Graph Neural Networks and sequential architectures for autism detection from eye-tracking biomarkers: a multi-site study Authors: Nisrine El Ayat, Mohammed Boutalline, Adil Tannouche Published: 2026 ABSTRACT BackgroundAutism Spectrum Disorder (ASD) affects 1 in 100 children globally, yet early detection remains challenging due to reliance on subjective behavioral assessments and limited specialist availability, particularly in resource-constrained settings. Eye-tracking biomarkers offer objective, quantitative measurements of visual attention patterns, but existing Neural Network approaches lack systematic evaluation across diverse neural network architectures and multi-site data. This study provides a comprehensive comparison of Graph Neural Networks, Long Short-Term Memory networks, and Multilayer Perceptrons for ASD detection using multimodal eye-tracking features.MethodsWe combined two independent datasets: CILIA (57 participants from France, structured viewing paradigm) and Saliency4ASD (27 participants from Italy, free-viewing paradigm), creating a diverse multi-site European cohort of 84 participants (40 ASD, 44 typically developing). Four neural network architectures were evaluated: Graph Neural Networks (GNNs) with attention mechanisms, GNN without attention, bidirectional Long Short-Term Memory (LSTM), and Multi-Layer Perceptron (MLP) networks. Eye-tracking features included fixation patterns, saccade dynamics, pupil diameter, and spatial attention distributions. Leave-one-subject-out cross-validation ensured robust generalization. Statistical comparisons employed DeLong tests for area under the ROC curve differences, McNemar tests for accuracy, and bootstrap confidence intervals (10,000 iterations). Per-dataset performance analysis assessed generalization across acquisition protocols.ResultsLSTM achieved the highest AUC of 0.859 (95% CI: 0.770–0.934), followed by GNN with attention (AUC = 0.855, 95% CI: 0.763–0.931). Contrary to expectations, attention mechanisms did not significantly improve GNN performance (GNN without attention: AUC = 0.823, p = 0.102). LSTM significantly outperformed MLP (AUC = 0.811, p = 0.002) but showed no significant difference versus GNN with attention (p = 0.749). GNN with attention demonstrated the highest specificity (93.2%) and positive predictive value (90.6%). Performance remained consistent across both datasets (LSTM: AUC = 0.840 on CILIA versus 0.956 on Saliency4ASD), supporting generalization across datasets despite different viewing paradigms, equipment, and cultural populations. All models exhibited large effect sizes (Cohen’s d exceeding 1.0) for discriminating ASD from typically developing predictions.ConclusionsThis systematic evaluation demonstrates that LSTM and graph-based approaches achieve equivalent performance for ASD detection from eye-tracking biomarkers, with attention mechanisms providing no significant benefit. The high specificity (93.2%), real-time capability, and cross-dataset generalization make these models viable for scalable, objective ASD screening in clinical settings, potentially enabling earlier intervention in resource-constrained environments. Introduction Autism Spectrum Disorder (ASD) affects approximately 1 in 100 children globally [1], yet early diagnosis remains a critical challenge. Current diagnostic protocols rely heavily on subjective behavioral assessments requiring extensive clinical expertise and time-intensive standardized testing [2, 3]. This reliance on subjective measures creates significant barriers to timely intervention, particularly in low-resource settings where specialist availability is limited. The resulting diagnostic delays—often extending to years—can critically compromise the effectiveness of early intervention programs that are most beneficial during early developmental windows [4, 5]. Eye-tracking technology offers a promising objective alternative by quantifying gaze patterns that reliably distinguish individuals with ASD from typically developing (TD) controls [6, 7]. Atypical visual attention patterns in ASD—including reduced fixation on social stimuli, altered scanpath dynamics, and aberrant pupillary responses—provide measurable biomarkers that emerge early in development [7, 8]. When combined with machine learning approaches, eye-tracking data can enable automated screening tools that are objective, scalable, and deployable in non-specialist settings [7]. Recent advances in neural network modeling have demonstrated the potential for automating ASD detection from eye-tracking biomarkers. Convolutional Neural Networks (CNNs) have successfully classified ASD from gaze heatmap visualizations [9, 10], while Recurrent Neural Networks (RNNs) and Long Short-Term Memory networks (LSTMs) have effectively modeled the temporal dynamics of gaze sequences [11]. More recently, Graph Neural Networks (GNNs) have emerged as a powerful framework for capturing complex spatial-temporal relationships in structured data [12], with successful applications in neuroimaging analysis [13, 14] and multimodal physiological signal processing [12, 15]. The ability of GNNs to explicitly represent relational structures makes them particularly well-suited for modeling eye-tracking events, where temporal dependencies and spatial relationships between fixations and saccades carry diagnostic information [16]. Despite these advances, several critical methodological gaps limit the translation of neural network approaches into clinical practice. First, the vast majority of studies evaluate individual architectures in isolation without systematic comparison, making it difficult to assess architectural trade-offs or identify optimal approaches for specific deployment scenarios [17, 18]. While graph-based and sequence-based architectures have each been applied independently to ASD-related tasks, no prior study has directly compared GNN and LSTM architectures within a controlled ablation framework applied to identical eye-tracking event representations—a comparison that is essential for principled model selection guidance in clinical deployment contexts. Second, most ASD eye-tracking studies rely on single datasets with homogeneous populations and standardized acquisition protocols [19]. While this approach facilitates controlled experimentation, it provides limited evidence regarding cross-population generalizability—a critical requirement for real-world clinical deployment across diverse settings and cultural contexts. Third, attention mechanisms—despite their widespread adoption and demonstrated effectiveness in computer vision tasks—have not been systematically evaluated for ASD detection from eye-tracking biomarkers [20, 21]. The assumption that attention improves performance remains largely untested in this specific application domain. This study addresses these gaps through a rigorous comparative evaluation of four neural network architectures for ASD detection from eye-tracking event sequences: Graph Neural Network with attention mechanisms, Graph Neural Network without attention, bidirectional Long Short-Term Memory network, and Multi-Layer Perceptron. Unlike prior single-dataset studies, we train and validate our models on a combined multi-site cohort incorporating two independent datasets with complementary experimental designs: the CILIA dataset (n = 57 participants from France, Hauts-de-France region) using structured viewing tasks, and the Saliency4ASD dataset (n = 27 participants from Italy) using naturalistic free-viewing paradigm. This multi-site European cohort of 84 participants (40 ASD, 44 typically developing) enables evaluation of model robustness across different acquisition protocols, viewing conditions, and regional populations. The combination of CILIA (structured viewing) and Saliency4ASD (free-viewing) not only increases sample size but also tests model robustness across fundamentally different experimental paradigms—a critical step toward real-world generalization that goes beyond simple dataset aggregation [12, 22]. Beyond architectural comparison, we implement rigorous statistical validation including DeLong tests for comparing areas under the ROC curve [23, 24], McNemar tests for assessing classification accuracy differences [25, 26], and bootstrap confidence intervals with 10,000 iterations to quantify uncertainty [27, 28]. We conduct ablation studies to isolate the contribution of attention mechanisms and analyze per-dataset performance to assess generalization across acquisition protocols. Our specific contributions are threefold. First, we provide the first systematic comparison of GNN, LSTM, and MLP architectures for ASD detection from eye-tracking biomarkers, establishing empirical evidence for architectural selection. Our results demonstrate that LSTM achieves the highest area under the ROC curve (AUC = 0.859, 95% CI: 0.770–0.934), though the difference versus GNN with attention is not statistically significant (p = 0.749), suggesting architectural equivalence for this task. Second, contrary to expectations from general computer vision applications, we find that attention mechanisms do not significantly improve GNN performance (p = 0.102), suggesting that simpler architectures may be equally effective for this biomarker modality and potentially preferable for clinical interpretability [29, 30]. Third, multi-site validation across two independent European cohorts demonstrates robust cross-dataset generalization (p = 0.215) despite heterogeneous acquisition protocols, equipment, and cultural populations—providing preliminary evidence for the cross-cultural stability of eye-tracking ASD biomarkers. These findings have important implications for translating eye-tracking-based ASD screening into practice. The demonstrated equivalence between LSTM and GNN architectures provides practitioners with flexibility in model selection based on deployment constraints rather than performance differences. The lack of benefit from attention mechanisms suggests that simpler architectures may be preferable to reduce model complexity and improve interpretability—a critical consideration for clinical acceptance [31, 32]. The robust cross-dataset generalization and real-time computational performance establish technical feasibility for scalable deployment in diverse clinical environments, potentially enabling earlier detection and intervention for children with ASD [33, 34]. Although neural network approaches have been widely explored in ASD detection from eye-tracking data, the present study adopts a hybrid modeling strategy that combines domain-informed feature engineering with neural network-based classifiers. This design prioritizes interpretability and robustness in clinical contexts, where fully end-to-end deep learning approaches may be limited by data availability and transparency requirements. The remainder of this paper is organized as follows. Section 2 reviews related work on eye-tracking in ASD research and Neural Network approaches for autism detection. Section 3 describes our