Advantages of combining multiple eye-tracking paradigms for distinguishing young autistic from non-autistic children Authors: Dan Xu, Lan Zhang, Xiangqin Wang, Xun Zeng, Linghong Huang, Yanan Qing, Menghan Zhou, Qin Li, Xiaojiao Yang, Weihua Zhao, Shuxia Yao, Jiao Le, Keith M Kendrick Published: 2026 ABSTRACT AbstractBackgroundSingle-paradigm, single-measure eye-tracking protocols have demonstrated utility in distinguishing between autistic and non-autistic children although effect sizes and reproducibility vary. There is a need for brief, scalable digital behavioral biomarkers that integrate complementary information from multiple eye-tracking paradigms and achieve improved accuracy in combination.Methods74 autistic and 63 non-autistic children aged 24–72 months performed five different eye-tracking paradigms (facial emotion processing, gaze-following, dynamic social versus geometric patterns, social interaction, and spinning) lasting 3.25 min in hospital or kindergarten settings. A broad set of fixation-based metrics was extracted from each paradigm. We compared discrimination performance of single-paradigm models versus a combined multi-paradigm model using random forest (RF) classifiers. In the autistic group, symptom severity was assessed using standardized clinical measures. We further evaluated whether paradigms contributed complementary information, estimated potential clinical value of multi-paradigm models, and conducted exploratory subgrouping analyses to examine whether eye-tracking–defined profiles aligned with symptom-based groupings.ResultsAll individual paradigms distinguished between autistic and non-autistic children, but RF models found a combined paradigm-based model performed best, achieving an AUC of 95% and an accuracy of 90%. Representational similarity analysis indicated that paradigms contributed partially distinct information rather than reflecting a single redundant dimension of social attention. Decision curve analysis demonstrated that the multi-paradigm model provided added net benefit across clinically relevant threshold probabilities compared with strategies based on treating all or no children as autistic. Clustering of eye-tracking features revealed three autistic subgroups with distinct visual preference profiles, whereas clustering based on clinical symptoms alone identified only two subgroups.LimitationsSex imbalance and group differences in developmental quotient may have confounded some effects. Models were only internally cross-validated in a single cohort and decision curve analysis relied on assumed clinic prevalence, so external validation and testing in larger, more diverse and high-risk samples, including other neurodevelopmental conditions, are needed.ConclusionsA brief multi-paradigm eye-tracking battery yields robust case–control discrimination and non-redundant behavioral readouts, suggesting that it may complement symptom-based approaches and provide a scalable behavioral framework for future studies seeking to bridge molecular findings with observable autistic profiles.Supplementary InformationThe online version contains supplementary material available at 10.1186/s13229-026-00718-z.Keywords: Autism spectrum disorder, Multi-eye-tracking paradigms, Visual preference, Social attention, Gaze pattern Background Autism is a highly heterogeneous condition. To better understand this diversity, studies have leveraged phenotypic/genetic [1], biological [2], neural [3, 4] or behavioral assessments [5] to identify autistic subgroups to aid diagnosis and inform therapy selection. However, the complex mapping from genes and brain function to behavioral performance has not yet been clarified. This unresolved mapping remains a central challenge in translational autism research [6–8]. Stable and quantifiable behavioral profiles may help address this gap. In this context, non-invasive eye tracking offers an objective, simple, and scalable tool for standardized measurements [9]. Prior work has demonstrated group-level differences between autistic children and non-autistic peers using single paradigms, usually only one eye-tracking measure, such as fixation duration, despite there being many other potentially informative measures [10]. To date, the main eye-tracking paradigms used include face scanning [11], ‘GeoPref’ (preference for geometric patterns) test [12–14], gaze following [15, 16], and biological motion perception (BM) [17, 18], which probe attention to salient facial features (e.g., the eyes) [19, 20], gaze-following [21, 22], and preferences for social versus non-social stimuli [10, 23, 24]. Autistic children often show less interest in the eyes of faces, gaze following [16, 25, 26], and in social relative to non-social stimuli [24, 27, 28]. However, while some studies have included a number of individual paradigms [24, 29] they rarely investigate whether each contributes nonredundant information [30], directly compare multi-paradigm models with single-paradigm models using rigorous cross-validation and statistical testing [31–33], or examine how paradigm-specific representational structure aligns with symptom dimensions. Furthermore, few studies quantify potential clinical utility [30] via decision-curve analysis (DCA) or evaluate whether data-driven subgroups derived from eye-tracking features recapitulate conventional symptom-based groupings [13]. Guided by this rationale, we selected five paradigms that map onto orthogonal dimensions of social cognition and restricted interests and included