI. Performance/analysis: AUC, area under the receiver operating characteristic curve; ROC, receiver operating characteristic; N/A, not available or not assessed. ML/methods: SVM, support vector machine; RBF, radial basis function; RF, random forest; NB, naïve Bayes; ANN, artificial neural network; CNN, convolutional neural network; LSTM, long short-term memory; GRU, gated recurrent unit; BiLSTM, bidirectional LSTM; CNN-LSTM, convolutional neural network with LSTM; KNN, k-nearest neighbours; LDA, linear discriminant analysis; J48, C4.5 decision tree (Weka); PART, partial decision tree; VAMs, visual attention models; GA, genetic algorithm; UASN, Uncertainty-inspired ASD Screening Network. Clinical tools: ADOS, Autism Diagnostic Observation Schedule; CARS, Childhood Autism Rating Scale; RRB, restricted, repetitive behaviours Table 1Hardware-based eye-tracking studies for autism spectrum disorder (ASD). Summary of included studies reporting sample characteristics, stimuli, gaze metrics, machine learning methods, and diagnostic performanceFirst Author (Year)Age Group (Mean)Sample Size (Total; ASD/Control)Home Based PotentialStimulus TypeKey MetricsSensitivity (%)/Specificity (%)AUC ScoreML AlgorithmRefAhmed (2023)7.88 y59 (29/30)NoDynamic/static stimuli (balloons, cartoons, facial images, objects)Visual patterns, eye contact, engagement, gaze focus97.25/98.940.98LSTM, CNN-LSTM, GRU, BiLSTM[63]Alarifi (2023)TD: 88.21 ± 4.12 y; ASD: 8 ± 3.89 y104 (63/41)NoVisual stimuli (human & animal cartoon faces)Dwell time, time to first fixationN/AN/ASVM, RF, LSTM, Logistic Regression[64]Al-Shaban (2023)7.7 y228 (144/84)NoSocial scenesDwell proportionN/A0.73N/A[65]Amestoy A (2021)8.8 ± 1.4, 14.1 ± 1.3, 29 ± 6.9 y129 (97/49)NoSaccade tasks (Gap, Step, Overlap, Antisaccade)Attention span, response time/gain, saccade accuracyN/AN/AN/A[66]Antolí A (2025)53.51 m93 (ASD:24; DLD:25/44)NoSocial & non-social image pairsVisual preferenceN/AN/ANaïve Bayes, Ripper, One Rule, Partial Decision Trees, C4.5, Logistic Model Trees[67]Bast (2021)ASD: 16.23 y, TD: 16.21 y284 (142/142)NoNaturalistic video stimuliFixation duration, fixations/sec, pupil dilation, saccade duration/amplitude/velocityN/A0.67N/A[68]Caldani (2020)12.1 y64 (32/32)NoSaccade paradigms (reflexive & voluntary)Saccade express/anticipatory/gain/error ratesN/AN/AN/A[69]Camero (2021)ASD: 21 m; TD: 20 m20 (10/10)NoFaces, pseudo-objects, pseudo-wordsFixation & gaze duration, pupil dilationN/AN/AN/A[70]Camero (2023)ASD: 27.9 m, TD: 14.4 m114 (57/57)NoCartoons, pseudo-words, pseudo-objectsGaze following & attention timeN/AN/AN/A[71]Chevallier C (2015)ASD: 12.2 y, TD: 14.9 y81 (59/22)NoStatic, dynamic & interactive visual explorationSocial prioritization scoreN/AInteractive: 0.721; Dynamic: 0.623; Static: 0.448N/A[72]Chukoskie (2018)13.9 y8 (8/0)YesVideo gamesFixation duration, saccade patterns, spatial attentionN/AN/AN/A[73]Cilia (2021)ASD: 7.58 y; non-ASD: 8 y59 (29/30)NoSocial vs. non-social videos/photos (joint attention)Visual patterns via scan paths83/800.9CNN[74]Congiu (2024)ASD: 56.63 m; TD: 60.42 m38 (19/19)NoHuman faces and objectsPrioritization, preference, sustained attentionN/AN/AN/A[75]de Belen (2023)ASD: 4.57 y; TD: 4.61 y77 (60/17)NoJoint attention task with actor & toysGaze shift scores, response timeN/AN/AN/A[27]de Belen (2024)ASD: 4.63 y; TD: 4.61 y74 (57/17)NoGeoPref Test dynamic stimulusVisual attention100/76.470.91CNN-LSTM[25]Fernandez-Lanvin (2024)14 m122NoSocial stimuli (video scenes)Social engagement, gaze exchanges, gaze following94.3/94.70.9845 (RF)RF, Naive Bayes, KNN, AdaBoost, ANN, SVM, J48, PART[76]Fletcher-Watson (2018)ASD: 23.06 m, TD: 23.25 m39 (19/20)YesApp screenshots, cartoons, photographsFixation duration, AOI, TFFN/AN/AN/A[77]Fong (2019)ASD: 68 m, TD: 61 m40 (20/20)NoColoured overlaysOcular performance (fixation duration/count, visit duration), reading speed/errorsN/AN/AN/A[78]Franco (2023)TD: 6.7 y, ASD: 6.8 y77 (33/44)NoSocial scenes, geometric scenesFixation points69/930.82RF, SVM, ANN[79]Frazier (2016)ASD: 4.9 y, TD: 5.8 y79 (40/39)YesStill pictures, dynamic videosFixation duration, fixation count, gaze distribution80/820.86N/A[6]Frazier (2018)ASD: 5.7 y, TD: 6.8 y201 (91/110)YesFacial affect, gaze following, joint attention, social images, naturalistic scenesGaze patterns, saccades, blinksN/A0.86N/A[80]Griffin (2020)TD: 13.9 y, ASD: 13.5 y70 (35/35)NoGaze following/perception tasksVisual attention, gaze shifts-N/AN/A[81]Harrop (2019)ASD (M): 9.53, ASD (F): 8.33, TD (M): 7.68, TD (F): 7.93 y74 (42/32)NoDyadic & parallel play scenesTotal fixation durationN/AN/AN/A[82]He (2021)HFA: 5.08 y, LFA: 4.98 y, TD: 5.24 y74 (26 HFA/24 LFA/24 TD)YesGaze stimuli (human, cartoon, finger, arrow)Reaction time to AOI, fixation length79.2/900.938KNN[83]Hudry (2025)TD: 44 m, ASD: 40 m196 (96/100)NoSocial/non-social scenes, faces, referential attentionPeople vs. geometry, eyes vs. mouth, PLD, RJA82/700.82Classification Algorithm[84]Ioannou (2020)N/A73 (12/25)NoStatic social images (low vs. high complexity)Fixation duration on faces, TFF90.0/78.0N/AN/A[85]Jones (2023)Discovery: 22.4 m, Replication: 25.4 m1089 (519/570)NoSocial interaction videosSocial visual engagementN/A0.89, 0.90N/A[86]Jones (2023)TD: 23.4 m, ASD: 24.9 m335 (150/185)NoSocial interaction videos (Eye-tracking index test)Social visual engagement78/85.40.9N/A[87]Kang (2020)ASD: 4.52 y, TD: 4.64 y157 (77/80)NoFace photos (own-race, other-race, familiar)Corneal reflections, gaze direction, fixation duration, saccade featuresN/A0.89K-means, linear SVM[88]Kang (2020)ASD: 4.29 y, TD: 4.26 y97 (49/48)NoFace photos (own-race young girl), dynamic kitten videoFixation time in AOIN/A0.93SVM[89]Keehn (2024)ASD: 2.6 m; TD: 2.4 m146 (102/44)NoGeoPref test, gap-overlap paradigm, PLR, animated cartoonNon-social preference, no-shift %, PLR latency/amplitude, fixation durations77.5/77.30.93NA[38]Kong (2022)Toddler ASD: 2.4 y; Toddler Control: 2.2 y; Preschool ASD: 4.6 y; Preschool control: 4.8 y173 (92/81)NoDynamic stimuli (rabbit, faces, biological motion, objects, infant face)Gaze patterns, fixation timeToddlers: 80/82; Pre-schoolers: 81/61N/ASVM[41]Kou (2019)ASD: 3.72 y; TD: 4.10 y66 (32/34)NoDynamic videos (dancing adult) & geometric shapesSocial attention via visual preferenceN/A0.764 (fixation counts for DSI)N/A[90]Krol (2019)ASD: 16.27 m; TD: 16.31 m44 (21/23)NoFacial features, emotional recognitionFace scanning patternsN/AN/AEmotion classification model[91]Król (2019)ASD: 16.27 m, TD: 16.31 m44 (21/23)YesFace images (FACES database, various expressions)Eye fixations, Scan Match similarityN/AN/ALDA[92]Lencastre (2024)7.74 y57 (28/29)NoNon-organic distractor (balloon), handwaving cartoonGaze analysis95.1/93.1N/ATime neural network[93]Lian (2023)ASD: 11.17 y, TD: 8.96 y49 (23/26)YesApp-based UI tasks (icons, buttons, labels)Number of fixations, TFF, total fixation durationN/AN/AN/A[94]Liu (2016)ASD: 7.90 y, TD: 7.86 y87 (29/29; 29)YesFacial images for recognitionFixation duration93.1/86.20.8963SVM (RBF kernel)[95]Lockwood (2024)ASD: 4.50 y; TD: 4.20 y78 (46/32) + 26 IDNoProsaccade/antisaccade stimuli (circles, animal animations)Prosaccades, antisaccadesN/ATask1: 0.82, Task 2: 0.74, Task 3: 0.74N/A[96]Lohan (2018)ASD: 128.63 m, TD: 126.38 m53 (24/29)YesFaces, houses, abstract patterns, avatar interactionPupil diameter, gaze location, AOIN/AN/AN/A[97]Meng (2023)39.70 m161 (117/44)NoVisual stimulusAttention patterns, engagement75/N/A0.81CNN, RF[98]Moore (2018)ASD: 30.0 m, TD: 29.7 m277 (76/51)NoDynamic social vs. geometric patternsFixation duration, saccade patterns35.00/94.000.75N/A[99]Muratori (2019)ASD: 25.1 m; TD: 31.7 m27 (12/15)NoJoint attention video clipsJoint AttentionN/AN/AN/A[100]Oliveira (2021)3–18 y106 (76/30)NoVideos of biological vs. geometric movementsFixation points69/930.822 (ANN with GA); 0.775 (SVM)SVM, ANN/VAMs[101]Ozdemir (2022)ASD: 34.85 m; TD: 32.90 m133 (61/72)NoSocial interaction vs. non-social toy/animation videosDifferential visual attention patternN/A0.708DT, NB, RF, SVM[102]Pierce (2016)ASD: 28 m, TD: 23.6 m444 (152/115)NoSide-by-side