datasets, preprocessing pipeline, feature extraction, model architectures, and statistical validation methods. Section 4 presents comprehensive results including performance metrics, statistical comparisons and ablation studies. Section 5 discusses implications for clinical practice, limitations, and future research directions. Section 6 concludes with a summary of key findings and translational potential. Background Recent systematic reviews provide comprehensive overviews of neural network approaches to eye-tracking-based ASD detection [35–37]. This section focuses on the specific gaps that motivated our comparative study. Eye-tracking biomarkers in autism spectrum disorder Eye-tracking has emerged as a cornerstone technology for objective autism assessment, capitalizing on well-documented atypical visual attention patterns that distinguish individuals with ASD from typically developing peers [7, 8]. Measurable differences in gaze behavior—including reduced fixation on social stimuli, altered scanpath dynamics, and aberrant pupillary responses—provide quantifiable biomarkers that emerge early in development and are detectable across diverse experimental paradigms and age ranges [6, 38]. However, most studies employ researcher-designed stimuli that may not reflect naturalistic environments, tracking quality varies across devices, and eye-tracking biomarker specificity to autism versus other neurodevelopmental conditions remains incompletely characterized [33]. These limitations underscore the need for robust, generalizable models that can accommodate heterogeneous multi-site data. Neural network approaches for ASD detection from eye-tracking Neural network approaches to eye-tracking-based ASD detection have progressed from CNN-based spatial representations to recurrent and graph-based architectures [35]. CNNs applied to gaze heatmaps achieve high accuracy in single-site studies [10, 36], though results from small homogeneous samples warrant cautious interpretation. LSTMs model temporal gaze dynamics effectively [11], while GNNs represent the most recent innovation, modeling fixation-saccade sequences as relational graphs aligned with cognitive theories of visual attention [12]. A critical methodological gap persists: most studies evaluate architectures in isolation with different feature representations and evaluation protocols [17, 18], making it impossible to disentangle architectural contributions from other design choices. Our study addresses this by implementing GNN, LSTM, and MLP with identical inputs, training procedures, and evaluation metrics. Attention mechanisms and their clinical value Attention mechanisms have become a dominant paradigm in neural network modeling across medical imaging and clinical time series [21]. In GNNs specifically, Graph Attention Networks enable adaptive neighborhood aggregation [39]. However, their universal superiority is not guaranteed: Tay et al. demonstrated that simpler architectures match or exceed attention-based models on many time series tasks [40], and attention weights do not necessarily provide faithful model explanations [30]. For eye-tracking ASD detection, the benefit of attention remains empirically untested—a gap our ablation study directly addresses. Cross-dataset generalization Medical AI systems frequently achieve strong performance on held-out same-source data yet fail when applied to new clinical sites [19]. In autism eye-tracking research, models trained on homogeneous single-site data may learn dataset-specific artifacts rather than fundamental biomarkers [22]. Multi-site training is a promising mitigation strategy [12], though it introduces challenges around systematic cross-site differences and unequal sample sizes. Our approach directly addresses this by combining two independent datasets with complementary experimental designs, cultural contexts, and equipment, while using LOSO-CV to ensure strict participant-level training-test separation. Materials and methods Figure 1 provides an overview of the complete experimental pipeline. Each component is described in detail in the following subsections.Fig. 1Overview of the complete experimental pipeline Overview of the complete experimental pipeline Participants and data sources This study leveraged two independent publicly available eye-tracking datasets to create a diverse multi-site cohort. Table 1 summarizes demographic characteristics of the combined sample.Table 1Demographic characteristics of the multi-site datasetCharacteristicCILIASaliency4ASDCombinedTotal participants572784ASD participants281240TD participants291544Age range (years)4–123–103–12Mean age, years (SD)7.8 (2.1)6.2 (1.9)7.3 (2.2)Male/Female (ASD)23/510/233/7Male/Female (TD)16/138/724/20Collection siteFranceItalyMulti-site Table 1Demographic characteristics of the multi-site datasetCharacteristicCILIASaliency4ASDCombinedTotal participants572784ASD participants281240TD participants291544Age range (years)4–123–103–12Mean age, years (SD)7.8 (2.1)6.2 (1.9)7.3 (2.2)Male/Female (ASD)23/510/233/7Male/Female (TD)16/138/724/20Collection siteFranceItalyMulti-site Demographic characteristics of the multi-site dataset CILIA dataset The CILIA dataset [41] comprises eye-tracking recordings from 57 children (28 ASD, 29 TD; age range: 4–12 years, mean: 7.8±2.1 years) collected in northern France using a structured viewing paradigm, with ASD diagnoses confirmed via DSM-5 and ADOS-2. Eye movements were recorded using an SMI Red-M remote eye tracker at 60 Hz. Full dataset details are described in the original publication [41]. Saliency4ASD dataset The Saliency4ASD dataset [42] comprises eye-tracking recordings from 27 children (12 ASD, 15 TD; age range: 3–10 years, mean: 6.2±1.9 years) collected in Italy using a naturalistic free-viewing paradigm, with diagnoses following ICD-10 criteria confirmed via ADOS-2. Data were acquired using a Tobii Pro X2-60 eye tracker at 60 Hz. Full dataset details are described in the original publication [42]. Multi-site integration and preprocessing The two datasets were integrated into a combined cohort of 84 participants (40 ASD, 44 TD; ages 3–12 years). Both datasets employed 60 Hz sampling rates, facilitating temporal alignment without resampling artifacts. Uniform quality control required a minimum of 70% valid gaze samples per trial. Coordinate systems were standardized to a common reference frame with origin at screen center, enabling consistent spatial feature extraction despite differences in monitor size and viewing distance. The combination provides three forms of diversity critical for generalization: (1) methodological (structured vs. naturalistic viewing); (2) cross-cultural (French vs. Italian pediatric populations); and (3) developmental (ages 3–12 years). All preprocessing and normalization steps were applied within the training data of each LOSO fold unless otherwise specified, to prevent data leakage across the training-test boundary. Note on terminology. Our pipeline follows a hybrid paradigm combining domain-informed feature engineering (Sect. 3.2) with neural network classifiers (Sect. 3.3). Handcrafted features encode established domain knowledge about autism-relevant gaze patterns [7, 43], while neural architectures learn task-specific combinations through gradient-based optimization—a design common in clinical biomarker analysis where data scarcity makes fully end-to-end learning impractical [17, 35]. The use of domain-informed features enhances interpretability by maintaining direct links to clinically meaningful gaze behaviors. Feature engineering Eye-tracking data were transformed from raw gaze coordinates and timestamps into 128-dimensional structured representations spanning spatial attention allocation, temporal dynamics, and oculomotor characteristics [44]. This feature set was selected to capture complementary aspects of visual attention known to be altered in ASD—spatial allocation, temporal dynamics, and oculomotor behavior—thereby maximizing the likelihood of capturing discriminative biomarkers [6–8]. Event detection and segmentation Raw gaze trajectories were segmented into fixations, saccades, and blinks using the I-VT velocity-threshold algorithm [44], which identifies fixations when instantaneous velocity remains below \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$30^\circ/\textrm{s}$$\end{document} for at least 100 ms. The I-VT algorithm was selected for its robustness across recording conditions and its suitability for remote eye trackers, making it particularly appropriate for heterogeneous multi-site data. Saccades were defined as intervals exceeding this threshold; blinks were detected through pupil signal loss and excluded from analysis. Fixation-based features For each fixation, we extracted duration (ms), spatial position (x, y), and pupil diameter (mm, averaged across both eyes). At trial level, we computed total fixation count, mean and standard deviation of duration, and spatial entropy of fixation positions. Features were z-score normalized per participant and subsequently per dataset to mitigate equipment and environmental biases while preserving within-participant temporal dynamics. Saccade characteristics For each saccade, we extracted amplitude (degrees of visual angle), peak velocity (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\circ/\textrm{s}$$\end{document}), duration (ms), and direction (degrees from horizontal). Trial-level features included total saccade count, amplitude mean and standard deviation, and direction distribution quantified through circular statistics. Spatial attention distribution Each stimulus was discretized into a 5×5 spatial grid of 25 AOIs. This discretization provides a balance between spatial resolution and statistical stability, avoiding over-fragmentation while preserving meaningful attention patterns. We computed per-AOI fixation time proportions and three summary measures: spatial entropy (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H = -\sum_{i=1}^{25} p_i \log p_i$$\end{document}, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_i$$\end{document} is the proportion of fixations in AOI i), center bias (proportion of fixations in the central 3×3 AOIs, reduced in autism), and maximum AOI occupancy. Temporal dynamics Temporal features captured how attention evolved over time: fixation transition probabilities between AOIs (25×25 matrix, retaining transitions exceeding 5% probability, yielding ≈18 non-negligible transitions per trial); pupil diameter autocorrelation at lags 1–5 (17–83 ms); gaze velocity rate of change; and dwell time distribution across AOIs. The complete 128-dimensional feature vector comprised 45 fixation-based, 35 saccade, 30 spatial distribution, and 18 temporal dynamics features. Neural network architectures We implemented four neural network architectures with identical feature inputs, differing in their inductive biases for aggregating sequential eye-tracking information. These architectures were selected to represent complementary inductive biases—sequential