a combination of multiple eye-tracking measures. Two conventional paradigms were included which assessed interest in facial stimuli [16, 34, 35] (a facial emotions paradigm, FE and a gaze following paradigm, GF). Three other paradigms assessed visual preferences for social versus non-social stimuli. First, a novel social interaction (SI) paradigm presented videos featuring a child playing with a parent and a table of unattended toys. A second novel paradigm (Spinning) contrasted paired videos of spinning children and objects since autistic children are often attracted to spinning objects and also like to spin themselves (stimming) [36]. A third established paradigm [8, 10, 20] contrasted paired videos of dancing children with geometric patterns (DSG), sometimes referred to as the ‘GeoPref’ test [12, 14, 24]. We excluded BM paradigms since BM’s discriminative value in autism is inconsistent and often confounded by low-level stimulus properties and complexity, with heterogeneity across samples contributing to variable effects [37, 38]. The total duration of the five paradigms is around 3.25 min, which can balance the feasibility of data collection and keep different features [29, 39]. Increasingly, data-driven approaches are being used to identify putative autistic subgroups based on genetic architecture [6, 40], neuroanatomy [41], functional connectivity [42] and cognitive profiles [43] and these may differ in terms of early adaptation and response to intervention. However, objective, scalable behavioral readouts have rarely been used as the primary basis for such stratification. Multi-paradigm eye-tracking profiles, in particular, may yield complementary subgroups defined by distinct visual attention patterns, which could have implications for both case–control discrimination and the characterization of within-autism variation. Only one single-paradigm study reported possible subgroups of dynamic geometry and social stimuli preferences in autistic children [14], highlighting the need to evaluate whether broader multi-paradigm profiles provide a more robust and interpretable view of heterogeneity. Taken together, prior eye-tracking findings in autism have been inconsistent, in part because most studies rely on a single paradigm and/or a single metric, while paradigms differ in the processes they probe and autism is highly heterogeneous. To address these limitations, the current study tests a multi-paradigm, multi-metric framework across five complementary eye-tracking paradigms. We ask whether integrating information across paradigms improves discrimination of autistic from non-autistic children, if different paradigms capture complementary dimensions of social-attention behavior and how these dimensions relate to symptom severity, if the resulting classifier provides potential clinical value and finally if data-driven subgroups derived from eye-tracking converge with symptom-based profiles. Methods Study design and participants This cross-sectional study analyzed eye-tracking and clinical assessment data from 24 to 72 -month-old children diagnosed with autism (n = 77, 64 males, mean age = 55.70, SD = 13.65) and non-autistic peers (n = 64, 41 males, mean age = 57.79, SD = 14.12). Groups were frequency-matched on age (t = -1.173, p = 0.243) but not on sex (χ2 = 6.674, p = 0.01, phi = -0.218). Four children could not be included in the final analysis due to data collection issues leaving a final total of 74 autistic and 63 non-autistic children (details are provided in Fig. S1). All analyses were adjusted for age and sex to account for residual imbalance. Autistic children were assessed with the Autism Diagnostic Observation Schedule–Second Edition [44] (ADOS-2). Both autistic and non-autistic individuals were assessed using the Gesell Developmental Schedules [45] (five subdomains). Caregivers of the autistic children also completed assessments including the Social Responsiveness Scale–Second Edition [46] (SRS-2) and Repetitive Behavior Scale–Revised [47, 48] (RBS-R). The Gesell score revealed expected differences in all developmental sub-scales for autistic relative to the non-autistic children (Table 1). The exclusion criteria were: (1) autistic children: ADOS-2 comparison score 5; (2) visual and auditory problems such as weak vision, color blindness, astigmatism, etc. Non-autistic children with visual problems, neurological disorders or mental disorders were excluded. The study was approved by the Ethics Committee of The University of Electronic Science and Technology of China (1061420210829007) and conducted in accordance with the latest revision of the Declaration of Helsinki. Written informed consent was obtained from caregivers in all cases. Table 1Demographic and participant characteristics informationAutistic children n = 74Non-autistic children n = 63t/χ2 p Cohen’s dSexF/M = 14/60F/M = 21/423.720.054-Age (months)54.9313.2657.7913.741.240.22− 0.212Gesell Motion57.7894.4-15.481< 0.001-2.654 Fine Motor51.2117.8794.8110.52-17.013< 0.001-2.916 Adaptive52.8818.2890.4610.29-14.479< 0.001-2.482 Language418.2492.1612.07-18.633< 0.001-3.194 Personal-social48.7818.2192.8512.12-16.363< 0.001-2.805ADOS-2 SA13.803.03- RRB2.741.59- Total score16.543.56-CARS34.843.45-SRS-2 total score102.4223.7942.6416.8816.585< 0.00012.871RBS-R total score21.4514.28Gesell: Chinese version of the Gesell Developmental Schedule; AODS-2: the Autism Diagnostic Observation Schedule, Second Edition; SA: social affect; RRB: repetitive and restrictive behavior; CARS: The