An AI model links gaze patterns—but its autism-screening claims need stronger tests

Turning eye movements into a learning problem

Where someone looks, how their gaze moves and how long it pauses can provide information about visual attention. Researchers are investigating whether such measurements could support autism assessment. But classifying recorded gaze patterns is not the same as establishing a diagnosis or showing that a screening tool works in everyday care.

In an arXiv preprint submitted on November 7, 2026, Quang-Anh N. D. and colleagues introduce ATHENA, short for Adaptive Threshold Hypergraph Enhanced Neural Architecture. Their system combines a convolutional neural network with a mechanism for connecting groups of learned features. The work reports computational experiments on existing eye-tracking datasets, not a prospective clinical screening trial. The supplied record does not establish peer-reviewed publication.

Learning connections between features

The network processes images representing eye movements rather than directly interpreting a child's behavior. Convolutional layers first extract patterns from those images. ATHENA then enriches the resulting internal feature maps using a hypergraph.

An ordinary graph connects pairs of nodes; a hypergraph can connect several nodes through one shared connection. Here, the nodes represent positions in learned feature maps, not children or anatomical structures. Grouping them lets the model combine information from multiple features with similar representations.

The proposed distinction is an adaptive distance threshold. Instead of always using the same cutoff to decide which features belong together, the architecture calculates a threshold from average and maximum feature responses. Its intended benefit is to adjust feature grouping to the input being processed.

A second component, Feature Map Mix, blends selected regions of internal representations from two training examples and mixes their labels accordingly. This is training-data augmentation: it creates additional combinations for learning, not additional independently observed participants.

What the datasets contain

The first dataset is described as 547 gaze-path images collected from 59 participants. Its original visualizations encode movement velocity, acceleration and jerk—the rate at which acceleration changes—using color channels.

The second, ASD-EyeMov, contains recordings from 14 autistic children and 14 typically developing children viewing 300 stimulus images across seven categories. Gaze was recorded at 120 samples per second during three-second presentations. The researchers construct trajectory images from fixation positions and durations, using transparency to encode normalized duration.

Preprocessing includes grayscale conversion and resizing to 100 by 100 pixels. Training also uses geometric image augmentation. Although the paper's title refers to toddlers, its dataset descriptions do not provide participant ages sufficient to establish toddler-specific performance.

High reported scores, with a metric discrepancy

For the first dataset, the authors report an AUC score of 0.9980 and accuracy of 98.45%. For ASD-EyeMov, they report an AUC score of 0.9986 and accuracy of 98.56%.

Conventionally, ROC AUC is the area under a curve plotting sensitivity—the true-positive rate—against the false-positive rate as the decision threshold changes. It measures how well a model ranks positive examples above negative ones. But Equation (20) in the supplied extract instead integrates positive predictive value, PPV, multiplied by the derivative of negative predictive value, NPV, with respect to the threshold. PPV and NPV describe the proportions of positive and negative predictions that are correct; they are not the sensitivity and false-positive rate used in the standard ROC integral.

This is an apparent methodological discrepancy. The supplied evidence does not establish whether it reflects a reporting or extraction error, or the calculation actually implemented. The reported scores therefore cannot be confirmed as standard ROC AUC and should not be treated as verified measures of near-perfect ranking performance. Nor does the reported 0.9986 mean that 99.86% of children would receive a correct diagnosis.

The reported sensitivity is approximately 98% on both datasets; specificity is 98.74% and 99.02%, respectively. These remain dataset evaluation results, not established population-screening rates.

A component comparison offers narrower evidence for ATHENA itself. Adding the module to a baseline CNN raises reported accuracy from 92.18% to 94.76%, a gain of 2.58 percentage points. That comparison does not isolate the contribution of every ingredient in the full system.

The missing test: unfamiliar children

The central uncertainty is how training and testing examples were separated. The paper does not specify the split procedure, whether participants were kept entirely separate, or how augmentation was confined to training data. When one person contributes multiple records, separating images without separating people can let a model exploit participant-specific patterns. The report does not establish whether that happened here.

The paper also does not provide confidence intervals or repeated-run variability for its headline scores. Dataset naming and comparison-table references are inconsistent, further complicating interpretation of claims that the system outperforms earlier methods.

The authors acknowledge the need for larger, more diverse and independent validation. The current contribution is an architecture with high reported benchmark scores whose interpretation remains limited. Establishing screening usefulness requires clarification of the AUC calculation, transparent participant-level evaluation and testing on genuinely new populations—not simply more images from a small group.