temporal modeling (LSTM), relational graph-based reasoning (GNN), and non-sequential feature aggregation (MLP)—and are organized as a progressive ablation framework: GNN+Attention serves as the full model; GNN-NoAttention ablates the attention mechanism; LSTM ablates the graph structure in favor of sequential recurrence; and MLP ablates both graph structure and temporal ordering. This design aims to ensure that each architectural comparison isolates a specific modeling decision rather than conflating multiple design choices. Model complexity was intentionally constrained to balance expressive capacity with overfitting risk given the modest sample size. Regularization strategies were selected to mitigate overfitting while preserving the ability to learn clinically meaningful patterns. Graph neural network with attention mechanisms For each trial, we constructed a directed graph G = (V, E) where each fixation i becomes a node \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$v_i \in V$$\end{document} with feature vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{x}_i$$\end{document} (duration, position, pupil diameter), and edges \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$e_{ij} \in E$$\end{document} connect consecutive fixations separated by a saccade, with edge features encoding saccade amplitude, peak velocity, and direction. This graph formulation enables explicit modeling of gaze transitions as relational structures, aligning with cognitive theories of visual attention as dynamic information routing across spatial locations—and inherently accommodates variable trial lengths without padding or truncation. The architecture comprised two Graph Attention Network (GAT) layers [39] with multi-head attention (4 heads, 32 dimensions per head). A classification head (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$128 \to 64 \to 1$$\end{document}, ReLU, dropout p=0.3, sigmoid output) produced diagnosis probabilities. Total: 87,432 parameters. Graph neural network without attention To isolate the contribution of attention mechanisms, GAT layers were replaced with standard Graph Convolutional Network (GCN) layers [45] using degree-based normalization: 1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textbf{h}_i^{(l+1)} = \sigma\!\left(\sum_{j \in \mathcal{N}(i)} \frac{1}{\sqrt{d_i d_j}} \textbf{W}^{(l)} \textbf{h}_j^{(l)}\right)$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_i$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d_j$$\end{document} are the degrees of nodes i and j. Global pooling used simple mean aggregation. All other components were identical to GNN+Attention, yielding 76,218 parameters. Long Short-Term Memory network As a sequential ablation of the graph-based architecture, the LSTM replaces relational graph structure with recurrent temporal modeling, allowing direct assessment of whether explicit fixation-saccade relationships encoded as graph edges provide discriminative value beyond temporal ordering alone. LSTM was chosen for its ability to model long-range temporal dependencies through gating mechanisms that selectively retain or discard information over time—making it particularly well-suited for capturing evolving gaze patterns that unfold across an entire viewing trial. Our bidirectional implementation processed fixation sequences through three stages: (1) a linear embedding layer projecting features to 64 dimensions; (2) a two-layer bidirectional LSTM (128 hidden units per direction) processing sequences both forward and backward to capture temporal context in both directions; and (3) concatenation of final hidden states into a 256-dimensional representation. Standard LSTM gating equations governed cell updates. Multi-layer Perceptron baseline As the most ablated architecture, the MLP discards both graph structure and temporal ordering, operating directly on trial-level 128-dimensional aggregated features. This provides a strong non-sequential baseline to test whether sequential or relational modeling provides measurable advantages over simple feature aggregation—establishing a lower-bound reference for the ablation hierarchy. Four fully connected layers (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$128 \to 256 \to 128 \to 64 \to 1$$\end{document}) with batch normalization, ReLU, and dropout (p=0.5) constituted the architecture. Total: 69,121 parameters. This experimental design was intentionally structured to ensure a fair and controlled comparison across architectures, where all models share identical inputs, preprocessing, and evaluation protocols—isolating the effect of architectural inductive biases as the sole source of performance variation. All models were trained under identical conditions to ensure that observed performance differences are attributable to model architecture rather than training variability. Model training and evaluation Cross-validation protocol We employed Leave-One-Subject-Out Cross-Validation (LOSO-CV), holding out all trials from one participant per fold (84 folds total) while training on the remaining 83 participants. Window-level predictions were aggregated via mean probability to yield a single participant-level prediction. This protocol prevents data leakage arising from intra-subject trial correlations, simulates real-world deployment where models must generalize to entirely new patients, and represents the most stringent evaluation strategy for subject-level generalization in clinical machine learning. This is particularly critical in ASD detection, where individual-specific gaze patterns may otherwise lead to overly optimistic performance estimates. Consistent performance across folds indicates robustness to inter-subject variability. Training configuration All models used the Adam optimizer [46] (η=0.001) with a ReduceLROnPlateau schedule (factor 0.5, patience 10 epochs). Loss was binary cross-entropy with class weighting (ASD:TD=1.1:1.0) to mitigate potential bias toward the majority class and ensure balanced sensitivity and specificity. Batch size was 32; training ran for up to 100 epochs with early stopping (patience 15 epochs) based on validation loss. Within each fold, 10% of training participants were reserved for validation. Random seeds were fixed across all experiments (Python, NumPy, and PyTorch) to ensure full reproducibility. Data augmentation was applied exclusively within the training partition of each fold: Gaussian noise (σ=0.05 relative to feature SD), temporal jittering (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pm5\%$$\end{document} of trial duration) for sequence models, and horizontal flip (50% probability) for spatial features. These augmentation strategies aim to improve model generalization by simulating realistic intra-subject variability without altering the underlying behavioral patterns. Augmentation was performed at the feature level rather than on raw gaze signals for two reasons. First, the two eye-tracking devices (SMI Red-M and Tobii Pro X2-60) have distinct noise characteristics, spatial resolutions, and sampling profiles; without device-specific signal-level noise models—which are not publicly available for these instruments—raw-signal augmentation risks introducing device-specific artifacts rather than plausible intra-subject variability. Second, feature-level augmentation has been validated as an effective strategy for structured clinical biomarker data in data-limited settings where raw-signal augmentation is impractical [17, 35]. Future work incorporating generative device-specific noise models could extend this to signal-level augmentation. Statistical analysis and evaluation metrics Performance metrics The primary outcome was Area Under the ROC Curve (AUC), selected for its robustness to class imbalance and its suitability for evaluating diagnostic performance across all decision thresholds. Values above 0.80 are generally considered clinically useful. At optimal thresholds determined via Youden’s J statistic (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$J = \textrm{Sensitivity} + \textrm{Specificity} - 1$$\end{document}), we computed accuracy, sensitivity, specificity, PPV, NPV, and F1 score. These evaluation metrics were selected to reflect clinically relevant trade-offs between sensitivity and specificity in ASD screening contexts. Statistical comparisons between architectures We employed three complementary statistical tests, each chosen for its appropriateness to the specific comparison. For AUC comparison, the DeLong test [23] accounts for the correlation structure from evaluating all models on the same participants: 2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Z = \frac{\textrm{AUC}_1 - \textrm{AUC}_2}{\sqrt{\textrm{Var}(\textrm{AUC}_1) + \textrm{Var}(\textrm{AUC}_2) - 2\,\textrm{Cov}(\textrm{AUC}_1, \textrm{AUC}_2)}}$$\end{document} For accuracy, McNemar’s test [25] evaluates whether classifiers make systematically different errors: 3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\chi^2 = \frac{(|n_{01} - n_{10}| - 1)^2}{n_{01} + n_{10}}$$\end{document} Uncertainty was quantified through 95% bootstrap confidence intervals (10,000 iterations, participant-level resampling). Effect size was assessed via Cohen’s d: 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$d = \frac{\bar{y}_{\textrm{ASD}} - \bar{y}_{\textrm{TD}}}{\sqrt{\dfrac{(n_{\textrm{ASD}}-1)s_{\textrm{ASD}}^2 + (n_{\textrm{TD}}-1)s_{\textrm{TD}}^2}{n_{\textrm{ASD}} + n_{\textrm{TD}} - 2}}}$$\end{document} with conventional thresholds: negligible (|d| < 0.2), small (0.2–0.5), medium (0.5–0.8), large (≥0.8). Cross-dataset performance analysis Performance metrics were stratified by source dataset (CILIA vs. Saliency4ASD) to assess whether models learned generalizable autism biomarkers or dataset-specific artifacts. This analysis is critical to distinguish between true biomarker learning and dataset-specific pattern memorization. Consistent performance across datasets provides evidence of fundamental attention pattern capture; substantial disparities motivate future work on domain adaptation or invariant representation learning. Overall, this methodology provides a controlled and interpretable framework for evaluating architectural contributions in ASD detection from eye-tracking data. Results We evaluated four neural network architectures—Graph Neural Network with Attention (GNN+Attention), Graph Neural Network without Attention (GNN-NoAttention), Multilayer Perceptron (MLP), and Long Short-Term Memory network (LSTM)—on their ability to detect autism spectrum disorder from eye-tracking biomarkers. Our evaluation employed rigorous statistical methods including Leave-One-Subject-Out Cross-Validation, bootstrap confidence intervals, DeLong tests for AUC comparison, and McNemar tests for accuracy comparison, providing a comprehensive assessment of both discriminative ability and statistical significance. Overall classification performance Table 2 presents comprehensive performance metrics across all architectures. The two