Hybrid Vision Transformer-Mamba Framework for Autism Diagnosis via Eye-Tracking Analysis

High diagnostic accuracy could improve access to early intervention: The model achieved 96% accuracy in detecting autism spectrum disorder from eye-tracking data, which could enable screening in remote or under-resourced clinical settings where specialist diagnosis is currently unavailable.

Explainable AI techniques enhance clinical trustworthiness: The framework incorporates explainable AI methods alongside its deep learning components, addressing a critical concern for healthcare applications—clinicians need to understand why the model makes diagnostic recommendations, not just accept its predictions.

Multimodal data integration captures complex behavioral patterns: By combining visual attention, speech, and facial cues through attention-based fusion, the model captures both spatial and temporal dynamics of behavior, moving beyond single-indicator screening methods that may miss important diagnostic signals.

High sensitivity (97%) is particularly important for screening: The model's sensitivity—its ability to correctly identify individuals with autism—is especially valuable for diagnostic screening, as missing cases could delay critical early interventions, though the 94% specificity suggests some false positives that would require clinical confirmation.

Transition from handcrafted to automated features reduces bias potential: Replacing manually-designed diagnostic features with learned representations from deep learning may reduce subjective bias in feature selection, though the model's performance should be validated across diverse populations to ensure equitable screening.