DGI & DSI imagesFixation duration/patterns, saccade numbers, AOI21/99N/AN/A[103]Robain (2022)ASD: 3.19 y, TD: 2.95 y110 (90/20)YesCartoon & real-life videosLooking time, social orienting, fixation duration/count, TFFN/AN/AN/A[104]Rudling (2024)10 m/14 m/18 m169 (EL-ASD: 35/EL-non-ASD: 94; TD: 40)NoDirect gaze interactive taskPercentage fixation, TFF latencyN/AN/AN/A[105]Samad (2018)ASD: 13 y, TD: 16 y16 (8/8)Yes3D facial expressions, facial recognitionGaze fixation duration, eye-hand coordination correlationN/AN/ASVM (RBF kernel)[106]Scherf (2018)10–18 y84 (34/50)NoVideo clips of actors expressing emotionsJoint attention, gaze trajectories, fixation durationN/AN/AN/A[107]Shic F (2022)ASD: 8.55 y; TD: 8.51 y399 (280/119)NoActivity, social, static, biomotion, PLR tasksOculomotor Index of Gaze to Human Faces (OMI)/Biomotion/PLRN/AN/AN/A[108]Sosnowski (2022)8.56 y54 (54/0)YesFacial emotion recognition tasksEmotion recognition scores pre/post-interventionN/AN/AN/A[109]Sun (2023)ASD: 3.1 y, TD: 2.9 y59 (32/27)NoNon-social objects; electronic & transport imagesGaze pattern & analysisADOS: 91/78.7, RRB: 90/77.40.812N/A[110]Sun (2024)ASD: 3.1 y; TD: 2.9 y59 (32/27)NoVisual stimulusGaze patterns66.7/70.60.75Multimodal Logistic Regression[111]Tenenbaum (2020)ASD: 65.65 m; TD: 49.95 m176 (93/40)YesSocial interaction (dyadic, joint attention, sandwich-making, moving toy)Frame-by-frame fixations, scan path distanceN/AN/AN/A[112]Thompson (2018)8.52 y16 (16/0)YesSinging (song) and speech (story)Fixation count, dwell timeN/AN/AN/A[113]Thorup (2016)HR: 10.25 m, LR: 10.27 m64 (47/17)YesLive interaction (experimenter, puppets)Gaze shifts, gaze fixationN/AN/AN/A[114]Tsuchiya (2021)ASD: 10.3 y; TD: 9.5 y141 (39/102)YesSocial vs non-social scenes, preferential paradigmsGaze fixation %, AOI74/800.84Best fit diagnostic algorithm[115]Vacas (2021)ASD: 55.89 m, TD: 53.53 m38 (19/19)NoStatic paired images (faces vs. CIO/non-CIO objects)Prioritization, preference, duration (sustained attention)N/AN/AN/A[116]Wadhera (2021)ASD: 16.32 m; TD: 25.43 m60 (30/30)YesMaze-based cognitive learning taskFixation duration, saccade amplitude, EEG metricsN/AN/AN/A[117]Wan (2019)ASD: 4.6 y, TD: 4.8 y74 (37/37)YesDynamic rabbit, video of a female mouthing alphabetGaze patterns, fixation time86.5/83.8N/AN/A[118]Wang (2018)ASD: 3.73 y, TD: 3.98 y39 (19/20)YesShort cartoon videosScreen dwelling, gaze positions, fixationsN/AN/AN/A[119]Wang (2024)ASD: 3.80 y; non-ASD: 3.30 y39 (22/17)NoDynamic videos (eyes vs mouth)AOI Switch Counts (ASC), Favoured AOI Shifts (FAS), AOI Vacancy Counts (AVC), Total Gaze Count (TGC), Total Fixation Time (TFT)88–100/80–88N/AN/A[42]Wang (2025)ASD: 4.3 y, TD: 4.46 y45 (22/23)YesSocial stimuliAttention & response100/1001ResNet18 + Bi-LSTM[120]Wei (2024)Training: 4.1 y; Validation: 3.1 y529 (290/239)YesSocial stimuliGaze movements & pattern82.2/69.40.849Random Forest[121]Wen (2022)24.37 m1863 (725/1138)NoDynamic social vs. geometric side-by-side videoFixation to geometric AOIs, ASDGeo subtype17.00/98.00 (69 fixations); 23.00/95.00 (CV); 33.00/95.20 (fixations + saccades)N/AN/A[122]Xiong (2024)18–24 m76 (41/35)NoVisual stimulusGaze movements & pattern87.8/45.70.684Binary Logistic Regression[123]Zacharov (2022)ASD: 53.54 m, TD: 49.50 m37 (13/24)NoAnimated cartoon train (A-not-B WM task)Looking duration at AOIs, first-gaze accuracyN/AN/AN/A[124]Zhang (2019)ASD: 59.40 m, TD: 62.30 m34 (15/19)NoTarget/distractor stimuli (squares, circles)Saccade amplitude, peak velocity, durationN/AN/AN/A[125]Zhang (2021)ASD: 71.67 m, TD: 70.21 m34 (15/19)NoEmotional face distractors (angry, happy, neutral)Saccadic errors, saccade latencies (RDP)N/AN/AN/A[126]Zhang (2021)3–6 y42 (21/21)NoET tasks (e.g., “NativeFace1/2”), static & dynamicET metrics (fixation duration, gaze entropy) + EEG featuresN/AN/ARF (ET feature selection) + GCN (fusion classifier)[127]Zhang (2022)ASD: 5.37 ± 0.74 y; TD: 5.54 ± 0.75 y; TD adults: 25.44 ± 2.62 y299 (143/113 children + 43 adults)NoDynamic video clips (actors pointing/looking)RJA events, valid sampling rateN/A (not reported)0.82Logistic Regression[128]Zhang (2025)ASD: 3.89 y; TD: 4.56 y59 (33/26)NoInteractive, visualEye movements97/88.50.978ResNet-50+ LSTM[129]Zhou (2018)ASD: 5.7 y; TD (5y): 5.6 y; TD (4y): 4.6 y75 (26/49)NoAuditory sentences + visual imagesAnticipatory/competitor looks, fixation latency, gaze proportionN/AN/AN/A[130]Note. Age units: m, months; y, years. Participant groups: ASD, autism spectrum disorder; TD, typically developing; HR, high risk; LR, low risk; M/F, male/female; Disc, discovery cohort; Repl, replication cohort; HFA, high-functioning autism; LFA, low-functioning autism; DLD, developmental language disorder; ID, intellectual disability. Stimulus: App UI, application user-interface tasks. Key metrics: AOI, area of interest; TFF, time to first fixation; EEG, electroencephalography RJA, responsive joint attention; PLR, pupillary light reflex; WM, working memory; DGI, dynamic geometric images; DSI, dynamic social images; Dwell time, total time the eyes fixate on a defined AOI. Performance/analysis: AUC, area under the receiver operating characteristic curve; ROC, receiver operating characteristic; N/A, not available or not assessed. ML/methods: SVM, support vector machine; RBF, radial basis function; RF, random forest; NB, naïve Bayes; ANN, artificial neural network; CNN, convolutional neural network; LSTM, long short-term memory; GRU, gated recurrent unit; BiLSTM, bidirectional LSTM; CNN-LSTM, convolutional neural network with LSTM; KNN, k-nearest neighbours; LDA, linear discriminant analysis; J48, C4.5 decision tree (Weka); PART, partial decision tree; VAMs, visual attention models; GA, genetic algorithm; UASN, Uncertainty-inspired ASD Screening Network. Clinical tools: ADOS, Autism Diagnostic Observation Schedule; CARS, Childhood Autism Rating Scale; RRB, restricted, repetitive behaviours Hardware-based eye-tracking studies for autism spectrum disorder (ASD). Summary of included studies reporting sample characteristics, stimuli, gaze metrics, machine learning methods, and diagnostic performance Note. Age units: m, months; y, years. Participant groups: ASD, autism spectrum disorder; TD, typically developing; HR, high risk; LR, low risk; M/F, male/female; Disc, discovery cohort; Repl, replication cohort; HFA, high-functioning autism; LFA, low-functioning autism; DLD, developmental language disorder; ID, intellectual disability. Stimulus: App UI, application user-interface tasks. Key metrics: AOI, area of interest; TFF, time to first fixation; EEG, electroencephalography RJA, responsive joint attention; PLR, pupillary light reflex; WM, working memory; DGI, dynamic geometric images; DSI, dynamic social images; Dwell time, total time the eyes fixate on a defined AOI. Performance/analysis: AUC, area under the receiver operating characteristic curve; ROC, receiver operating characteristic; N/A, not available or not assessed. ML/methods: SVM, support vector machine; RBF, radial basis function; RF, random forest; NB, naïve Bayes; ANN, artificial neural network; CNN, convolutional neural network; LSTM, long short-term memory; GRU, gated recurrent unit; BiLSTM, bidirectional LSTM; CNN-LSTM, convolutional neural network with LSTM; KNN, k-nearest neighbours; LDA, linear discriminant analysis; J48, C4.5 decision tree (Weka); PART, partial decision tree; VAMs, visual attention models; GA, genetic algorithm; UASN, Uncertainty-inspired ASD Screening Network. Clinical tools: ADOS, Autism Diagnostic Observation Schedule; CARS, Childhood Autism Rating Scale; RRB, restricted, repetitive behaviours The major differences between the studies were the sample size, the stimuli or setting of the test and finally the end result accuracy of differentiating between autistic versus control groups. Between 2015 and 2025, one of the largest studies was by Wen et al., which validated the use of Geometric Preference (GeoPref) test, an eye-tracking measure of social attention, as a biomarker for early autism detection. The study recruited a total of 1,863 toddlers with