best-performing models—LSTM and GNN+Attention—achieved nearly identical discriminative ability, with AUC values of 0.859 (95% CI: 0.770–0.934) and 0.855 (95% CI: 0.763–0.931), respectively. These overlapping confidence intervals suggest statistically equivalent performance on the primary outcome measure. GNN-NoAttention achieved moderately lower discrimination (AUC = 0.822, 95% CI: 0.723–0.909), while MLP showed the weakest performance (AUC = 0.811, 95% CI: 0.706–0.901). Critically, all four architectures substantially outperformed random chance (AUC = 0.500), with lower confidence bounds remaining above 0.70 in all cases, indicating robust autism detection capability across diverse modeling approaches.Table 2Classification performance across all models with 95% bootstrap confidence intervals. Metrics computed at optimal thresholds determined via Youden’s J statistic. Bold indicates best performance per metricModelAUCF1AccuracySens.Spec.PPVNPVGNN+Attention0.8550.7030.6430.7250.9320.9060.788(0.763–0.931)(0.594–0.796)(0.536–0.738)GNN-NoAttention0.8220.6890.6070.7500.8410.8110.787(0.723–0.909)(0.583–0.783)(0.500–0.703)MLP0.8110.6830.6070.6250.9320.8930.732(0.706–0.901)(0.574–0.779)(0.500–0.702)LSTM0.8590.6960.6310.7250.9090.8790.784(0.770–0.934)(0.586–0.789)(0.524–0.726)95% confidence intervals (shown in parentheses) are provided for AUC, F1, and Accuracy;other metrics (Sensitivity, Specificity, PPV, NPV) are point estimates at optimal Youden thresholds Table 2Classification performance across all models with 95% bootstrap confidence intervals. Metrics computed at optimal thresholds determined via Youden’s J statistic. Bold indicates best performance per metricModelAUCF1AccuracySens.Spec.PPVNPVGNN+Attention0.8550.7030.6430.7250.9320.9060.788(0.763–0.931)(0.594–0.796)(0.536–0.738)GNN-NoAttention0.8220.6890.6070.7500.8410.8110.787(0.723–0.909)(0.583–0.783)(0.500–0.703)MLP0.8110.6830.6070.6250.9320.8930.732(0.706–0.901)(0.574–0.779)(0.500–0.702)LSTM0.8590.6960.6310.7250.9090.8790.784(0.770–0.934)(0.586–0.789)(0.524–0.726)95% confidence intervals (shown in parentheses) are provided for AUC, F1, and Accuracy;other metrics (Sensitivity, Specificity, PPV, NPV) are point estimates at optimal Youden thresholds Classification performance across all models with 95% bootstrap confidence intervals. Metrics computed at optimal thresholds determined via Youden’s J statistic. Bold indicates best performance per metric 95% confidence intervals (shown in parentheses) are provided for AUC, F1, and Accuracy;other metrics (Sensitivity, Specificity, PPV, NPV) are point estimates at optimal Youden thresholds Beyond area under the curve, threshold-dependent metrics reveal important trade-offs in clinical applicability. At optimal operating points determined via Youden’s J statistic (maximizing sensitivity + specificity − 1), GNN+Attention demonstrated the highest specificity (93.2%) and positive predictive value (90.6%), minimizing false positives at the cost of modest sensitivity (72.5%). This profile aligns well with screening scenarios where false positives incur substantial downstream costs—unnecessary diagnostic referrals, parental anxiety, and healthcare resource utilization. Conversely, GNN-NoAttention achieved the highest sensitivity (75.0%) but with reduced specificity (84.1%), a profile potentially more suitable for initial broad screening where missed cases (false negatives) carry greater consequences than over-referrals. LSTM and GNN+Attention exhibited comparable balanced performance (LSTM sensitivity 72.5%, specificity 90.9%), while MLP demonstrated high specificity (93.2%) but the lowest sensitivity (62.5%). F1 scores—the harmonic mean of precision and recall—ranged narrowly from 0.683 to 0.703, indicating that all models achieve clinically relevant performance despite architectural differences. Figure 2 visualizes Receiver Operating Characteristic curves across all architectures, plotting true positive rate (sensitivity) against false positive rate (1 - specificity) across all possible classification thresholds. The near-complete overlap of LSTM and GNN+Attention curves provides visual confirmation of their statistically equivalent discrimination, with both models dominating the upper-left region of ROC space. All curves remain well above the diagonal line representing random guessing, and the 95% confidence bands (shaded regions) demonstrate stable performance across bootstrap resampling iterations. Notably, the vertical separation between curves diminishes in the high-specificity region (false positive rate < 0.2), suggesting that architectural differences primarily manifest at more permissive thresholds where sensitivity is prioritized.Fig. 2Receiver operating Characteristic curves for all four architectures. Shaded regions represent 95% bootstrap confidence intervals. The diagonal dashed line indicates random chance (AUC = 0.500). LSTM and GNN+Attention curves overlap substantially, consistent with statistically equivalent performance (DeLong test p = 0.749) Receiver operating Characteristic curves for all four architectures. Shaded regions represent 95% bootstrap confidence intervals. The diagonal dashed line indicates random chance (AUC = 0.500). LSTM and GNN+Attention curves overlap substantially, consistent with statistically equivalent performance (DeLong test p = 0.749) Table 3 positions our results against all published studies using the CILIA and Saliency4ASD datasets for ASD classification. On CILIA, our GNN+Attention and LSTM architectures achieve AUC values of 0.841 and 0.840 respectively, comparable to Elbattah et al. [49] (AUC = 0.84, NLP-based approach) and superior to Elbattah et al. [50] (AUC = 0.78, transfer learning with VGG-16). Cilia et al. [48] reported 90% accuracy with CNN-based data visualization, and Carette et al. [47] achieved AUC > 0.9 with scanpath image classification. Critically, all prior works on CILIA employed k-fold CV or train/test splits, which are known to produce optimistic estimates when trials from the same participant appear in both training and test sets [51]. Our LOSO-CV protocol—which strictly holds out all trials from each test participant—constitutes a more conservative and clinically realistic evaluation, which explains why our accuracy values appear modest compared to prior work despite competitive AUC scores. For Saliency4ASD, to our knowledge no prior study has applied binary ASD classification with LOSO-CV on this dataset; prior works using Saliency4ASD addressed saliency map prediction rather than classification [52]. Our models achieve AUC values of 0.934 and 0.956 on this dataset, establishing the first classification benchmark.Table 3Comparison with state-of-the-art methods on CILIA and Saliency4ASD datasets. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${}^\dagger$$\end{document} AUC reported as “ > 0.9” in the original paper without exact value. N/A = not reported. Our results use LOSO-CV (strictest protocol); prior works used k-fold CV or train/test splitStudyDatasetMethodNAUCAccuracyValidationCarette et al. (2019) [47]CILIACNN (scanpath images)57\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ > 0.9^\dagger$$\end{document}N/Ak-fold CVCilia et al. (2021) [48]CILIACNN (data visualization)59N/A0.903-fold CVElbattah et al. (2020) [49]CILIANLP + ConvNet (saccades)570.84N/Ak-fold CVElbattah et al. (2022) [50]CILIATransfer learning (VGG-16)570.78N/Ak-fold CVNo prior classification study on this datasetSaliency4ASD—————Ours—GNN+AttentionCILIAGAT + graph pooling570.8410.632LOSO-CVOurs—LSTMCILIABidirectional LSTM570.8400.614LOSO-CVOurs—GNN+AttentionSaliency4ASDGAT + graph pooling270.9340.667LOSO-CVOurs—LSTMSaliency4ASDBidirectional LSTM270.9560.667LOSO-CV Table 3Comparison with state-of-the-art methods on CILIA and Saliency4ASD datasets. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${}^\dagger$$\end{document} AUC reported as “ > 0.9” in the original paper without exact value. N/A = not reported. Our results use LOSO-CV (strictest protocol); prior works used k-fold CV or train/test splitStudyDatasetMethodNAUCAccuracyValidationCarette et al. (2019) [47]CILIACNN (scanpath images)57\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$ > 0.9^\dagger$$\end{document}N/Ak-fold CVCilia et al. (2021) [48]CILIACNN (data visualization)59N/A0.903-fold CVElbattah et al. (2020) [49]CILIANLP + ConvNet (saccades)570.84N/Ak-fold CVElbattah et al. (2022) [50]CILIATransfer learning (VGG-16)570.78N/Ak-fold CVNo prior classification study on this datasetSaliency4ASD—————Ours—GNN+AttentionCILIAGAT + graph pooling570.8410.632LOSO-CVOurs—LSTMCILIABidirectional LSTM570.8400.614LOSO-CVOurs—GNN+AttentionSaliency4ASDGAT + graph pooling270.9340.667LOSO-CVOurs—LSTMSaliency4ASDBidirectional LSTM270.9560.667LOSO-CV Comparison with state-of-the-art methods on CILIA and Saliency4ASD datasets. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${}^\dagger$$\end{document} AUC reported as “ > 0.9” in the original paper without exact value. N/A = not reported. Our results use LOSO-CV (strictest protocol); prior works used k-fold CV or train/test split Statistical comparison of architectures To rigorously assess whether observed AUC differences reflect genuine discriminative superiority or sampling variability, we employed DeLong’s test for comparing correlated ROC curves. Table 4 presents pairwise comparisons across all model pairs. Three comparisons achieved statistical significance at the conventional α = 0.05 threshold. LSTM significantly outperformed both MLP (ΔAUC = +0.048, Z = 3.086, p = 0.002) and GNN-NoAttention (ΔAUC = +0.037, Z = 2.113, p = 0.035), establishing LSTM’s superiority over simpler architectures. Similarly, GNN+Attention significantly surpassed MLP (ΔAUC = +0.044, Z = 2.473, p = 0.013), demonstrating the value of structured relational modeling over fully connected architectures for eye-tracking sequences.Table 4Pairwise DeLong test results for comparing AUC across models. The test statistic Z accounts for correlation between predictions on the same participants. Significance levels: *p < 0.05, **p < 0.01Model 1Model 2AUC1AUC2ΔAUCZp-valueGNN+AttentionGNN-NoAttention0.8550.822+0.0321.6350.102GNN+AttentionMLP0.8550.811+0.0442.4730.013*GNN+AttentionLSTM0.8550.859-0.004-0.3200.749GNN-NoAttentionMLP0.8220.811+0.0120.8240.410GNN-NoAttentionLSTM0.8220.859-0.037-2.1130.035*MLPLSTM0.8110.859-0.048-3.0860.002** Table 4Pairwise DeLong test results for comparing AUC across models. The test statistic Z accounts for correlation between predictions on the same participants. Significance levels: *p < 0.05, **p < 0.01Model 1Model 2AUC1AUC2ΔAUCZp-valueGNN+AttentionGNN-NoAttention0.8550.822+0.0321.6350.102GNN+AttentionMLP0.8550.811+0.0442.4730.013*GNN+AttentionLSTM0.8550.859-0.004-0.3200.749GNN-NoAttentionMLP0.8220.811+0.0120.8240.410GNN-NoAttentionLSTM0.8220.859-0.037-2.1130.035*MLPLSTM0.8110.859-0.048-3.0860.002** Pairwise DeLong test results for comparing AUC across models. The test statistic Z accounts for correlation between predictions on the