Childhood Autism Rating Scale; SRS-2: Social Responsiveness Scale, Second Edition; RBS-2: the Repetitive Behavior Scale Revised-2 Demographic and participant characteristics information Gesell: Chinese version of the Gesell Developmental Schedule; AODS-2: the Autism Diagnostic Observation Schedule, Second Edition; SA: social affect; RRB: repetitive and restrictive behavior; CARS: The Childhood Autism Rating Scale; SRS-2: Social Responsiveness Scale, Second Edition; RBS-2: the Repetitive Behavior Scale Revised-2 Eye-tracking paradigms and AOIs Five eye-tracking paradigms were administered: FE, GF, SI, Spinning and DSG, for details see Supplementary Materials. Gaze was recorded with a 300 HZ Tobii TX300 eye tracker. Areas of interest (AOIs) were defined a priori for each paradigm; for FE, eyes, nose, and mouth; for GF: face, body, target consistent area (TCA), target inconsistent area (TIA); SI, parent upper/lower areas (Up_Pa, Bot_Pa), child upper/lower areas (Up_Chi, Bot_Chi) and toys; Spinning, rotating children and objects and DSG, dancing children and geometric patterns, see Fig. 1A top-row. The second and third rows of Fig. 1A show illustrative heatmaps for the non-autistic and autistic groups, respectively. These heatmaps are based on cumulative fixation counts (i.e., the number of fixations accumulated within the same stimulus exemplar within the same condition using the a priori AOI definitions) and do not incorporate fixation duration. They are provided as an initial qualitative visualization rather than inferential evidence; statistical comparisons are reported in the Results and Table 2. For each paradigm, we derived multiple established metrics: total fixation duration percentage (TFD%), fixation count percentage (FC%), time to first fixation (TFF), and first fixation duration (FFD). In GF, dynamic indices captured the proportion of trials with correct gaze-following (“following rate”), latencies to initiate following (“point latency”) and looking-back behavior via the EyetrackingR package (version 0.2.2, Dink & Ferguson, 2015, https://cran.r-project.org/web/packages/eyetrackingR). Each clip in GF lasts 5 s and is divided into 250 time bins, with the first second serving as the baseline. Since cue-onset time points were not precisely marked, this study employed derivative and curvature methods to identify wave peaks and valleys. The time of the first peak after baseline correction is defined as the cue response latency. The time window from the first peak to the first valley is defined as the looking-back behavior. The average fixation ratio between the first and last peaks is defined as the gaze-following ratio. Methodological details are provided in the supplementary materials. Trials where stimuli were viewed < 25% of time were excluded from data analysis similar to previous studies [24, 49, 50]. Flowchart of Eye-tracking paradigms, AOI fixation-count heatmaps, and analysis pipeline. (A) The first row shows five paradigms in the following order: facial emotional processing (FE), gaze-following with eye- or point-cue (GF), social interaction (SI), spinning children versus objects (Spinning), and dancing people versus geometric patterns (DSG). The second and third rows show group-level fixation-count heatmaps for non-autistic and autistic children, respectively. Heatmaps reflect cumulative fixation counts aggregated across participants for the same stimulus exemplar within each condition, aligned to the representative example shown in the ROI panel. Pooling across counterbalanced stimulus variants was avoided to preserve consistent AOI positioning. Heatmaps are provided for qualitative illustration; statistical comparisons are reported in subsequent analyses. (B) shows the steps (1→4) of building random forest model within-sets, then calculating model performance (aggregation), paradigm contribution analyses including cumulative paradigm learning and permutation importance. Representational similarity analysis compared the alignment between paradigm-paradigm, paradigm-label, and paradigm-symptom (SRS-2 & ADOS-2). Decision curve analysis (DCA) measured the efficacy of paradigms at threshold probabilities corresponding to plausible specialty-clinic prevalence (10%-20%). Exploratory cluster analysis on all eye-tracking features and symptoms, respectively for subgrouping Table 2Summary of primary group differences across five eye-tracking paradigmsParadigmAOIPrimary metricsAutistic children (%)Non-autistic children (%)t/WEffect size (g)pFDRFEEyeTFD%15.0 ± 1.5928.5 ± 1.59-6.111-1.063< 0.0001FEEyeFC%19.9 ± 1.5127.6 ± 1.52-3.616-0.6290.0007FENoseTFD%24.2 ± 1.5917.6 ± 1.592.9370.5110.0065FENoseFC%19.1 ± 1.5710.0 ± 1.625.2910.921< 0.0001GFPointPeak latency(s)2.0 ± 0.1281.46 ± 0.111538.5-< 0.001GFGazePeak latency(s)1.79 ± 0.151.76 ± 0.151981.5-0.866GFPointLooking-back0.25 ± 0.040.49 ± 0.043571.5-< 0.001GFPointGaze-following0.22 ± 0.020.26 ± 0.0142788.5-0.069SIParents_UpperTFD%35.5 ± 1.9161.7 ± 1.96-9.683-1.667< 0.0001SIParents_UpperFC%34.1 ± 1.7456.2 ± 1.78-8.977-1.545< 0.0001SIToysTFD%28.3 ± 1.9111.1 ± 1.966.3651.096< 0.0001SIToysFC%28.5 ± 1.7413.2 ± 1.786.2281.072< 0.0001SpinningObjectsTFD%61.1 ± 1.9843.7 ± 2.036.2721.101< 0.0001SpinningObjectsFC%56.2 ± 1.8948.8 ± 1.942.8090.4930.0086DSGDancing childrenTFD%42.5 ± 2.6267.1 ± 2.64-6.75-1.184< 0.0001DSGDancing childrenFC%49.7 ± 2.0968.8 ± 2.11-6.544-1.148< 0.0001SpinningChildrenTFF(s)1.12 ± 0.050.64 ± 0.0506.9181.11< 0.0001DSGGeometric patternsTFF(s)2.23 ± 0.260.55 ± 0.2574.7010.72< 