mean age of 24.4 months including ASD, ASD-feature, developmental delay, language delay, typically developing, and typical sibling of subject with ASD. The study concluded that children with autism are accurately identified using the 69% fixation threshold of GeoPref test with specificity 98% and sensitivity of 17% when increase saccades data was included [122]. Another large study was by Jones et al., who investigated social visual engagement markers for autism assessment in toddlers under 3 years. Across discovery (n = 719) and replication (n = 370) cohorts totalling 1,089 children, the index test showed high diagnostic accuracy (AUCs 0.90 and 0.89; sensitivities 81.9% and 80.6%; specificities 89.9% and 82.3%). Social visual engagement measures, incorporating social disability, verbal ability, and nonverbal cognition, were benchmarked against reference standard diagnoses. The results support eye-tracking objective measures as a valid aid for autism diagnosis and assessment [87]. Other researchers have focused on developing machine-learning models to improve screening accuracy. Wei et al. trained classifiers on 529 children (290 autistic, 239 controls; ages 1.5–6 years), using 449 participants for training and 80 for testing. The best model achieved an AUC of 0.85 with 76.5% accuracy, though the small test set highlights the need for validation [121]. Pierce et al. recruited 334 toddlers across ASD, developmental delay, other medical conditions, and typical development, and merged these with a prior cohort of 110, yielding a sample of 444 toddlers. The study demonstrated that a subset of toddlers with ASD showed a strong preference for geometric over social images, with fixation > 69% of the time on geometric images with 21% sensitivity and 99% specificity, suggesting the (GeoPref) test detects homogenous subtypes within the spectrum [103]. Across these studies, autistic participants of heterogenous age groups consistently showed reduced attention to social stimuli. The specific tasks varied (viewing images/videos versus live interactions like name-calling or joint attention), but all supported the potential of eye-tracking as a screening tool for ASD. The specific tasks varied, but collectively these studies support eye-tracking as a candidate digital biomarker for ASD. Most employed case-control classification designs with internal validation; only a small subset used prospective or externally validated designs approaching screening-level evidence. Many models demonstrated high classification accuracy, but often in small samples or in older children, indicating the need for larger validation studies in younger toddlers [68]. Notably, some high-accuracy models (>85%) were developed in samples of older children (mean ages 7–8 years) [63, 74, 95, 131], whereas very few focused on children aged 4–5 years [83, 89]. A handful of studies did focus on toddlers (10 months to 5 years) with modest sample sizes and varying accuracy [41, 70, 77, 110, 114, 123]. Wearables in eye-tracking Recently, there has been a shift towards wearable technologies, representing a significant step in moving eye-tracking beyond the confines of the laboratory into more naturalistic and interactive settings. Table 2 summarizes 8 out of 90 studies employing wearable eye trackers; these range from case-control classification studies to exploratory group-difference analyses and feasibility evaluations, and should not be interpreted as prospective screening studies. The complete wearable dataset is provided in Supplementary Table S3. The studies mainly differed in the stimuli used as they ranged from image or video-based tasks, to live interaction or face to face conversations usually integrated into the ADOS-2 assessment, or finally in a virtual reality (VR) environment.Table 2Wearable-based eye-tracking studies for autism spectrum disorder (ASD). Summary of included studies reporting sample characteristics, stimuli, gaze metrics, machine learning methods, and diagnostic performanceFirst Author (Year)Age Group (Mean)Sample Size (Total; ASD/Control)Home Based PotentialStimulus TypeKey MetricsSensitivity (%)/Specificity (%)AUC ScoreML AlgorithmRefAhn (2023)39.97 m66 (61/-)NoSmile behaviour/Social gaze occurrencesSocial gaze detection, eye contactN/AN/ACNN[132]Boyd (2022)9–18 y11 (11/0)YesStill images and luminance-manipulated photosChanges in gaze fixations54.2/62.5N/AN/A[133]Freeth (2019)ASD: 35.1 y, TD: 34.8 y25 (12/13)YesLive human interaction (dialogue)Gaze direction, % of time spent looking at face, eyes and mouth while speakingN/AN/AN/A[134]Minissi (2024)ASD: 53.14 m, TD: 57.88 m81 (39/42)NoSocial stimuliAttention and eye movements57/750.75Linear SVM[135]Nag (2020)ASD: 12.13 y, TD: 11.53 y33 (16/17)YesImages from the CAFÉ dataset (emotions)Gaze patternN/AN/AElastic Net Classifier, Baseline Classifier[136]Schaller (2021)ASD: 11.25; Non-ASC: 12.21; NTD: 12.41y65 (45/20)NoLive interaction with clinician (ADOS-2 session)Dwell time on eyes, face, mouthN/AN/AN/A[137]Zhao (2021)ASD: 99.6 m, TD: 108. 8 m39 (19/20)NoLive face-to-face conversational interactionVisual fixation, session length84.21/100.000.92SVM (best); LDA; decision tree; random forest[138]Zhao (2023)ASD: 99.6 m; TD: 108.8 m39 (19/20)NoLive face-to-face conversational interactionASD-like segments; eye movement entropy; fixation to eyes/mouth89.00/85.00N/ASVM[139]Note. Age units: m, months; y, years. Participant groups: ASD, autism spectrum disorder; TD, typically developing; Non-ASC, Non-autism spectrum condition; NTD, Non-typically developing. Stimulus/resources: CAFÉ, Child Affective Facial Expression dataset. Performance reporting: AUC, area under the receiver operating characteristic curve; N/A, not available or not assessed. ML/methods: SVM, support vector machine; LDA, linear discriminant analysis; RF, random forest. Clinical tools, ADOS-2, Autism Diagnostic Observation Schedule, Second Edition Table 2Wearable-based eye-tracking studies for autism spectrum disorder (ASD). Summary of included studies reporting sample characteristics, stimuli, gaze metrics, machine learning methods, and diagnostic performanceFirst Author (Year)Age Group (Mean)Sample Size (Total; ASD/Control)Home Based PotentialStimulus TypeKey MetricsSensitivity (%)/Specificity (%)AUC ScoreML AlgorithmRefAhn (2023)39.97 m66 (61/-)NoSmile behaviour/Social gaze occurrencesSocial gaze detection, eye contactN/AN/ACNN[132]Boyd (2022)9–18 y11 (11/0)YesStill images and luminance-manipulated photosChanges in gaze fixations54.2/62.5N/AN/A[133]Freeth (2019)ASD: 35.1 y, TD: 34.8 y25 (12/13)YesLive human interaction (dialogue)Gaze direction, % of time spent looking at face, eyes and mouth while speakingN/AN/AN/A[134]Minissi (2024)ASD: 53.14 m, TD: 57.88 m81 (39/42)NoSocial stimuliAttention and eye movements57/750.75Linear SVM[135]Nag (2020)ASD: 12.13 y, TD: 11.53 y33 (16/17)YesImages from the CAFÉ dataset (emotions)Gaze patternN/AN/AElastic Net Classifier, Baseline Classifier[136]Schaller (2021)ASD: 11.25; Non-ASC: 12.21; NTD: 12.41y65 (45/20)NoLive interaction with clinician (ADOS-2 session)Dwell time on eyes, face, mouthN/AN/AN/A[137]Zhao (2021)ASD: 99.6 m, TD: 108. 