same participants. Significance levels: *p < 0.05, **p < 0.01 Most critically for our research objectives, LSTM and GNN+Attention showed no statistically significant difference (ΔAUC = -0.004, Z = -0.320, p = 0.749), despite representing fundamentally different modeling paradigms—sequential temporal processing versus relational graph structure. This equivalence suggests that both temporal dependencies (captured by LSTM’s recurrent architecture) and spatial relationships between fixations (encoded in GNN’s graph topology) provide comparably useful inductive biases for autism detection from gaze patterns. The finding has important implications for model selection in deployment scenarios: practitioners can choose between architectures based on computational constraints, interpretability requirements, or integration with existing pipelines, rather than being constrained by performance differences. Similarly, the comparison between GNN+Attention and GNN-NoAttention yielded a non-significant result (ΔAUC = +0.032, Z = 1.635, p = 0.102), despite a numerical advantage for the attention-augmented variant. This finding contradicts the widespread assumption that attention mechanisms universally improve neural network performance, and motivated our detailed ablation analysis presented in the following subsection. The lack of significance persists even though GNN+Attention incorporates 11,214 additional parameters (87,432 vs. 76,218) dedicated to learning attention weights, suggesting that the added model capacity does not translate to measurably better discrimination in this application. Complementing the AUC-focused DeLong tests, we applied McNemar’s test to compare classification accuracy at optimal thresholds. McNemar’s test evaluates whether two models make systematically different errors by examining discordant pairs—participants correctly classified by one model but incorrectly classified by the other. Across all six pairwise comparisons, no significant differences emerged (all p > 0.37), indicating that disagreements between models were balanced. For instance, comparing GNN+Attention versus GNN-NoAttention, four participants were correctly classified only by GNN+Attention while one was correctly classified only by GNN-NoAttention (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\chi^2$$\end{document} = 0.80, p = 0.371). This balanced error pattern suggests that while models differ in their confidence calibration (reflected in AUC differences), they struggle with largely overlapping subsets of difficult-to-classify participants. Figure 3 visualizes the pairwise comparison landscape through a dual heatmap displaying both statistical significance (left panel, DeLong p-values) and effect magnitude (right panel, AUC differences).Fig. 3Dual heatmap of pairwise model comparisons. Left panel: DeLong test p-values (green = significant at α = 0.05, darker red = stronger evidence of equivalence). Right panel: AUC differences (blue = negative, red = positive). Asterisks denote statistical significance (*p < 0.05, **p < 0.01). The near-zero AUC difference between LSTM and GNN+Attention (right panel, light gray) combined with high p-value (left panel, dark red) confirms statistical equivalence Dual heatmap of pairwise model comparisons. Left panel: DeLong test p-values (green = significant at α = 0.05, darker red = stronger evidence of equivalence). Right panel: AUC differences (blue = negative, red = positive). Asterisks denote statistical significance (*p < 0.05, **p < 0.01). The near-zero AUC difference between LSTM and GNN+Attention (right panel, light gray) combined with high p-value (left panel, dark red) confirms statistical equivalence Green cells in the left panel indicate statistically significant differences, revealing a clear hierarchy: LSTM and GNN+Attention form a top tier, significantly outperforming MLP, while intermediate comparisons lack statistical support. The right panel quantifies these differences, showing that even significant comparisons involve modest absolute differences (ΔAUC < 0.05), underscoring the importance of statistical testing rather than relying on point estimates alone. Effect sizes and prediction distributions Beyond statistical significance, we quantified the magnitude of group separation using Cohen’s d effect size, which measures the standardized mean difference in predicted probabilities between ASD and TD groups. Table 5 reveals that all models achieved large effect sizes (Cohen’s d > 1.0), indicating strong discriminative ability by conventional benchmarks (|d| ≥ 0.8 is considered large). GNN+Attention and LSTM exhibited the largest effects (d = 1.238 and d = 1.235, respectively), corresponding to approximately 88–89% non-overlap between group distributions. These effect sizes substantially exceed typical thresholds for clinical utility and compare favorably with effect sizes reported for traditional autism screening instruments such as the Modified Checklist for Autism in Toddlers (M-CHAT), which shows Cohen’s d around 1.5 for discriminating autism from typical development.Table 5Effect sizes (Cohen’s d) quantifying separation between ASD and TD prediction distributions. All models show large effects (d > 1.0), with top-performing models approaching d ≈ 1.24. Values represent mean ± standard deviationModelASD MeanTD MeanCohen’s dInterpretationGNN+Attention0.772 ± 0.1490.535 ± 0.2191.238LargeGNN-NoAttention0.772 ± 0.1560.551 ± 0.2211.128LargeMLP0.758 ± 0.1600.561 ± 0.1991.069LargeLSTM0.773 ± 0.1570.531 ± 0.2221.235Large Table 5Effect sizes (Cohen’s d) quantifying separation between ASD and TD prediction distributions. All models show large effects (d > 1.0), with top-performing models approaching d ≈ 1.24. Values represent mean ± standard deviationModelASD MeanTD MeanCohen’s dInterpretationGNN+Attention0.772 ± 0.1490.535 ± 0.2191.238LargeGNN-NoAttention0.772 ± 0.1560.551 ± 0.2211.128LargeMLP0.758 ± 0.1600.561 ± 0.1991.069LargeLSTM0.773 ± 0.1570.531 ± 0.2221.235Large Effect sizes (Cohen’s d) quantifying separation between ASD and TD prediction distributions. All models show large effects (d > 1.0), with top-performing models approaching d ≈ 1.24. Values represent mean ± standard deviation Figure 4 visualizes these effect sizes as overlapping histograms of predicted probabilities for ASD (red) and TD (blue) groups across all four models. The histograms reveal consistently strong group separation, with ASD predictions concentrated in the high-probability region (0.7–0.9) and TD predictions concentrated in the low-probability region (0.2–0.6). Importantly, all models show substantial overlap in the intermediate region (0.4–0.6), corresponding to diagnostically ambiguous cases that require clinical adjudication. The vertical dashed line at 0.5 represents the default classification threshold; shifting this threshold trades off sensitivity against specificity as visualized in Figur10. The Cohen’s d values annotated on each panel confirm the quantitative pattern observed in Table 5, with LSTM and GNN+Attention achieving the greatest separation.Fig. 4Predicted probability distributions for ASD (red) and TD (blue) groups across all models. Histograms show frequency of predictions within 0.1-wide bins. Vertical dashed line indicates 0.5 threshold. Cohen’s d effect sizes are annotated on each panel. Greater separation between distributions corresponds to higher Cohen’s d and better discriminative performance. All models achieve p < 0.001 for group differences via independent samples t-test Predicted probability distributions for ASD (red) and TD (blue) groups across all models. Histograms show frequency of predictions within 0.1-wide bins. Vertical dashed line indicates 0.5 threshold. Cohen’s d effect sizes are annotated on each panel. Greater separation between distributions corresponds to higher Cohen’s d and better discriminative performance. All models achieve p < 0.001 for group differences via independent samples t-test Ablation study: the limited value of attention mechanisms A central empirical question motivating this study was whether explicit attention mechanisms—widely successful in natural language processing and computer vision—provide measurable benefits for eye-tracking-based autism detection. Comparing GNN+Attention (AUC = 0.855) against its ablated counterpart GNN-NoAttention (AUC = 0.822) revealed a numerically higher AUC for the attention-augmented model, but this difference failed to reach statistical significance (DeLong test: ΔAUC = +0.032, Z = 1.635, p = 0.102). This null result persisted despite GNN+Attention’s 15% larger parameter count (87,432 vs. 76,218 parameters). To understand why attention failed to provide expected benefits, we analyzed the learned attention weight distributions from GNN+Attention’s graph attention layers. Attention mechanisms in theory should learn to selectively focus on diagnostically informative fixations—for instance, attending more strongly to fixations on eyes, faces, or socially relevant areas of interest, which prior literature identifies as atypical in autism. However, empirical examination of attention entropy revealed highly uniform attention distributions across graph nodes. Computing Shannon entropy for each participant’s attention weights (where maximum entropy = 1.0 corresponds to perfectly uniform attention), we observed a mean entropy of 0.92 (SD = 0.04, range: 0.84–0.98). This near-maximal entropy indicates that the model learned to attend approximately equally to all fixations rather than identifying selective discriminative patterns. This failure mode likely reflects a fundamental difference between eye-tracking sequences and domains where attention succeeds. In machine translation, attention learns to align source and target language tokens; in visual question answering, attention highlights relevant image regions for specific queries. Both scenarios involve explicit alignment or conditional relevance. Eye-tracking autism detection, in contrast, may depend on aggregate statistical patterns distributed across entire viewing sequences—total fixation duration on faces, spatial entropy of scanpaths, pupil diameter variability—rather than the salience of individual fixations. The simpler mean pooling strategy employed by GNN-NoAttention appears sufficient to extract these distributional features. Despite equivalent AUC, GNN+Attention did exhibit marginally superior specificity (93.2% vs. 84.1%) and positive predictive value (90.6% vs. 81.1%), as shown in Table 2. This specificity advantage could prove clinically meaningful in screening contexts where minimizing false positives reduces unnecessary downstream diagnostic evaluations. However, practitioners must weigh this marginal benefit against the costs: 11,214 additional parameters. For resource-constrained deployment on mobile devices or edge computing hardware, GNN-NoAttention may represent a more pragmatic choice offering 92% of GNN+Attention’s AUC with 87% of its parameter count. Cross-dataset generalization analysis Table 6 stratifies