0.0001SIParents_bottomTFF(s)1.88 ± 0.173.26 ± 0.173-5.770.962< 0.0001FEEyeFFD(s)0.25 ± 0.150.29 ± 0.091690-0.01DSGGeometric patternsFFD(s)0.91 ± 0.120.39 ± 0.1223.0610.4960.016SIChildren_bottomFFD(s)2.12 ± 0.181.13 ± 0.193.849-0.440.0005Note: TFD%: percentage of total fixation duration; FC%: percentage of fixation counts; TFF: time to first fixation; FFD: first fixation duration. P values (pFDR) are Benjamini–Hochberg false discovery rate–adjusted within each paradigm and dependent variable Summary of primary group differences across five eye-tracking paradigms Note: TFD%: percentage of total fixation duration; FC%: percentage of fixation counts; TFF: time to first fixation; FFD: first fixation duration. P values (pFDR) are Benjamini–Hochberg false discovery rate–adjusted within each paradigm and dependent variable Data analysis pipeline Group differences All dependent variables (TFD%, FC%, TFF, FFD) were compared for between-group differences using group (autistic/non-autistic) * AOIs mixed-effects models with regressed age and sex by the R package (nlme 3.1–167: 10.32614/CRAN.package.nlme). For GF, we also compared temporal differences after the gaze/pointing cues see supplementary material. Nonparametric analyses were conducted for group differences in the proportion of correct gaze-following (TCA) across different time phases. Multiple comparisons were controlled using Benjamini-Hochberg false discovery rate (FDR) correction. Within each paradigm and outcome (TFD%, FC%, TFF, FFD), Benjamini-Hochberg FDR was applied to all post hoc pairwise EMM contrasts for the group × condition effect. Random Forest (RF) model processing pipeline and following analyses were conducted as depicted in Fig. 1B. All features were z-standardized within the training sets to prevent information leakage. Age and sex effects were removed from predictor variables by linear regression fit in the training data and applied to both training and held-out sets. For principal component analysis (PCA), loadings were estimated in the training sets and applied to transform held-out data. We trained RF classifiers to distinguish autistic from non-autistic individuals using PCA-reduced features. Within each of 100 repetitions of 10-fold stratified cross-validation, we: (1) fit PCA on the residualized training data and retained components explaining > 90% cumulative variance; (2) trained a RF model (500 trees; mtry = ⌊√p⌋) on the PCA scores; and (3) generated predictions for the held-out set. We evaluated area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, precision, negative predictive value (NPV), and F1 score across (a) single-paradigm models and (b) a combined model that concatenated features from all five paradigms. AUCs were compared using DeLong tests with FDR adjustment for multiple pairwise contrasts. Confidence intervals (CIs) were estimated by stratified bootstrap resampling of held-out predictions within repetitions. Paradigm-level contribution analyses To quantify the incremental utility of paradigms, we conducted a group-level accumulation analysis. Cumulative models were built by sequentially adding paradigm-specific feature blocks in a pre-specified order (GF → Spinning → DSG → FE → SI), repeating the full cross-validation procedure at each step. For each addition, we computed ΔAUC relative to the preceding model and tested significance via paired bootstrap of the cross-validated predictions across repetitions. At the feature level, permutation importance was estimated by randomly permuting each feature within held-out sets, p-values were obtained from permutation-based null distributions and FDR-adjusted. Detailed rankings are provided in Fig. S6. Representational similarity analyses To assess non-redundancy among paradigms, we constructed representational dissimilarity matrices (RDMs) from subject-wise feature profiles. Pairwise dissimilarity was defined as 1 − Pearson correlation. Three RDM variants were computed: (1) pooled RDM that included all participants, (2) group-demeaned RDM in which between-group mean differences were removed, and (3) within-group RDM computed separately within the autistic and non-autistic groups. To probe clinical relevance, we performed two alignment analyses. Paradigm–label alignment quantified the Spearman correlation between each paradigm’s RDM and a binary group-membership matrix. Paradigm–symptom alignment quantified the correlation between each paradigm’s RDM and symptom-distance matrices derived from SRS-2, ADOS-2 scores, including total and social affect, SA. This analysis explicitly quantifies paradigm–clinical relevance by testing whether the between-subject dissimilarity structure within each paradigm covaries with symptom-derived dissimilarity structure, beyond group-level separation. Significance was assessed via permutation test for inter-paradigm dissimilarities, labels, and symptom scores with FDR control. We also contrasted autistic-within vs. non-autistic-within paradigm–label distances to quantify heterogeneity. More details are in the Supplementary Materials. Decision-curve analysis Clinical utility was examined via DCA [51]. Because early intervention is a clinical priority and may yield greater lifelong benefits, we calculated net benefit across threshold probabilities from 0.01 to 0.50 for each model. The maximum tolerable false positive rate is set at 99% (). Although the 2022 autism and developmental disabilities monitoring network survey found approximately 1 in 31 children under age 8 had an autism