Inspiration
ADHD affects over 360 million people worldwide, yet diagnosis often takes months of subjective behavioral assessments and expensive clinical evaluations. We asked: what if we could use the eyes—windows to cognitive function—to detect attention disorders in minutes?
Research shows that individuals with ADHD exhibit distinct eye movement patterns: more frequent saccades, reduced fixation stability, and altered pupil responses. We built EXCITE to harness this science.
What it does
EXCITE is a non-invasive ADHD screening tool that analyzes eye movements in real-time. Users simply watch dots on a screen while our system:
- Tracks eye movements using computer vision (pupil detection, gaze estimation)
- Extracts biomarkers like saccade velocity, fixation stability, and gaze entropy
- Runs AI inference through a Transformer neural network trained on 12,000+ eye-tracking recordings
- Delivers results with probability scores and detailed analytics
How we built it
- Computer Vision Pipeline: Custom OpenCV-based pupil detection with ellipse fitting and glint tracking
- ML Model: PyTorch Transformer encoder pre-trained on GazeBase dataset, fine-tuned on ADHD-specific data
- Web Interface: Flask backend with a modern, animated frontend featuring real-time visualization
- Edge Deployment: Optimized to run on Raspberry Pi (~$50 hardware)
Challenges we ran into
- Achieving accurate pupil detection under varying lighting conditions
- Balancing model complexity with inference speed for real-time analysis
- Creating a UI that's both beautiful and functional for clinical settings
- Training on limited ADHD-labeled data while maintaining generalization
Accomplishments we're proud of
- End-to-end pipeline from raw video to ADHD probability in under 30 seconds
- Beautiful, YC-caliber web interface with neural network visualization
- Privacy-first design—all processing happens locally on device
- Portable solution that can be deployed in schools, clinics, or homes
What we learned
- The fascinating connection between eye movements and cognitive function
- Transfer learning techniques for medical AI with limited labeled data
- How to build production-ready ML pipelines that work on edge devices
What's next for EXCITE
- Clinical validation studies to establish sensitivity/specificity benchmarks
- Mobile app version for broader accessibility
- Integration with existing telehealth platforms
- Expanding to other attention-related conditions
Log in or sign up for Devpost to join the conversation.