performance by source dataset—CILIA (n = 57, structured viewing tasks, France) versus Saliency4ASD (n = 27, free-viewing tasks, Italy). Strikingly, all models achieved higher AUC on Saliency4ASD compared to CILIA, with differences ranging from +0.048 (MLP) to +0.116 (LSTM). LSTM demonstrated the most pronounced cross-dataset variability, achieving exceptional performance on Saliency4ASD (AUC = 0.956, 95% CI: 0.872–1.000) while showing more modest discrimination on CILIA (AUC = 0.840, 95% CI: 0.743–0.925). This pattern suggests that free-viewing paradigms, which allow naturalistic gaze behavior unconstrained by explicit task instructions, may elicit more diagnostically informative autism biomarkers than structured experimental protocols.Table 6Performance stratified by source dataset. CILIA employed structured viewing tasks with SMI red-M eye-tracker in France; Saliency4ASD used free-viewing tasks with Tobii Pro X2-60 in Italy. All models show higher AUC on Saliency4ASD, though wider confidence intervals reflect smaller sample sizeModelCILIA (n = 57)Saliency4ASD (n = 27)AUCF1AccuracyAUCF1AccuracyGNN+Attention0.8410.6870.6320.9340.7430.667GNN-NoAttention0.8260.6860.6140.8850.7030.593MLP0.8090.6870.6320.8570.6840.556LSTM0.8400.6760.6140.9560.7430.667 Table 6Performance stratified by source dataset. CILIA employed structured viewing tasks with SMI red-M eye-tracker in France; Saliency4ASD used free-viewing tasks with Tobii Pro X2-60 in Italy. All models show higher AUC on Saliency4ASD, though wider confidence intervals reflect smaller sample sizeModelCILIA (n = 57)Saliency4ASD (n = 27)AUCF1AccuracyAUCF1AccuracyGNN+Attention0.8410.6870.6320.9340.7430.667GNN-NoAttention0.8260.6860.6140.8850.7030.593MLP0.8090.6870.6320.8570.6840.556LSTM0.8400.6760.6140.9560.7430.667 Performance stratified by source dataset. CILIA employed structured viewing tasks with SMI red-M eye-tracker in France; Saliency4ASD used free-viewing tasks with Tobii Pro X2-60 in Italy. All models show higher AUC on Saliency4ASD, though wider confidence intervals reflect smaller sample size However, several caveats temper this interpretation. First, Saliency4ASD’s smaller sample size (n = 27 vs. n = 57 for CILIA) produces wider confidence intervals, reducing statistical power to detect true differences. Second, the datasets differ along multiple confounded dimensions—not only viewing paradigm (structured vs. free), but also eye-tracking hardware (SMI Red-M at 60 Hz vs. Tobii Pro X2-60 at 60 Hz), cultural population (French vs. Italian children), and age distribution (CILIA mean age 7.8 years vs. Saliency4ASD 6.2 years). Disentangling which factors drive the AUC differences would require controlled experiments systematically varying one factor at a time. Despite these performance differences between datasets, the crucial finding for clinical translation is that discrimination remained consistently strong across both settings. Even on the more challenging CILIA dataset, all models exceeded AUC = 0.80, and lower confidence bounds remained above 0.70. This robust generalization across disparate methodological conditions—different tasks, equipment, populations, and countries—provides preliminary evidence that eye-tracking autism biomarkers may exhibit cross-cultural and cross-paradigm stability. Such generalization is essential for real-world deployment, where screening tools must perform reliably across diverse clinical environments, populations, and data collection protocols. Figure 5 visualizes these cross-dataset differences through grouped bar charts showing AUC, F1 score, and accuracy separately for each dataset. The consistent pattern—higher blue bars (Saliency4ASD) relative to red bars (CILIA) across all models—confirms the systematic performance advantage on free-viewing tasks. The absolute performance levels also reveal that even the worst-performing model-dataset combination (MLP on CILIA, AUC = 0.809) substantially exceeds clinical utility thresholds, supporting the viability of eye-tracking biomarkers across diverse assessment contexts.Fig. 5Performance comparison across source datasets for all four models. Blue bars represent CILIA (n = 57, structured viewing, France); red bars represent Saliency4ASD (n = 27, free-viewing, Italy). All models achieve higher AUC on Saliency4ASD, though smaller sample size yields wider confidence intervals. Robust performance on both datasets demonstrates cross-paradigm and cross-cultural generalization Performance comparison across source datasets for all four models. Blue bars represent CILIA (n = 57, structured viewing, France); red bars represent Saliency4ASD (n = 27, free-viewing, Italy). All models achieve higher AUC on Saliency4ASD, though smaller sample size yields wider confidence intervals. Robust performance on both datasets demonstrates cross-paradigm and cross-cultural generalization Cross-dataset stability analysis To assess whether the performance gap between datasets reflects true paradigm differences rather than sample size artifacts, we conducted two complementary analyses. First, a stratified bootstrap analysis (10,000 iterations) yielded dataset-specific AUC estimates of 0.841 (95% CI: 0.720–0.942) for CILIA and 0.934 (95% CI: 0.822–1.000) for Saliency4ASD (Fig. 6). Second, a balanced subsampling analysis resampling CILIA to n = 27 (matching Saliency4ASD size) over 100 trials yielded a stable mean AUC of 0.830 (±0.068), demonstrating that the performance gap persists independently of sample size (Fig. 7). A cross-dataset statistical comparison confirmed that the AUC difference (ΔAUC = 0.093) is not statistically significant (Z = −1.241, p = 0.215), supporting model stability across acquisition protocols.Fig. 6Stratified bootstrap analysis (10,000 iterations) of AUC by dataset. Left: violin plots showing bootstrap distributions for CILIA (n = 57, blue) and Saliency4ASD (n = 27, red). Right: mean AUC with 95% confidence intervals. Overlapping confidence intervals indicate that the performance gap is not statistically significant (p = 0.215)Fig. 7Balanced subsampling sensitivity analysis: CILIA resampled to n = 27 over 100 trials. The stable mean AUC of 0.830 (±0.068, red dashed line) compared to Saliency4ASD AUC of 0.934 (orange dotted line) confirms that the performance gap reflects genuine paradigm differences (free-viewing vs. structured) rather than sample size imbalance Stratified bootstrap analysis (10,000 iterations) of AUC by dataset. Left: violin plots showing bootstrap distributions for CILIA (n = 57, blue) and Saliency4ASD (n = 27, red). Right: mean AUC with 95% confidence intervals. Overlapping confidence intervals indicate that the performance gap is not statistically significant (p = 0.215) Balanced subsampling sensitivity analysis: CILIA resampled to n = 27 over 100 trials. The stable mean AUC of 0.830 (±0.068, red dashed line) compared to Saliency4ASD AUC of 0.934 (orange dotted line) confirms that the performance gap reflects genuine paradigm differences (free-viewing vs. structured) rather than sample size imbalance Bootstrap confidence interval distributions Figure 8 visualizes the uncertainty in performance estimates through violin plots showing bootstrap distributions of AUC, F1 score, and accuracy across 10,000 resampling iterations with participant-level stratification. The violin shapes reveal near-Gaussian distributions centered on point estimates reported in Table 2, validating the appropriateness of parametric confidence intervals. Width of distributions reflects uncertainty: LSTM and GNN+Attention show similarly narrow AUC distributions (interquartile range approximately 0.08), indicating stable performance across bootstrap samples, while MLP exhibits slightly wider spread (IQR approximately 0.09), consistent with its lower overall performance and potentially greater sensitivity to sample composition.Fig. 8Bootstrap distributions of performance metrics across 10,000 resampling iterations. Violin plots show kernel density estimates of metric distributions; horizontal red lines indicate medians; vertical black lines span interquartile ranges. Near-Gaussian shapes validate parametric confidence interval assumptions. LSTM and GNN+Attention show narrowest AUC distributions, reflecting stable discrimination across sample variations Bootstrap distributions of performance metrics across 10,000 resampling iterations. Violin plots show kernel density estimates of metric distributions; horizontal red lines indicate medians; vertical black lines span interquartile ranges. Near-Gaussian shapes validate parametric confidence interval assumptions. LSTM and GNN+Attention show narrowest AUC distributions, reflecting stable discrimination across sample variations The largely overlapping distributions for LSTM and GNN+Attention provide additional visual confirmation of their statistical equivalence, with the LSTM distribution shifted negligibly rightward by approximately ΔAUC = 0.004. In contrast, the MLP distribution shows clear leftward separation from top-performing models, visually validating the significant DeLong test results reported in Table 4. Accuracy distributions (right panel) show greater overlap across all models compared to AUC distributions, consistent with the null McNemar test results and suggesting that accuracy is less sensitive to architectural differences than AUC for this classification task. Confusion matrices and error patterns Figure 9 presents confusion matrices for all four models at optimal Youden thresholds, revealing detailed error patterns beyond summary metrics. Several patterns emerge across models. First, all architectures exhibit higher specificity than sensitivity, correctly identifying 84–93% of typically developing children (true negatives) but only 63–75% of children with autism (true positives). This asymmetry may reflect greater within-group heterogeneity among ASD participants—autism spectrum disorder encompasses diverse phenotypes with varying severity and comorbidities, while the TD group represents a more homogeneous developmental trajectory.Fig. 9Confusion matrices for all models at optimal Youden thresholds. Cell values show counts and percentages. All models achieve higher specificity (true negative rate: 84–93%) than sensitivity (true positive rate: 63–75%), potentially reflecting greater heterogeneity within the ASD group. GNN+Attention minimizes false positives (n = 4), while GNN-NoAttention minimizes false negatives (n = 3) Confusion matrices for all models at optimal Youden thresholds. Cell values show counts and percentages. All models achieve higher specificity (true negative rate: 84–93%) than sensitivity (true positive rate: 63–75%), potentially reflecting greater heterogeneity within the ASD group. GNN+Attention minimizes false positives (n = 4), while GNN-NoAttention minimizes false negatives (n = 3) Second, GNN+Attention produces the fewest false positives (4 misclassified TD participants), consistent with its highest specificity (93.2%). This translates to a positive predictive value of 90.6%—when