8 m39 (19/20)NoLive face-to-face conversational interactionVisual fixation, session length84.21/100.000.92SVM (best); LDA; decision tree; random forest[138]Zhao (2023)ASD: 99.6 m; TD: 108.8 m39 (19/20)NoLive face-to-face conversational interactionASD-like segments; eye movement entropy; fixation to eyes/mouth89.00/85.00N/ASVM[139]Note. Age units: m, months; y, years. Participant groups: ASD, autism spectrum disorder; TD, typically developing; Non-ASC, Non-autism spectrum condition; NTD, Non-typically developing. Stimulus/resources: CAFÉ, Child Affective Facial Expression dataset. Performance reporting: AUC, area under the receiver operating characteristic curve; N/A, not available or not assessed. ML/methods: SVM, support vector machine; LDA, linear discriminant analysis; RF, random forest. Clinical tools, ADOS-2, Autism Diagnostic Observation Schedule, Second Edition Wearable-based eye-tracking studies for autism spectrum disorder (ASD). Summary of included studies reporting sample characteristics, stimuli, gaze metrics, machine learning methods, and diagnostic performance Note. Age units: m, months; y, years. Participant groups: ASD, autism spectrum disorder; TD, typically developing; Non-ASC, Non-autism spectrum condition; NTD, Non-typically developing. Stimulus/resources: CAFÉ, Child Affective Facial Expression dataset. Performance reporting: AUC, area under the receiver operating characteristic curve; N/A, not available or not assessed. ML/methods: SVM, support vector machine; LDA, linear discriminant analysis; RF, random forest. Clinical tools, ADOS-2, Autism Diagnostic Observation Schedule, Second Edition The first study to use wearables for screening of autism was by Nag et al. in 2020 who developed Google Glass fitted with Microsoft LifeCam HD-6000 Webcam to examine gaze and emotion recognition in 33 children (16 ASD and 17 neurotypical, aged 6–17 years). The stimuli comprised images from the Child Affective Facial Expression (CAFÉ) dataset on screen (happy, sad, angry, scared, disgust, surprised, and calm). Gaze and emotion recognition features distinguished ASD from neurotypical children with 81.2% accuracy but did not significantly outperform classifiers using only age and gender. Despite capturing relevant behaviours, classification performance did not surpass simple demographic models, limited by substantial technical challenges including device bulk, calibration burden, noisy data, and non-naturalistic stimuli [136]. Another study by Boyd et al. 2022 recruited only 11 autistic children aged 9–18 years and measured the effect of manipulated image luminance on gaze and verbal behaviours. The participants had a head mounted eye tracker, and eye gaze fixation along with verbal responses were measured. Results indicated that darker non-salient image regions increased fixations on salient areas, while lighter backgrounds enhanced global verbal responses. Interaction effects with spatial frequency and age-related differences were noted [133]. The study by Minissi et al. in 2024 employed a novel VR system called CAVE environment while using the Tobii Pro Glasses 2 for eye-tracking. The study was done on 81 children aged 3–7 (39 ASD, 42 control) and the eye-tracking models underperformed due to technical challenges especially when compared to what the motor bio signals classification yielded [135]. The next group of studies utilised live interactions as the test stimuli rather than image/video-based tasks. Schaller et al. used SMI mobile eye-tracking glasses to classify 65 children and adolescents (ages 8–18) (45 suspected Autism Spectrum Condition (ASC), 20 controls) conducted within a live clinician-administered ADOS-2 session, this study is best characterised as a diagnostic differentiation study rather than a scalable screening design. The significant difference in this study compared to others is that the measures were taken during live interaction with a clinician during the standard ADOS-2 assessment, integrating eye-tracking into the evaluation. The study measured dwell time on the face (eyes and mouth) and concluded that neurotypical individuals fixated on the central facial region more than those with ASD, and these metrics correlated with clinical ratings [137]. Another study by Zhao et al. used real life interviews with Tobii Pro Glasses 2 in 39 participants (19 ASD and 20 control), aged 6–13 years, after 4 of the 43 initially enrolled participants were excluded because of technical eye-tracking problems. They paired eye-tracking metrics based on visual fixation across AOIs with session-length features in a machine learning model and were able to differentiate the ASD group from controls with 92.3% classification accuracy [138]. Interestingly, a study by Ahn et al. used adult-worn Pivot head and Orca camera glasses rather than child worn glasses to capture continuous, high-resolution egocentric video of children during ADOS-2 assessment, minimizing participant burden and avoiding intrusive wearable placement (observer effect). The study recruited 66 children aged 24–68 months with suspected and subsequently confirmed ASD diagnosis (61 confirmed ASD, 5 developmental delay). The automated measurements the proportion of time spent in social gaze with and without smiling towards parents predicted ADOS-2 severity [132]. Zhao et al. used Tobii Pro Glasses 2 with 43 children aged 6–13 years (20 ASD, 23 control). The study measured oculomotor behaviours by analysing the range of gaze movement during a structured face-to-face conversation and the duration of the session. The machine learning classifier achieved high accuracy (87.18%), sensitivity (89.47%), and specificity (85.00%) in distinguishing ASD from TD based on short gaze behaviour segments. The method preserved data integrity and showed classification performance comparable to gold-standard tools (ADOS-2, ADI-R) [139]. Webcam-based eye-tracking Since COVID-19 pandemic, the interest in accessible remote screening tools has grown, spurring the use of webcam-based solutions. Table 3 summarizes the 5 out of 90 studies that have investigated the use of webcam-based solutions applied to autism-related gaze research, encompassing platform feasibility studies, proof-of-concept classification studies, and one screening-adjacent study combining gaze preference with M-CHAT-R [143]. The complete webcam dataset is provided in Supplementary Table S4.Table 3Webcam-based eye-tracking studies for autism spectrum disorder (ASD). Summary of included studies reporting sample characteristics, stimuli, gaze metrics, machine learning methods, and diagnostic performanceFirst Author (Year)Age Group (Mean)Sample Size (Total; ASD/Control)Home Based PotentialStimulus TypeKey MetricsSensitivity (%)/Specificity (%)AUC ScoreML AlgorithmRefJensen (2021)ASD:60 m; TD:42.8 m101(28/73)NoSocial scenes (children playing) & abstract scenes (moving shapes)Gaze preferenceVisit1- 64/92. Visit 2- 63/90GP-MCHAT-R model: Visit 1 - 0.89, Visit 2 - 0.88.N/A[140]Strobl (2019)18+ y17 (0/17)YesFace images and circular gaze tracking taskEye versus mouth fixation accuracy97.25/98.94N/ASVR[141]Vargas-Cuentas (2017)48.71 m31 (8/23)YesSocial vs moving shapes scenes (split-screen)Left/Right gaze classification accuracyN/AN/AN/A[142]Wang (2019)Adults: 25 y (average);Children: 2 y (average)Adults: 10 (0/10); Children: 7 (5/2)YesResponse-to-name videosFacial landmarks, expressions and gaze directionsN/AN/AResNet18 + BiLSTM[143]Zheng (2016)ASD: 2.19 y, TD: 1.33 y16 (8/8)NoSocial orienting cues, visual stimuli, name-calling stimuliGaze measurement, gaze directionN/AN/AN/A[144]Note. Age units: m, months; y, years. Participant groups: ASD, autism spectrum disorder; TD, typically developing. Stimulus/resources: GP-MCHAT-R, Gazefinder-Modified Checklist for Autism in Toddlers, revised; MCHAT-R: Modified Checklist for Autism in Toddlers, Revised. Performance/analysis: AUC, area under the receiver operating characteristic curve; N/A, not available or not assessed. ML/methods: SVR, support vector regression; ResNet, residual neural network; BiLSTM, bidirectional long short-term memory Table 3Webcam-based eye-tracking studies for autism spectrum disorder (ASD). Summary of included studies reporting sample characteristics, stimuli, gaze metrics, machine learning methods, and diagnostic performanceFirst Author (Year)Age Group (Mean)Sample Size (Total; ASD/Control)Home Based PotentialStimulus TypeKey MetricsSensitivity (%)/Specificity (%)AUC ScoreML AlgorithmRefJensen (2021)ASD:60 m; TD:42.8 m101(28/73)NoSocial scenes (children playing) & abstract scenes (moving shapes)Gaze preferenceVisit1- 64/92. Visit 2- 63/90GP-MCHAT-R model: Visit 1 - 0.89, Visit 2 - 0.88.N/A[140]Strobl (2019)18+ y17 (0/17)YesFace images and circular gaze tracking taskEye versus mouth fixation accuracy97.25/98.94N/ASVR[141]Vargas-Cuentas (2017)48.71 m31 (8/23)YesSocial vs moving shapes scenes (split-screen)Left/Right gaze classification accuracyN/AN/AN/A[142]Wang (2019)Adults: 25 y (average);Children: 2 y (average)Adults: 10 (0/10); Children: 7 (5/2)YesResponse-to-name videosFacial landmarks, expressions and gaze directionsN/AN/AResNet18 + BiLSTM[143]Zheng (2016)ASD: 2.19 y, TD: 1.33 y16 (8/8)NoSocial orienting cues, visual stimuli, name-calling stimuliGaze measurement, gaze directionN/AN/AN/A[144]Note. Age units: m, months; y, years. Participant groups: ASD, autism spectrum disorder; TD, typically developing. Stimulus/resources: GP-MCHAT-R, Gazefinder-Modified Checklist for Autism in Toddlers, revised; MCHAT-R: Modified Checklist for Autism in Toddlers, Revised. Performance/analysis: AUC, area under the receiver operating characteristic curve; N/A, not available or not assessed. ML/methods: SVR, support vector regression; ResNet, residual neural network; BiLSTM, bidirectional long short-term memory Webcam-based eye-tracking studies for autism spectrum disorder (ASD). Summary of included studies reporting sample characteristics, stimuli, gaze metrics, machine learning methods, and diagnostic performance Adults: 25 y (average); Children: 2 y (average) Note. Age units: m, months; y, years. Participant groups: ASD, autism spectrum disorder; TD, typically developing. Stimulus/resources: GP-MCHAT-R, Gazefinder-Modified Checklist for Autism in Toddlers, revised; MCHAT-R: Modified Checklist for Autism in Toddlers, Revised. Performance/analysis: AUC, area under the receiver operating characteristic curve; N/A, not available or not assessed. ML/methods: SVR, support vector regression; ResNet, residual neural network; BiLSTM, bidirectional long short-term memory The initial study in 2016 by Zheng et al. included only 16 toddlers (12–24 months), 8 with ASD diagnosis and 8 control. They employed 5 mounted monitors that comprised an intelligent learning environment where the prompt level required for participants to respond to their name (name calling prompts) and the time taken from initial prompt to target fixation (Animated attention attractors) was measured using Logitech webcams. The study validated the use of such extensive system to administer social orienting cues focusing on early social orienting skills [144]. Another study in 2017 used a camera to record the child’s face and later on ocular movements were extracted from the video using an algorithm to depict visual preferences. The study included 23 control children and 8 ASD children all aged between 2 and 6 years. The study used the recorded video and the algorithm developed to detect child’s face, process the image frame, undergo feature extraction of the iris and the pupil and finally identify the gaze and hence the gaze preference. The study concluded that autistic children spent significantly less time viewing social scenes than control children while both manual and algorithmic evaluations showed this difference, with an error in the algorithm of maximum 1.52% [142]. In 2019, Strobl et al. conducted a proof of principle study on 17 healthy adult volunteers to assess the feasibility and accuracy of a smartphone-based eye-tracking software. The iTracker software was able to successfully distinguish between gaze directed towards the eye and the mouth of a face displayed on the smartphone screen, however, some biases and classification errors are still under study and development [141]. In 2021, an interesting study by Jensen et al. investigated the effectiveness of gaze preferences along with M-CHAT-R as a screening tool for ASD in children aged between 36 and 99 months of age. The study included 73 control and 28 autistic children at the initial visit with a follow-up visit one month later with 1 drop out in the autistic group. Gaze preference was manually coded from the initial 15 seconds of the video as they were more sensitive to differences than the whole 50 seconds video and quantified as the proportion of time the child focused on a social scene, an abstract scene, the centre of the screen, or being distracted. The combined tool demonstrated better performance of 0.89 AUC compared to each tool alone, with the autistic children spending significantly less time gazing at the social scene compared to control group [140]. Another study by Wang et al. in 2021 utilised the response to name protocol integrated into a multi sensor vision system for automated screening where 10 control adults (average 25 years), 2 control children and 5 autistic children (average 2 years) were recruited. The metric measured was the child’s response to their name being called from different positions, quantified using head pose estimation and gaze estimation detected using Kinect sensor and two red, green, blue wavelength (RGB) cameras. The classification score for the system was 92.7% in evaluating the response to name task, factors like rapid head movements, minimal head turns or inconsistent call intervals by the implementer resulted in some errors. Despite these errors, the system’s scores showed similar results with minor differences compared to the clinical assessments and closely mirrored clinical observations: typically developing participants exhibited a much higher frequency of social orienting than the ASD group [143]. Home-based eye-tracking Home-based eye-tracking represents an advance in decentralized assessments as it overcomes the conventional drawbacks of conventional clinical-based ASD screening methods. Following recent literature, three recent studies (3.3% out of 90 papers) were concluded to offer distinct methodologies and advance the field through innovative approaches and critical reflection on implementation challenges relevant to home-based ASD screening feasibility. Importantly, none of these studies reported externally validated diagnostic accuracy; each addresses feasibility or proof-of-principle objectives and should be read as providing evidence that home deployment is technically achievable, not that it is clinically validated [145–147]. The home-deployed dataset is provided in Table 4 and expanded in Supplementary Table S5.Table 4Home-based eye-tracking studies for autism spectrum disorder (ASD). Summary of included studies reporting sample characteristics, stimuli, gaze metrics, machine learning methods, and diagnostic performanceFirst Author (Year)Age Group (Mean)Sample Size (Total; ASD/Control)Home Based PotentialStimulus TypeKey MetricsSensitivity (%)/Specificity (%)AUC ScoreML AlgorithmRefBradshaw (2023)4–8 m40 (25/15)YesSocial interactions (parent play, toy play)Gaze location, distribution & behavioursNANAN/A[145]Kim (2024)ASD: 32.1 y; TD: 33 y37 (15/22)YesVisual stimulus (sitcoms,interviews, trailers)Gaze duration & behavioursNANAN/A[146]Varma (2022)2–15 y95 (68/27)YesInteractive games (Guess What mobile app)Gaze fixation patterns, visual scanning patternsNANALSTM[147]Note. Age units: y, years. Participant groups: ASD, autism spectrum disorder; TD, typically developing. Performance/analysis: AUC, area under the receiver operating characteristic curve; N/A, not available or not assessed. ML/Methods: LSTM, long short-term memory Table 4Home-based eye-tracking studies for autism spectrum disorder (ASD). Summary of included studies reporting sample characteristics, stimuli, gaze metrics, machine learning methods, and diagnostic performanceFirst Author (Year)Age Group (Mean)Sample Size (Total; ASD/Control)Home Based PotentialStimulus TypeKey MetricsSensitivity (%)/Specificity (%)AUC ScoreML AlgorithmRefBradshaw (2023)4–8 m40 (25/15)YesSocial interactions (parent play, toy play)Gaze location, distribution & behavioursNANAN/A[145]Kim (2024)ASD: 32.1 y; TD: 33 y37 (15/22)YesVisual stimulus (sitcoms,interviews, trailers)Gaze duration & behavioursNANAN/A[146]Varma (2022)2–15 y95 (68/27)YesInteractive games (Guess What mobile app)Gaze fixation patterns, visual scanning patternsNANALSTM[147]Note. Age units: y, years. Participant groups: ASD, autism spectrum disorder; TD, typically developing. Performance/analysis: AUC, area under the receiver operating characteristic curve; N/A, not available or not assessed. ML/Methods: LSTM, long short-term memory Home-based eye-tracking