diagnosis, representing about 3.2% [52], neurodevelopmental clinics diagnose approximately 29.3% of autistic children [53]. This study evaluates the net benefit of the combined-paradigm model across three referral rates (10%, 15%, 20%) at different risk thresholds. Net benefit was defined as: where TP and FP are true/false positives, N is sample size, and pt is the risk threshold. Exploratory clustering for subgrouping Unsupervised clustering was performed separately for (1) eye-tracking features from the combined-paradigm model and (2) clinical measures (ADOS-2 total, SA, repetitive and restrictive behavior, comparison score; SRS-2; RBS-R). Features were z-standardized. We applied agglomerative hierarchical clustering with Ward.D2 linkage on Euclidean distances [54]. The optimal number of clusters was selected using a consensus of internal criteria (silhouette width and NbClust majority vote). Cluster characteristics were compared using ANOVA or Kruskal–Wallis tests with FDR correction. To select the top-3 features of each subgroup, ANOVAs were performed on all features with FDR correction. As a sensitivity analysis, we also performed joint clustering of eye-tracking and clinical measures under multiple weighting schemes; results were not robust to weighting and showed weak separation (Table S4). Therefore, we retained separate clustering as the primary analysis. Results Group differences across paradigms Across paradigms, autistic and non-autistic children showed robust between-group differences in primary gaze allocation (TFD%) across key AOIs (Fig. 2; Table 2). In the FE paradigm, autistic children spent less time viewing the eye region (TFD%: 15.0 ± 1.59% vs. 28.5 ± 1.59%; t132 = − 6.111, p < 0.0001, d = − 1.063) but more time viewing the nose (TFD%: 24.2 ± 1.59% vs. 17.6 ± 1.59%; t132 = 2.937, p = .0065, d = 0.511). In the GF paradigm, autistic children exhibited slower responses to pointing (W = 1538.5, p < 0.001) but not gaze (W = 1981.5, p = 0.866), reduced looking-back (W = 3571.5, p < 0.001), and a trend toward increased gaze-following (W = 2788.5, p = 0.069). In the SI paradigm, autistic children allocated less attention to the adult playing with the child (TFD%: 35.5 ± 1.91% vs. 61.7 ± 1.96%; t135 = − 9.683, p < 0.0001, d = − 1.667) and more attention to toys on the table (TFD%: 28.3 ± 1.91% vs. 11.1 ± 1.96%; t135 = 6.365, p < 0.0001, d = 1.096). In the Spinning paradigm, autistic children showed greater attention to spinning objects (TFD%: 61.1 ± 1.98% vs. 43.7 ± 2.03%; t132 = 6.272, p < 0.0001, d = 1.101). In the DSG paradigm, autistic children showed reduced attention to dancing children (TFD%: 42.5 ± 2.62% vs. 67.1 ± 2.64%; t133 = − 6.750, p < 0.0001, d = − 1.184). Corresponding results for FC% (and full AOI-by-paradigm details) are summarized in Table 2. Results of group differences based on mixed-effect model with regressed age and sex. In the FE paradigm, compared to non-autistic, autistic children had significantly less fixation duration and fewer counts on eyes, but more on nose regions. For the GF paradigm under the pointing condition, the left panel shows the dynamic changes of looking proportion to the target-congruent area (TCA), with dots and triangles indicating peak and valley points. The right panel shows that, compared with non-autistic children, autistic children had a significant longer latency to the first peak, a significant lower probability of looking back to the model’s face between the first peak and first valley, and a marginally lower probability of gaze following between the first and last peak. For SI, autistic compared to non-autistic children, showed significantly lower fixation duration and counts on interacting parents but higher on non-social toys. For the Spinning paradigm autistic children showed significantly greater fixation durations and counts on spinning objects and on geometric patterns in the DSG paradigm, while showing significantly less attention to social stimuli (Spinning or dancing children). FDR corrected p: * p < 0.05, ** p < 0.01, *** p < 0.001, the same below TFF and FFD showed paradigm-dependent group differences (Fig. S3; Table 2). After FDR correction, group differences in TFF were confined to specific AOIs: autistic children showed longer latency for first fixation on spinning children and geometric patterns, but shorter latency for the lower-body region during parent–child interaction; latencies for other AOIs did not differ reliably (Table 2). For FFD, autistic children showed longer durations for the facial eye region, geometric patterns, and the lower-body region during interaction, whereas no other AOIs showed significant differences after correction (Table 2). Combined-paradigm model performance PCA-based RF results, incorporating all eye-tracking measures and AOIs, indicated that the combined-paradigm model significantly surpassed performance of single-paradigm models (Fig. 3A, accuracy = 90.0%, AUC = 95.3%, DeLong test after correction: ps < 0.05, Fig. S5). The model performance metrics are detailed in Table S2. The relative contribution of each paradigm was evaluated by paradigm accumulation experiments, and when GF, Spinning, DSG, FE and SI tasks were added sequentially, the AUC of the model showed a stepwise significant increase (ΔAUC of each step was statistically significant, P < .05), and the learning curve is shown in Fig. 3B. Model comparisons for paradigms, their complementarity, clinical utility and subgrouping. (A) AUC for combined paradigms reached 