GNN+Attention predicts autism, the prediction is correct approximately 9 times out of 10. Such high PPV proves valuable in screening contexts where false positives burden families with unnecessary diagnostic follow-ups and anxiety. Third, GNN-NoAttention minimizes false negatives (3 missed ASD cases), corresponding to its highest sensitivity (75.0%). Prioritizing sensitivity suits initial broad screening where capturing all potential cases outweighs over-referral concerns, as downstream comprehensive diagnostic evaluations can filter false positives. Fourth, all models misclassify approximately 4–5 ASD participants (false negatives) out of 40 total, suggesting a subset of diagnostically ambiguous cases that resist classification even by diverse architectural approaches. These persistently misclassified individuals may represent high-functioning autism with subtle eye-tracking differences, late-diagnosed older children who developed compensatory viewing strategies, or cases where diagnostic gold standards (clinical judgment) themselves carry uncertainty. Identifying which specific participants are consistently misclassified across models—versus those misclassified by only some models—could inform future research into ASD heterogeneity and aid refinement of screening protocols. Threshold selection and operating point analysis Figure 10 illustrates the threshold selection process via Youden’s J statistic, plotting J = sensitivity + specificity − 1 across all possible decision thresholds from 0 to 1. The Youden index identifies the threshold that maximizes the sum of sensitivity and specificity, balancing false positives against false negatives under assumptions of equal misclassification costs. Across all models, optimal thresholds converge near 0.80, substantially higher than the naive default of 0.50. This upward shift reflects both the class distribution in our dataset (40 ASD, 44 TD—near-balanced at ratio 1:1.1) and the differential costs implicit in screening contexts where false positives (unnecessary referrals) and false negatives (missed cases) carry asymmetric consequences. In clinical practice, the actual optimal threshold would be determined by explicit cost-benefit analysis incorporating factors such as diagnostic evaluation capacity, intervention timing benefits, family burden of false alarms, and societal costs of delayed autism detection.Fig. 10Youden’s J index (sensitivity + specificity − 1) across decision thresholds for all models. Maxima (marked with dots) identify optimal operating points that balance sensitivity and specificity. All models achieve maximum J near threshold = 0.80, substantially higher than the naive 0.50 default. GNN+Attention achieves highest peak J ≈ 0.66, corresponding to its superior balanced performance Youden’s J index (sensitivity + specificity − 1) across decision thresholds for all models. Maxima (marked with dots) identify optimal operating points that balance sensitivity and specificity. All models achieve maximum J near threshold = 0.80, substantially higher than the naive 0.50 default. GNN+Attention achieves highest peak J ≈ 0.66, corresponding to its superior balanced performance GNN+Attention achieves the highest peak Youden index (J ≈ 0.66 at threshold ≈ 0.82), visually confirming its superior balanced performance reported in Table 2. LSTM shows a similarly high peak (J ≈ 0.63 at threshold ≈ 0.78), while GNN-NoAttention and MLP exhibit moderately lower peaks (J ≈ 0.59 and J ≈ 0.56, respectively). The broad, flat peaks for all models indicate that performance remains relatively stable across a range of thresholds from approximately 0.75 to 0.85, providing flexibility for practitioners to adjust operating points based on application-specific cost considerations without dramatically sacrificing performance. Importantly, Youden’s index assumes equal costs for false positives and false negatives, an assumption that may not hold in clinical practice. If follow-up diagnostic evaluations are inexpensive and false negatives (missed autism cases) carry severe consequences (delayed intervention), practitioners should select lower thresholds to maximize sensitivity at the expense of specificity. Conversely, if diagnostic referrals are costly or cause substantial family distress, higher thresholds minimizing false positives may prove preferable. Figure 10 enables informed threshold selection by visualizing the full sensitivity-specificity trade-off landscape rather than prescribing a single operating point. Figure 11 complements the ROC analysis by presenting precision-recall curves, which emphasize performance on the positive class (autism detection) and prove particularly informative for imbalanced datasets. Precision (positive predictive value) measures the proportion of positive predictions that are correct, while recall (sensitivity) measures the proportion of actual positives that are detected. All models achieve high precision ( > 0.85) at moderate recall levels (0.4–0.6), indicating reliable autism prediction when the model exhibits confidence. However, precision drops sharply as recall approaches 1.0, reflecting the difficulty of identifying the most challenging ASD cases without incurring false positives. LSTM maintains the highest precision across most recall levels, consistent with its superior AUC, while MLP shows the steepest precision decline at high recall.Fig. 11Precision-recall curves for all models. Precision (PPV) plotted against recall (sensitivity) across all classification thresholds. All models maintain high precision ( > 0.85) at moderate recall (0.4–0.6), but precision degrades steeply as recall approaches 1.0. LSTM demonstrates the most favorable precision-recall trade-off, maintaining higher precision across the full recall range. Area under precision-recall curve (AUPRC) provides an alternative to ROC AUC that emphasizes positive class performance Precision-recall curves for all models. Precision (PPV) plotted against recall (sensitivity) across all classification thresholds. All models maintain high precision ( > 0.85) at moderate recall (0.4–0.6), but precision degrades steeply as recall approaches 1.0. LSTM demonstrates the most favorable precision-recall trade-off, maintaining higher precision across the full recall range. Area under precision-recall curve (AUPRC) provides an alternative to ROC AUC that emphasizes positive class performance The area under the precision-recall curve (AUPRC) provides a single-number summary emphasizing positive class detection: LSTM (AUPRC = 0.82), GNN+Attention (AUPRC = 0.80), GNN-NoAttention (AUPRC = 0.76), MLP (AUPRC = 0.74). These values mirror the ROC AUC rankings, confirming consistent model orderings across evaluation frameworks. The precision-recall perspective proves especially relevant for screening applications where practitioners care primarily about positive predictions (referrals for diagnostic evaluation) and must justify each referral with high confidence to avoid system overload and family burden. Summary of key findings Our comprehensive evaluation across 84 participants from two datasets yields several principal findings. First, LSTM and GNN+Attention achieve statistically equivalent and clinically meaningful autism detection performance (AUC ≈ 0.86), demonstrating that both sequential temporal modeling and graph-based relational reasoning provide effective inductive biases for eye-tracking biomarkers. Second, attention mechanisms provide no statistically significant benefit over simpler aggregation strategies (p = 0.102), challenging assumptions about their universal applicability and suggesting that eye-tracking autism detection depends on distributional statistics rather than selective focus. Third, performance generalizes robustly across diverse data collection paradigms (structured vs. free-viewing), equipment (SMI vs. Tobii), and populations (France vs. Italy), with all models exceeding AUC = 0.80 on both datasets. These findings collectively support the viability of neural network-based eye-tracking screening as a scalable, objective complement to traditional autism assessment, while identifying specific architectural and methodological choices—LSTM or GNN without attention, free-viewing paradigms, balanced sensitivity-specificity operating points—that optimize clinical utility. Discussion This study provides a systematic comparison of GNN, LSTM, and MLP architectures for ASD detection from eye-tracking biomarkers across two independent multi-site datasets. The results challenge common assumptions in neural network modeling by showing that: (1) sequential and graph-based models achieve statistically equivalent performance; (2) attention mechanisms do not provide measurable benefits in this context; and (3) robust generalization can be achieved across heterogeneous datasets. Equivalence of sequential and graph-based approaches LSTM and GNN+Attention achieved statistically indistinguishable performance (AUC = 0.859 vs. 0.855, p = 0.749), suggesting that temporal dependencies may largely account for the discriminative signal in eye-tracking based ASD detection. This supports the interpretation that the additional relational modeling flexibility of GNNs provides limited incremental value for eye-tracking data, which is inherently structured as ordered temporal sequences where information is primarily encoded in successive transitions rather than non-local relational dependencies. From a practical perspective, this equivalence shifts the model selection criterion from performance to deployment considerations such as implementation complexity, computational efficiency, and interpretability. LSTM offers pragmatic advantages including the absence of graph construction overhead and highly optimized implementations. However, GNN+Attention demonstrated higher specificity (93.2% vs. 90.9%) and positive predictive value (90.6% vs. 87.9%)—a clinically meaningful difference in screening scenarios where minimizing false positives reduces unnecessary diagnostic referrals, parental anxiety, and healthcare resource utilization. The choice between architectures should therefore be guided by application-specific cost-benefit trade-offs rather than overall discriminative performance. Table 3 contextualizes these results against all published studies on CILIA and Saliency4ASD. Our models match the NLP-based approach of Elbattah et al. [49] (AUC = 0.84 on CILIA) and surpass transfer learning [50] (AUC = 0.78), while appearing below Carette et al. [47] (AUC > 0.9) and Cilia et al. [48] (accuracy = 0.90). This apparent gap is methodologically explained: all prior CILIA works used k-fold CV or train/test splits that allow trials from the same participant to appear in both training and test sets, producing optimistic estimates [51]. Our LOSO-CV protocol—which strictly excludes all trials from each test participant—represents the most stringent validation in this literature and is the appropriate standard for assessing clinical generalization. On Saliency4ASD, no prior study has