studies for autism spectrum disorder (ASD). Summary of included studies reporting sample characteristics, stimuli, gaze metrics, machine learning methods, and diagnostic performance Visual stimulus (sitcoms, interviews, trailers) Note. Age units: y, years. Participant groups: ASD, autism spectrum disorder; TD, typically developing. Performance/analysis: AUC, area under the receiver operating characteristic curve; N/A, not available or not assessed. ML/Methods: LSTM, long short-term memory To begin with, Varma et al. introduced a game-based mobile app that coded gaze behaviour during natural family interactions using computer vision algorithms in 2022. This approach enabled collection from children with a wider developmental range (ages 2–15), adopted privacy-centred automated annotation, and resulted in statistically significant gaze differences between cohorts. However, dependence on indirect gaze estimation increased its susceptibility to noise, interpretability was stalled by changing camera positions, predictive accuracy was biased by imbalance in groups, and external validation against gold-standard metrics was missing. Hence, the reported machine learning model exhibited only moderate predictive power [147]. Bradshaw et al. established the use of head-mounted eye-tracking (HMET) for infants as young as four months in home settings, including infants at elevated genetic risk for ASD in 2023. The design highlighted custom-fitted equipment revealing high feasibility and retention, which reinforces the probability for rich longitudinal data collection within naturalistic setting. Interestingly, the proportion of successful data acquisition was exceptional, and parental acceptance was high. However, the study encountered shortcomings in gaze estimation accuracy within younger infants, difficulties in calibration depending on pupil-only methods, and like other methods, relied heavily on parental reports for behavioural assessment. The exclusion of between-group analyses and motor/postural factors limited the ability to relate findings directly to atypical developmental pathways [145]. Finally, Kim et al. in 2024 provided the first proof-of-principle evidence for smartphone-based gaze estimation in home environments. In a small study limited to high-IQ adults (n = 37, mean age 32 years), the system reproduced expected group-level gaze differences between ASD and typically developing participants, with privacy-preserving video processing addressing data security considerations. However, no diagnostic accuracy metrics were reported, no children were included, and generalisation to paediatric or community samples cannot be assumed [146]. Building on the three home-deployed studies collectively, home-based eye-tracking can be conceptualised as a three-phase process: at-home gaze data acquisition, AI-powered analysis, and clinician-facing outputs as illustrated in Fig. 4. Supplementary Table S6 synthesises gaze features across all 90 studies into nine categories: (1) social attention to faces/eyes (25+ studies, AUC 0.82–0.98); (2) social versus geometric preference (specificity 94–99%, sensitivity 17–38%); (3) joint attention gaze shifts (AUC 0.68–0.82); (4) saccade dynamics (strong discriminators but requiring ≥ 250 Hz, incompatible with webcam deployment); (5) pupillary responses; (6) temporal gaze sequences (highest reported discrimination, AUC 0.87–1.0, from small samples); (7) dwell time/AOI engagement (broadest webcam deployability); (8) biological motion response; and (9) pupillary light reflex (AUC 0.93; requires controlled-luminance hardware). For each category, the table documents diagnostic performance ranges, home-deployability, deployment constraints, and best-performing ML approaches, providing a structured evidence map for feature selection in next-generation system design.Fig. 4Conceptual framework for home-based AI-Enabled autism screening. This framework illustrates a proposed three-phase pipeline. In phase 1, a child engages with webcam-based visual tasks from as early as 18 months of age, while parents complete structured developmental questionnaires. In phase 2, raw gaze plots and clinical data are securely uploaded for feature extraction and behavioural analysis using machine learning models trained on ASD datasets. The system generates risk scores categorized as low, moderate, or high. In phase 3, results are presented on a secure clinician dashboard, where findings are reviewed in context and linked to outcomes. This workflow emphasizes non-invasiveness, ecological validity, privacy-preserving design, and scalability, providing an accessible at-home screening pathway Conceptual framework for home-based AI-Enabled autism screening. This framework illustrates a proposed three-phase pipeline. In phase 1, a child engages with webcam-based visual tasks from as early as 18 months of age, while parents complete structured developmental questionnaires. In phase 2, raw gaze plots and clinical data are securely uploaded for feature extraction and behavioural analysis using machine learning models trained on ASD datasets. The system generates risk scores categorized as low, moderate, or high. In phase 3, results are presented on a secure clinician dashboard, where findings are reviewed in context and linked to outcomes. This workflow emphasizes non-invasiveness, ecological validity, privacy-preserving design, and scalability, providing an accessible at-home screening pathway Discussion The evidence across 90 studies traces a clear technological trajectory from dedicated laboratory hardware to home-deployed consumer systems. Each stage solved specific limitations of its predecessor while introducing new ones, as summarised below. Hardware-based systems established the gaze biomarker evidence base through high spatial and temporal resolution, enabling the most robust and replicated findings in the literature to date [148, 149]. Their non-portability, high cost, and dependence on specialist operation in tightly controlled environments fundamentally preclude clinical scaling [148, 149]. The principal design lesson is that calibration-free, ambient-light-robust systems are needed to translate these findings beyond the laboratory [42, 74, 145, 150]. Advances have produced more compact, screen-mounted eye trackers that reduce costs and bulkiness of the test, but these still require expert personnel, extensive calibration, and controlled environment [42, 141, 151, 152]. Wearable technology introduced naturalistic, egocentric measurement of gaze in ecologically valid settings, uniquely capturing both gaze direction and the participant’s concurrent visual scene. High device cost, mechanical complexity, and poor tolerability in young children constrain the evidence base to only eight included studies, limiting generalisability. While data richness is high, home scalability remains low; wearable systems are best positioned as specialised research instruments rather than scalable screening tools, and further miniaturisation and child-appropriate design are prerequisites for broader deployment [138, 153]. Webcam-based systems made the pivotal contribution of consumer-device accessibility at no additional hardware cost, substantially lowering the barrier to large-scale deployment; demand accelerated further during the COVID-19 era as remote and home-based assessment became a clinical priority. Reduced spatial precision relative to dedicated eye trackers, combined with sensitivity to ambient lighting, head pose variability, and environmental distraction, remains the principal technical barrier to clinical-grade application [74, 141, 152, 154]. The main design lesson is that appearance-based deep learning models are rapidly narrowing the accuracy gap, positioning webcam-based systems as the most plausible platform for population-level screening, provided robustness is outside controlled environments is formally demonstrated [141, 154]. Findings from Clin et al. further reinforce that social attention varies meaningfully with real-life context and familiarity, underscoring