0.95 compared to SI (0.87), FE (0.84), DSG (0.82), Spinning (0.78) and GF (0.68). (B) Cumulative paradigm learning curve calculated by adding paradigms in order GF-Spinning-DSG-FE-SI, and shaded area shows bootstrap 95% CIs. (C) Paradigm complementarity and alignment with labels. Top-row: Subject representational dissimilarity matrices (RDMs) compare feature profiles under 3 configurations: pooled (labels ignored), group-demeaned (autistic/non-autistic means removed), and within-group (autistic-only, non-autistic-only). Across configurations, between-paradigm dissimilarity is consistently higher, supporting complementarity. Bottom-row: Correlations between each paradigm’s RDM and a binary diagnostic label matrix (autistic vs. non-autistic) under the same configurations shows strong between-group alignment for DSG and SI; in within-group alignment, SI contributes the strongest within-paradigm structure, followed by FE, Spinning and GF. (D) Net-benefit (NB) across threshold probabilities (pt = 0.01–0.50) for the combined-paradigm model under assumed prevalence π = 10%, 15%, 20%. The optimal pt = 5% (labeled with max NB and ΔNB) gives Se ≈ 0.95–0.97, Sp ≈ 0.79–0.83, and 8.7–18.5 net true positives per 100. (E) Illustrates autistic subgroups derived from visual preference and symptom profiles. Visual-preference clustering yielded three subgroups: subgroup 1 relatively preserved across paradigms except FE; subgroup 2 best in SI then GF; subgroup 3 selective FE preference), whereas symptom-based clustering identified two, with subgroup 1 showing greater severity than subgroup 2 across all domains Non-redundancy of information in paradigms RSA results indicate significant dissimilarity across different eye-tracking paradigms (ps < 0.001). Even after controlling for group differences, this cross-paradigm dissimilarity remained significant (Fig. 3C top row) only when comparing within the same group (autistic-autistic or non-autistic - non-autistic), suggesting that each paradigm captures non-redundant information. Further, paradigm-label alignment analysis revealed that the intergroup separation was most pronounced between SI and DSG (SI:0.407 > DSG:0.268), while the other three paradigms showed moderate alignment. After de-centering, DSG classification is almost entirely driven by mean differences (ρ=-0.035); whereas GF/Spinning/FE, along with a small portion of SI, retain non-redundant label-related structures beyond mean differences. Finally, SI captures intragroup heterogeneity most strongly (ρ = 0.334); followed by FE/Spinning/GF (with intragroup non-redundant information); DSG is weak (Fig. 3C bottom row). Fig. S7A shows that within-group paradigm-label distances were generally larger in autistic than non-autistic children. In the full sample, paradigm-symptom alignment with SRS-2 was strongest for the SI paradigm, see Fig. S7B. For autistic children , SI showed the strongest alignment with ADOS-2 total scores, see Fig. S7D. Detailed task–clinical mapping results are provided in the Supplementary Materials (Fig. S8–S9). Clinical decision curve and subgrouping analysis In decision-curve analysis the combined-paradigm model dominated both treat-all and treat-none across thresholds from pt = 0.05 to 0.50 for each assumed prevalence scenario (π = 10%, 15%, 20%, Fig. 3D). The optimal threshold was 5% in all scenarios where sensitivity and specificity were 0.952/0.828 (π = 10%), 0.952/0.793 (π = 15%), and 0.968/0.793 (π = 20%). The corresponding net benefit (NB) values were 0.087, 0.134, and 0.185, respectively—interpretable as 8.7, 13.4, and 18.5 net true positives per 100 patients relative to treat-none. Furthermore, exploratory cluster analysis based on eye-tracking features partitioned autistic individuals into three distinct clusters (subgroups), whereas clustering based on clinical symptoms only identified two clusters (Fig. 3E). Subgroup 1 children showed greater preference for dynamic social cues within the DSG paradigm in terms of significantly greater fixation duration (d = 1.3, ) and fixation frequency ratios (d = 1.21, ) toward the “dancing” social scene compared to other subgroups. In the Spinning paradigm, they also showed more fixations on social stimuli (d = 1.16, ), revealing a relatively preserved social orientation. Subgroup 2 children showed a preference for geometric/non-social stimuli, exhibiting the highest proportion of fixation duration (d = 1.35, ) in the Spinning non-social condition and the DSG geometric condition (d = 1.19, ). In the gaze-following task, individuals in this subgroup also tended to fixate on body parts (d = 0.859, ) rather than faces, reflecting a pronounced non-social orientation and neglect of key social cues. Subgroup 3 children showed markedly atypical facial scanning patterns, concentrating attention almost exclusively on the nasal region during the FE paradigm. Both the frequency and duration of nasal fixations were significantly elevated (TFD%: d = 2.35, FC%: d = 2.52, ), accompanied by prolonged orientation latencies toward key interactive regions in social interaction paradigm (d = 1.30, ). These subgroup findings are hypothesis-generating and require replication and longitudinal validation. Discussion This work found robust group-level visual preferences across all five eye-tracking paradigms used between autistic and non-autistic children and across different eye-tracking measures. The combined-paradigm model which aggregated features from individual paradigms showed improved performance, reaching 