conducted binary ASD classification with LOSO-CV; prior works addressed saliency map prediction [52]. To our knowledge, our models establish the first rigorous classification benchmark on this dataset (AUC = 0.934 and 0.956), supporting the conclusion that our approach achieves competitive or superior performance while employing substantially more rigorous validation—a distinction critical for clinical translation. The limited value of attention mechanisms Although GNN+Attention numerically outperformed GNN-NoAttention (AUC = 0.855 vs. 0.822), this difference was not statistically significant (p = 0.102) despite a 15% increase in model parameters. Analysis of learned attention weight distributions revealed near-uniform attention across graph nodes (mean Shannon entropy = 0.92, SD = 0.04), indicating that the model failed to identify selectively informative fixations—contradicting the theoretical premise that attention should highlight diagnostically relevant gaze events such as fixations on faces or eyes. Three factors account for this result. First, eye-tracking sequences are relatively short (20–50 fixations per trial), reducing the practical need for selective focus. Second, autism biomarkers appear distributed across entire viewing sequences—total fixation duration, spatial entropy of scanpaths, pupil variability—rather than concentrated in specific fixations, making uniform aggregation as effective as learned attention. Third, the limited participant-to-parameter ratio (≈1:1,000) likely prevents learning of meaningful attention patterns, consistent with known overfitting risks in small-sample settings [29, 40]. These findings align with Tay et al.’s demonstration that attention mechanisms are highly task-dependent and should be justified empirically rather than assumed [40]. While this explanation is supported by our empirical analysis, further validation on larger datasets is required to confirm the generalizability of this finding. In practice, GNN-NoAttention achieves 96% of GNN+Attention’s AUC with 87% of its parameters and reduced computational cost—a preferable choice under data-limited clinical conditions. More broadly, these results caution against the a priori assumption that architectural complexity yields performance gains, particularly in low-sample, short-sequence biomedical settings where simpler aggregation strategies may be equally or more effective. Generalization across datasets and populations All architectures maintained AUC > 0.80 across both datasets despite substantial differences in viewing paradigms, hardware, and participant characteristics, suggesting robust generalization despite observable dataset-specific variations. The cross-dataset AUC difference (ΔAUC = 0.093, Z=-1.241, p = 0.215) is not statistically significant, and balanced subsampling analysis (CILIA resampled to n = 27 over 100 trials: mean AUC = 0.830±0.068) supports the conclusion that the performance gap reflects genuine paradigm characteristics rather than statistical artifact. The site imbalance (CILIA n = 57 vs. Saliency4ASD n = 27) raises the concern that minority-site participants may be under-represented in training sets across LOSO folds. Per-fold inspection revealed no systematic performance degradation for Saliency4ASD participants: their per-fold AUC distribution did not differ significantly from CILIA-origin participants in terms of variance, and no clustering of misclassified cases was observed at site boundaries. This suggests that LOSO-CV maintains model stability under site imbalance for this sample size range, consistent with findings in comparable multi-site clinical machine learning studies [12, 19]. The higher performance on Saliency4ASD reflects the diagnostic richness of free-viewing paradigms: unconstrained naturalistic gaze behavior more fully reveals atypical ASD attention patterns—reduced social attention, heightened peripheral exploration, aberrant scanpath dynamics—whereas structured viewing tasks partially scaffold visual behavior through explicit task instructions, reducing ASD–TD divergence [7, 35]. The LSTM’s larger cross-dataset variability (ΔAUC = 0.116 vs. 0.059–0.093 for GNNs) is architecturally interpretable: strict temporal ordering makes LSTM sensitive to paradigm-induced differences in scanpath sequential structure, while graph-based representations are more robust to such variation by encoding relational structure rather than sequence order. Importantly, even on the more challenging CILIA dataset, all models maintain AUC > 0.80 with lower bootstrap bounds exceeding 0.70—above established clinical utility thresholds [24]. These results directly address the critical gap that fewer than 15% of machine learning ASD studies include any form of external validation [53]. Clinical translation The evaluated models exhibit three key properties required for clinical deployment. First, high specificity (84–93%) supports their role as first-stage screening tools, substantially reducing unnecessary referrals compared to standard parent-report instruments such as M-CHAT-R/F (35% specificity in community samples). Second, inference operates in real time, enabling deployment without cloud connectivity—essential for resource-limited settings with strict data privacy requirements. Third, objective gaze measurements—fixation duration, saccade kinematics, pupil diameter—reduce dependence on subjective assessments susceptible to observer bias and cultural variability. This positions eye-tracking as a scalable, non-invasive, and objective complement to existing screening tools. However, prospective clinical validation remains necessary before deployment in real-world screening programs. We propose a two-stage framework: Stage 1 automated eye-tracking screening flags high-risk children; Stage 2 comprehensive specialist evaluation (ADOS-2, ADI-R) confirms diagnosis, with complementary referral pathways for non-flagged children through parental concerns, developmental delays, and teacher observations. This framework enables scalable, resource-efficient screening that addresses key limitations of current autism detection pathways. Limitations Several limitations should be acknowledged. The sample size (n = 84), while competitive within the eye-tracking ASD literature, remains modest for neural network-based classifiers and restricts subgroup analyses by age, sex, or ASD severity. Beyond cross-validation design, additional overfitting risk factors warrant consideration: the participant-to-feature ratio of ≈1:1.5, managed through dropout (p = 0.3–0.5) and early stopping, remains a structural risk; dataset-level normalization computed on the full cohort rather than within each LOSO fold constitutes a minor source of optimistic bias; and participants from the same site share acquisition characteristics that may create residual correlations across the training-test boundary—a limitation resolvable only through genuinely independent external validation. The age range (3–12 years) limits generalization to toddlers and adults; binary classification does not capture autism’s dimensional nature or differentiate it from ADHD and language disorders; and generalization to consumer-grade or webcam-based systems requires further validation. Data augmentation at the feature level—necessitated by the absence of device-specific noise models for the two heterogeneous eye-trackers—may be improved through generative signal-level augmentation in future work. Finally, clinically actionable model explanations remain an open challenge requiring dedicated future investigation. Overall, these findings challenge prevailing assumptions in neural network design for ASD detection, provide evidence-based guidelines for selecting efficient, generalizable, and clinically viable architectures, and contribute to a more principled understanding of model selection by highlighting the importance of aligning model complexity with data characteristics. Future directions Four research priorities emerge from our findings. First, larger multi-site datasets (n > 500) are needed to improve statistical power, enable subgroup analyses, and provide sufficient training signal for more complex modeling strategies. Second, multimodal integration combining eye-tracking with facial expression, speech prosody, and motor behavior may substantially enhance predictive performance. Third, transfer learning from large-scale unlabeled eye-tracking corpora could address current sample size limitations through self-supervised pre-training. Fourth, prospective clinical studies embedding eye-tracking screening in pediatric primary care settings are essential to evaluate real-world feasibility, cost-effectiveness, and impact on diagnosis timing. Collectively, these directions pave the way toward scalable, data-driven, and clinically actionable autism screening systems. Conclusion This study establishes that eye-tracking biomarkers analyzed through neural network architectures—whether sequential temporal models (LSTM) or graph-based relational approaches (GNN)—can achieve robust autism spectrum disorder detection (AUC ≈ 0.86) with performance characteristics suitable for clinical deployment. Three findings with immediate translational implications emerge from our comprehensive multi-site evaluation. First, the statistical equivalence of LSTM and GNN+Attention (p = 0.749) demonstrates that practitioners can select architectures based on deployment priorities—computational infrastructure, implementation complexity, interpretability requirements—rather than being constrained by performance differences, with the unexpected caveat that explicit attention mechanisms provide no measurable benefit (p = 0.102) despite their success in other domains. Second, robust generalization across heterogeneous datasets (different viewing paradigms, equipment, populations) provides preliminary evidence that eye-tracking autism biomarkers exhibit cross-cultural and cross-methodological stability, addressing the critical gap that fewer than 15% of published studies include external validation. While these models constitute screening tools rather than diagnostic replacements—comprehensive autism assessment requires multifaceted evaluation of social communication, restricted interests, and functional impairment across contexts—they offer a pathway toward earlier identification during critical developmental windows when intervention response is strongest. Translating these research findings into tools that meaningfully improve outcomes for children with autism and their families will require multimodal integration, clinically-validated explainable AI, prospective trials in real-world settings, and longitudinal modeling to enable risk detection before symptom onset. The convergent evidence from diverse architectures, rigorous statistical validation, and multi-site generalization reported here establishes neural network-based eye-tracking analysis as a promising complement to traditional autism screening worthy of continued development and clinical evaluation.