the importance of ecologically valid assessment settings [155]. Home-deployed studies provide direct empirical evidence of ecological validity and the practical feasibility of unsupervised screening, capturing children’s naturalistic behaviour with reduced stress, engaging caregivers as active collaborators, and enabling repeated longitudinal measurements [145, 156–159]. Real-time progress tracking through regular monitoring enables personalised assessments [76]. The evidence base currently available lacks externally validated diagnostic accuracy reporting, which substantially limits the conclusions that can be drawn about real-world performance. The existing evidence is insufficient to support deployment recommendations; prospective, multisite validation in diverse and uncontrolled home environments, with prespecified endpoints and rigorous calibration-robustness testing, represents the most critical next step for the field [145–147]. A technically important distinction for system design is between summary gaze statistics and temporally structured gaze features. Most included studies extracted summary metrics such as fixation proportion, dwell time, time to first fixation that collapse gaze dynamics into scalar values, discarding information about transition order, orientation latency, and scanning regularity. Between 8 and 10 studies employed sequence-aware models (LSTM, BiLSTM, CNN-LSTM, GRU) that preserve these temporal dynamics; more broadly, machine learning approaches have been applied to improve diagnostic assessment of ASD in the presence of differential and co-occurring diagnoses. Such models generally reported higher discrimination (AUC 0.87–1.0 versus 0.71–0.90 for summary-statistic classifiers), though direct comparison is confounded by differences in sample size and validation design [129, 160, 161]. Supplementary Table S6 explicitly classifies each feature category as summary statistic or temporal dynamic. For home-deployed systems, temporal models demand higher and more consistent sampling rates, greater sensitivity to webcam frame dropout, and greater computational overhead, all constraints that are more acute outside controlled laboratory environments. As synthesised in Supplementary Table S6, social attention features represent the most replicated and home-deployable category, while saccade dynamics and PLR although highly discriminative, remain hardware-bound. This mapping has direct implications for next-generation system design. A growing body of work on calibration-free gaze estimation is directly relevant to home-based feasibility, given that toddlers, who are the primary target population, are the least likely to tolerate conventional multi-point calibration. Deep-learning appearance models can estimate gaze without explicit calibration; the Pupil Labs Neon system achieved acceptance rates of 74–94% in infants aged 3–24 months [162]. For webcam-based systems, WebGazer.js self-calibrates from cursor–gaze coordination, eliminating explicit calibration sessions entirely [163]. Kaduk et al. (2023) demonstrated that a modern webcam-based system achieves 1.4° accuracy, approaching laboratory-grade standards [152]. The emergence of calibration-free approaches specifically reduces the compliance burden for toddlers [78, 88, 145, 146]. Beyond calibration, the transition from laboratory to home introduces a distribution shift that degrades model performance. Laboratory-trained classifiers face systematic domain shifts in home environments where lighting, viewing distance, and hardware vary freely. Distribution shift also operates at the person level such as variability in facial morphology, skin tone, and head movement can reduce gaze estimation accuracy for underrepresented populations [164, 165]. Among the 90 included studies, very few explicitly evaluated cross-setting or cross-population generalisation. The largest prospective study (Jones et al., 2023; n = 499) employed a six-site design, but did not study laboratory-to-home transfer [87]. However, Kim et al. (2024) was the only paper to perform a direct cross-method comparison between smartphone-based gaze estimation in laboratory and home settings and a conventional laboratory eye-tracking benchmark. The study found low gaze-estimation error in both smartphone conditions and replicated reduced face-looking in autistic participants, but was limited to a small, high-IQ adult sample [146]. Multi-environment validation designs and domain adaptation techniques will be needed to maintain performance across the heterogeneous conditions characteristic of home screening. Diagnostic metrics reported across included studies are not directly comparable, owing to four sources of heterogeneity: (1) validation strategy, ranging from internal cross-validation to prospective multisite design; (2) comparator standard (ADOS-2, DSM-5, M-CHAT-R/F, or clinical judgement); (3) population composition (balanced case-control versus community-screened); and (4) task design and feature extraction. Only a small subset employed externally validated designs, notably Jones et al. (2023; AUC 0.90, n = 499) [87], Frazier et al. (2018) [80], and Keehn et al. (2024; AUC 0.93, PPV 88.8%) [38, 166]. Supplementary Tables S2–S5 provide validation method, comparator tool, PPV, NPV, and confidence intervals for contextualised interpretation. An important distinction exists between screening, case-control classification, and feasibility evaluation, as each implies different data requirements, validation standards, deployment expectations, and analytical frameworks [167]. Screening demands high sensitivity at low community base rates with brief, non-specialist-administered tools; case-control classification tests discrimination under optimised conditions; and feasibility studies evaluate acceptability and compliance without diagnostic endpoints. High accuracy in a balanced case-control design does not predict real-world screening performance, where community base rates are low; for example, recent U.S. surveillance estimates place ASD prevalence at about 2.8%, substantially reducing positive predictive value [2]. In Supplementary Tables S2–S5, the ‘Study Design’ column classifies each study by purpose, and the ‘Biomarker Relevance’ column grades evidence as: validated biomarker (externally validated, prospective), candidate biomarker (internally validated), exploratory marker (group-level differences only), or non-diagnostic (feasibility). The broader literature on multivariate biomarker-based diagnosis of ASD similarly emphasises that convergent evidence across multiple marker modalities is needed before any single biomarker can be considered clinically sufficient [168]. Based on the Biomarker Relevance classification in Supplementary Tables S2–S5, only 2 studies meet the validated biomarker threshold; both prospective evaluations of the same eye-tracking index test [86, 87]. Additionally, 42 are classified as candidate biomarkers (internally cross-validated), 30 as exploratory markers (group-level differences only), and 16 as non-diagnostic (feasibility or intervention studies without classification outcomes). Reported sensitivity, specificity, and AUC values should therefore be interpreted in the context of each study’s design tier rather than as population-level screening performance estimates. Limitations of this scoping review should be considered when interpreting its conclusions. Restricting searches to PubMed and Scopus and to English-language publications published between 2015 and 2025 may have excluded relevant studies, particularly conference proceedings and preprints that are common in fast-moving engineering fields; given the pace of innovation in gaze estimation and edge computing, the grey literature likely contains relevant feasibility work not captured here. The absence of a formal search of grey literature and non-indexed sources therefore likely biases the evidence map toward published, peer-reviewed, and positive results. Consistent with scoping methodology, no formal quality appraisal was conducted, meaning