Eye Gaze and Attention Patterns Power a New AI Screen for Autism in Indian Children

Researchers in India have built a two-part artificial intelligence system that screens for autism spectrum disorder by studying where children look and how long they stay engaged, reporting accuracy above 97 percent on a dataset of Indian children. The study, published in Complex & Intelligent Systems, positions eye-tracking-style image analysis and video-based attention measurement as an alternative to the questionnaire-based screenings that dominate early autism detection today. Traditional screening tools such as the Modified Checklist for Autism in Toddlers, or M-CHAT, depend heavily on parent reports and standardized questions.

The authors argue that these instruments carry inherent subjectivity, culture-specific biases, and parent-report biases that can delay or distort a diagnosis. Their proposed alternative replaces subjective judgment with measurable visual behavior: photographs capturing eye positioning toward regions of interest, and videos capturing attention during live activities. The primary dataset was small and carefully balanced, consisting of 14 autistic and 14 non-autistic Indian children.

From this cohort, the team constructed a multitask classification framework. The first task classifies children as autistic or typically developing based on their eye positioning patterns toward regions of interest, using a hybrid architecture that combines ConvNeXtBase, a modern convolutional neural network, with LightGBM, a gradient-boosted decision tree model known for its efficiency on tabular features. On the primary Indian dataset, the hybrid model achieved an accuracy of 97.65 percent, precision of 99 percent, recall of 94.29 percent, an F1-score of 97.06 percent, specificity of 99 percent, and a ROC-AUC of up to 98 percent.

Because a single train-test split can flatter a model by chance, the researchers validated reliability using Cohen's Kappa coefficient, a statistic that measures agreement beyond what random guessing would produce. Values of 0.95 for a 90-10 split, 0.92 for an 80-20 split, and 0.95 for a 70-30 split indicate almost perfect agreement across different data partitions. A paired t-test on the performance comparisons yielded t = 14.4 with p < 0.001, which the authors cite as evidence that the observed improvements are statistically significant rather than artifacts of a lucky split.

The second task tackles a different behavioral signal: attention. Using histogram of oriented gradients, a classic computer vision technique that describes the structure of edges in an image, paired with a linear support vector machine classifier, the system quantifies each child's attention rate and engagement ratio during live activities. The attention results were striking in their contrast.

Autistic children in the study exhibited 36.46 percent attention and 72.98 percent engagement during the recorded activities, while typically developing children demonstrated 72.53 percent attention and 96.26 percent engagement. In other words, typically developing children attended to the task roughly twice as much as their autistic peers, and their engagement was more than 23 percentage points higher. All of these performance metrics were validated with paired t-tests, according to the paper.

The two tasks complement each other in a clinically meaningful way. Combining a static image modality with a dynamic video modality means the system does not rely on a single snapshot of behavior, and the multitask structure allows each signal to be evaluated with the method best suited to it. The authors conclude that eye positioning and attention patterns are promising indicators for ASD screening, and that deep learning approaches offer automatic, efficient, and culturally unbiased alternatives to traditional methods.

By training and validating on local data, the team aims to mitigate the culture-specific biases embedded in imported screening instruments. Important caveats remain. The dataset comprised only 28 children, so the reported metrics describe performance on this specific cohort and will need replication on larger and more diverse samples before any clinical deployment.

The study establishes feasibility and statistical robustness within its scope rather than diagnostic validity in real-world clinics. Even so, the work points toward a future where computer vision-based eye gaze analysis could help identify autism traits at early stages and reduce overreliance on human judgment, complementing rather than replacing clinicians in the screening pipeline.

Subject of Research: Multimodal deep learning screening for autism spectrum disorder using eye-position images and attention videos from Indian children

Article Title: Eyes don’t lie: a multitask and multimodal classification for ASD screening on Indian data

Article References: Bidwe, R. V., Mishra, S., Bajaj, S., Vaidya, S., Revadkar, Y., & Kotecha, K. (2026). Eyes don’t lie: a multitask and multimodal classification for ASD screening on Indian data. Complex & Intelligent Systems. https://doi.org/10.1007/s40747-026-02544-5

Image Credits: AI Generated

DOI: 10.1007/s40747-026-02544-5

Keywords: autism spectrum disorder, ASD screening, deep learning, eye gaze, ConvNeXt, LightGBM, attention measurement, computer vision, HOG, support vector machine, Indian children, multimodal classification