90% accuracy, 95.3% AUC, 91.7% precision, and 90.8% NPV (individual paradigm accuracy: FE, 75%; GF, 66%; SI, 82%; Spinning, 73%; DSG, 75%). RSA revealed that paradigms provide unique and complementary information, demonstrating the advantage of multi-dimensional evaluation. Paradigm–symptom alignment further suggested differential sensitivity of paradigms to dimensional variation in social communication (e.g., SRS-2, ADOS-2), and DCA demonstrates the potential value of the comprehensive model in aiding clinical diagnosis (NB > 0). Finally, eye-tracking characteristics reveal finer distinctions in autistic subgroups than clinical symptoms, suggesting a diversity of social attentional profiles. The consistently observed group differences across all five different paradigms and measures align with, but substantially extend, previous literature that has typically focused on single paradigms and single measures. Meta-analyses consistently document reduced attention to social vs. non-social information in autism [55, 56], but with heterogeneity across paradigms, measures and regions of interest [57]. The DSG paradigm serves as a robust identifier for autistic children, potentially capturing core and shared atypical preferences for geometric patterns, thereby providing strong intergroup separation. This aligns with past DSG studies consistently identifying a well-described subgroup [13, 14]. Furthermore, the DSG paradigm has been associated with default mode network (DMN)-Occipito-Temporal cortex hypoconnectivity in independent research [2], supporting the value of non-social, dynamic geometric content as an informative probe. In contrast, paradigms like GF, FE, and Spinning appeared more sensitive to within-group micro-level variation. The GF paradigm involves atypicalities in initiating/responding to joint attention and altered accuracy/latency profiles, which relate to cognition and symptom severity [15, 58–60]. The variable response to emotional faces in the FE paradigm across autistic individuals is consistent with known heterogeneity in facial emotion processing [61, 62]. The Spinning paradigm, incorporating a mix of social and non-social repetitive motions, may probe circumscribed interests, with recent meta-analytic evidence suggesting that attention in autism is disproportionately captured by stimuli related to specific interests rather than non-social objects generically [63], but this atypical visual attention to non-social objects may impact social affect in naturalistic settings. The SI paradigm, which simulates dynamic, complex social interaction, was unique in capturing both between-group and within-group variance, likely because it concurrently challenges multiple social cognitive processes [64, 65] such as rule understanding [66], attention shifting [67], imitating the dyadic interaction [68], and detail-focused processing bias [69]. Thus, the SI paradigm is most sensitive to the variation in ADOS-2 and SRS-2 clinical symptom scores. High heterogeneity among autistic individuals is well-established [70], manifesting in core symptom structures [71, 72], cognitive [73], verbal-nonverbal abilities [74], sensory-interest patterns [75], and responsiveness to rehabilitation interventions [76]. Multi-paradigm, multi-indicator combinations can reveal attentional stickiness, dynamic changes over time, and more refined attentional exploration patterns [77], aspects that may be undetectable using single-gaze proportion metrics. Similar to Frazier’s ARI [78] and ABC-CT’s OMI metrics (The Autism Biomarkers Consortium for Clinical Trials, ABC-CT) [79], integrating multi-paradigm, multi-indicator data synthesizes social attention patterns across diverse contexts, condensing multi-task information into composite indices reflecting symptom load. Consequently, models combining multiple paradigms and indicators are more likely to enhance individual-level discrimination, and provide complementary evidence for evaluating intervention efficacy. Due to uneven distribution of regional medical resources and limitations in clinical experience, delayed interventions resulting from missed or delayed diagnoses persist [80, 81]. Recent work in primary care and specialty settings demonstrates that eye-tracking–based indices can augment diagnostic workflows when implemented with standardized protocols and leakage-controlled validation [30, 82, 83]. Our DCA indicates that the highest benefit occurs at pt = 5% across different prevalence rates π (10%-20%), demonstrating the model’s robustness in epidemiological contexts. At an assumed prevalence of 10%, screening 100 children would avoid approximately 65 unnecessary interventions compared with a treat-all strategy. This makes it especially suitable as a risk screening tool in pediatric primary care settings to support clinical decision-making. Furthermore, the brevity, scalability, objectivity, quantifiability, and non-intrusiveness of eye-tracking-based assessments, combined with their ability to reveal nuanced processing patterns, make them a promising candidate for monitoring responses to interventions over time [50, 77]. The partial correspondence between gaze-based and symptom-based subgrouping highlights a key result of this study: multi-paradigm eye-tracking may capture within-autism heterogeneity that is not fully reflected by global clinical scores. Clustering based on eye-tracking features identified three distinct subgroups, whereas clustering based on clinical symptoms yielded only two. This