that studies with weak designs contribute equally to the evidence map as those with strong ones; the Biomarker Relevance grading in Supplementary Tables S2–S5 partially mitigates this by signalling validation tier, but does not substitute for a systematic quality assessment. Finally, methodological heterogeneity precluded quantitative synthesis, and the translational conclusions drawn are necessarily interpretive rather than statistically derived. Towards clinical translation Translating this evidence into clinical practice requires addressing three sequential challenges: validation, regulation, and deployment. The regulatory landscape is addressed first, as it defines the evidence thresholds that validation studies must meet; clinical integration and deployment considerations follow. The predominantly internal cross-validation evidence base underscores the distance between current performance and regulatory-grade deployment. Pathways differ across jurisdictions. In the United States, FDA De Novo or 510(k) clearance applies to software-as-a-medical-device screening tools; the EarliTec EarliPoint Evaluation received FDA 510(k) clearance in June 2023, the first eye-tracking device cleared for ASD evaluation, providing a regulatory benchmark [86, 169]. In Europe, CE marking under EU MDR 2017/745 requires a conformity assessment with clinical investigation data. In the UAE, MOHAP and Dubai Health Authority regulate medical devices under an EU MDR-aligned framework requiring compliance with Dubai Clinical Practice Guidelines for ASD [170, 171]. In Australia, regulation falls under the Therapeutic Goods Administration. These distinctions should directly inform the design of future prospective validation studies, aligning endpoints with jurisdiction-specific requirements. Building on the webcam advances noted above, clinical integration success depends on standardised protocols, caregiver training, and reliable data collection and interpretation [120, 129, 146, 172]. Although most of the reviewed studies were done in labs, adapting these methods for home use will require usability testing, technical support, clear procedures for parents, and toddler-friendly interfaces and stimuli [83]. Machine learning and deep learning pipelines, particularly calibration-free appearance models and uncertainty-aware classification frameworks, are central to this adaptation, as they automate feature extraction, compensate for reduced hardware precision, and enable real-time risk stratification on consumer devices [129, 160]. Integrating telehealth systems with electronic health records will enable remote monitoring and timely clinical decision-making, consistent with the growing evidence base for telehealth interventions in ASD communities [173]. On a public health level, home-based eye-tracking presents an affordable way to expand ASD screening, cut costs, and support early interventions with lasting developmental and economic benefits [76, 145, 150]. Several deployment considerations are relevant for home-based systems. First, on-device (edge) inference avoids uploading video data to remote servers, preserving privacy and enabling use in low-bandwidth or offline settings. Lightweight architectures such as MobileNet and quantised CNN-RNN pipelines now support real-time gaze estimation on consumer smartphones and single-board edge devices such as NVIDIA Jetson and Raspberry Pi, with model quantisation reducing inference time by approximately 20% without meaningful accuracy loss [174, 175]. Cloud inference introduces bandwidth requirements and data governance obligations that may be prohibitive in underserved settings. Second, home-based assessment requires caregivers as co-administrators, shifting control from trained personnel to non-expert users. Standardised instruction materials, built-in quality-assurance feedback, and automated gating for low-confidence recordings are essential design requirements. Third, device heterogeneity across socioeconomic groups may introduce systematic bias if systems are optimised for high-end hardware, and equity considerations should be explicitly addressed. The three home-deployed studies collectively demonstrate that ecological feasibility and clinical validity are separate challenges requiring sequential rather than simultaneous resolution. All three studies successfully demonstrated technical and caregiver acceptability as infants tolerated head-mounted systems [145], adults completed unsupervised smartphone assessments [146], and caregivers administered game-based mobile tasks with children [147], yet none reported externally validated diagnostic accuracy metrics. This pattern suggests that the first prerequisite for home-based deployment has been met: the technology can be administered outside a clinic. The outstanding challenge is clinical validation: demonstrating that gaze patterns captured in these conditions meet the sensitivity, specificity, and reproducibility thresholds required for screening use. Future work must therefore move beyond feasibility studies to prospective diagnostic accuracy trials conducted in home settings, with pre-specified endpoints aligned to the regulatory benchmarks discussed above. Drawing on the evidence from 90 studies, a next-generation home-based ASD screening system should combine: (1) calibration-free webcam or tablet front-camera gaze estimation; (2) brief (<5 min), age-appropriate video stimuli targeting social attention and biological motion, the most replicated and home-deployable feature categories; (3) on-device inference for privacy and low-bandwidth operation; (4) hybrid feature extraction combining summary gaze statistics with temporal sequence models, aligned with emerging multimodal frameworks that integrate visual, linguistic, and behavioural features [176]; (5) adaptive quality-assurance gating for low-confidence recordings; (6) parent-facing interfaces with standardised instructions and real-time engagement feedback; and (7) prospective, multisite, cross-cultural validation against a clinical gold standard with pre-specified endpoints before regulatory submission. No existing system meets all seven criteria. The EarliTec EarliPoint Evaluation, the only FDA-cleared eye-tracking ASD assessment tool, satisfies criteria (2), and (7) but uses proprietary tablet hardware rather than a consumer webcam. Conclusion This scoping review mapped 90 studies across four eye-tracking platform stages, hardware (n = 74), wearable (n = 8), webcam (n = 5), and home-deployed (n = 3) to assess the translational readiness of gaze-based digital biomarkers for early ASD screening in home settings. The evidence establishes that social attention, gaze-following, and biological motion response carry robust discriminative signal for ASD, but this evidence derives almost entirely from laboratory-based, case-control studies with internal validation. Only three studies collected data in genuinely home-deployed settings, and none reported externally validated diagnostic accuracy. The field’s central challenge is therefore not whether eye-tracking biomarkers discriminate but whether that discrimination can be reproduced on consumer-grade, calibration-free platforms in uncontrolled home environments with diverse populations. Priorities for the next phase include: (1) prospective, multisite diagnostic accuracy trials conducted in home settings with prespecified endpoints; (2) standardised stimulus protocols and reporting metrics to enable cross-study comparison; (3) evaluation of model robustness to distribution shift across devices, lighting, and populations; and (4) navigation of jurisdiction-specific regulatory pathways, informed by the EarliPoint FDA 510(k) precedent. Until these gaps are addressed, home-based eye-tracking should be regarded as a promising adjunct to existing screening pathways rather than a replacement, with integration into clinical practice contingent on the evidence standards outlined in this review. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary material 1 Supplementary material 1