A Multi-Modal Agentic Framework for Dyslexia Screening by Real-Time Gaze Pattern Recognition
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
The increase in number of consumer-grade imaging devices and advances in vision-language models (VLMs) present unprecedented opportunities for accessible healthcare diagnostics. This paper introduces a novel agentic framework leveraging vision-language models for automated dyslexia detection through eye-tracking pattern analysis. The system employs consumer-grade infrared cameras (GC0308 0.3MP, 120FPS) integrated with OpenCV-based pupil tracking to generate comprehensive attention maps during reading tasks. These visual representations are subsequently analyzed by specialized VLM agents that identify four critical dyslexic indicators: erratic eye movements, excessive regressions, word skipping, and abnormal attention clustering. The agentic architecture implements a multi-stage analysis pipeline where specialized agents evaluate different aspects of reading behavior, with a central orchestrating agent synthesizing individual assessments into comprehensive diagnostic reports. Our framework achieves 80.6% accuracy in dyslexia classification while maintaining real-time performance (15fps per frame) on consumer hardware. For enhanced accessibility, we implemented MediaPipe-based head pose estimation as an alternative to specialized cameras, achieving less accuracy but hardware independent using standard webcams. This work contributes to the democratization of specialized healthcare diagnostics through intelligent consumer electronics, reducing assessment costs from thousands to under twenty dollars demonstrating promising accuracy suitable for preliminary screening.
Fields of Study & Subjects
Gaze Tracking and Assistive Technology
Reading and Literacy Development
Handwritten Text Recognition Techniques
Publication details
- Journal / source
- International Conference on Consumer Electronics
- Pages
- 1–6
- Publication date
- 3 February 2026
- Language
- EN