suggests that attention-based behavioral profiles may index distinct underlying dimensions [2] and capture variance not reflected in aggregate symptom scores [84], supporting the potential value of richer behavioral phenotyping for precision-oriented research and, ultimately, personalized intervention strategies [85]. For instance, one subgroup showed strong visual preferences in DSG, GF, and Spinning paradigms but weak engagement with FE, a pattern distinct from other subgroups. From a translational perspective, this complementary behavioral construct may be useful in future studies aimed at aligning clinical stratification with genetic, neurological, or treatment response heterogeneity. Importantly, these subgroups are exploratory and hypothesis-generating; their stability, generalizability, and clinical utility require replication in independent cohorts and longitudinal validation. Limitations Several limitations must be acknowledged. The sex imbalance and significant difference in developmental quotients between groups, though common in autism research, and controlled for in the case of sex, may confound some of the observed effects. While internal validation via cross-validation was robust, the model requires validation in an independent, larger sample to ensure generalizability. Our DCA relied on plausible specialty-clinic prevalence assumptions [86]; site-specific calibration and prospective impact evaluation are needed prior to clinical deployment. Finally, while we emphasized five complementary paradigms, future work should test portability across laboratories and extend to other neurodevelopmental conditions to examine specificity and to assess their utility in identifying ‘at risk’ children who subsequently receive an autism diagnosis. Conclusions This study establishes a multi-paradigm eye-tracking framework as a powerful and objective approach to support earlier recognition of autistic children. More importantly, by delineating the unique cognitive contributions of different paradigms and their specific links to clinical symptoms, this work moves beyond simple classification toward a more nuanced, quantitative understanding of autistic heterogeneity. This work lays the foundation for future research integrating eye-tracking with genetic, neural, and longitudinal data to advance stratified assessment and more tailored support for autistic individuals. Electronic Supplementary Material Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements We thank all families who participated in our study and the support of Chengdu Maternal and Children’s Hospital and Shuangqiao Campus of Chengdu No. 3 Kindergarten. Abbreviations The autism biomarkers consortium for clinical trials Autism Diagnostic Observation Schedule-Second Edition Areas of interest Area under curve Biological motion perception Children lower area Parent bottom area Confidence intervals Decision-curve analysis Default mode network Dynamic social versus geometric patterns paradigm Fixation count percentage False discovery rate Facial emotions paradigm First fixation duration False positives Gaze following paradigm Net benefit Negative predictive value Principle component analysis Repetitive Behavioral Scale-Revised Representational dissimilarity matrices Random forest Repetitive and restrictive behavior Representational similarity analysis Social affect Dynamic spinning child versus object Social Responsiveness Scale-Second Edition Target consistent area Total fixation duration percentage Time to first fixation Target inconsistent area True positives Child upper area Parent upper area Author contributions Drs Kendrick and DX take responsibility for the integrity of the data and the accuracy of the data analysis. Concept and design: Kendrick, Le, and DX. Acquisition, analysis, or interpretation of data: All authors. Drafting of the manuscript: DX, Le and Zhang. Critical revision of the manuscript for important intellectual content: All authors. Statistical analysis: DX, XQ Wang Obtained funding: Kendrick and Le. Administrative, technical, or material support: Yao, Zhao, Li, Zhou, Qing, Huang. Supervision: Kendrick, Le. Funding This work was supported by the National Natural Science Foundation of China (NSFC) [grant number 82301732 – JL]; Yingcai Scheme, Chengdu Women’s and Children’s Central Hospital (CWCCH) [grant number YC2023016 – JL]; Key Technological Projects of Guangdong Province “Development of New Tools for Diagnosis and Treatment of Autism” grant number 2018B030335001 (K.M.K.); and Sichuan Province Key Research and Development Project [grant number 2023YFWZ0003 - KMK]. Data availability The dataset supporting the conclusions of this article is available on reasonable request by email directed to the corresponding author and subject to agreement of the funder. Declarations Ethics approval and consent to participate This study was approved by the Ethics Committee of the University of Electronic Science and Technology of China (1061420210829007) and legal guardians of all autistic and non-autistic children signed informed consent. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Dan Xu and Lan Zhang Joint first authors. Contributor Information Jiao Le, Email: lejiao-89@163.com. Keith M. Kendrick, Email: k.kendrick.uestc@gmail.com References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The dataset supporting the conclusions of this article is available on reasonable request by email directed to the corresponding author and subject to agreement of the funder.