2025 · Research lead — pipeline, model, evaluation
Autism Detection via Eye Tracking
A non-invasive screening pipeline that reads gaze behaviour instead of administering a questionnaire.
- Accuracy
- 95%
- Published
- IEEE OTCON-2025
- Pipeline
- ResNet18 → Random Forest
the problem
Autism screening depends on structured clinical observation, which is slow, unevenly available, and hard to scale. Gaze carries a well-documented behavioural signal — children on the spectrum distribute attention differently across faces and scenes — but reading that signal normally requires a specialist.
the approach
A two-stage pipeline. ResNet18 handles deep feature extraction from eye-tracking frames, learning representations of gaze patterns and fixation duration as behavioural biomarkers. A Random Forest then classifies on those features. Splitting the work this way keeps the classifier small and inspectable — the deep network does perception, and a model you can actually interrogate does the decision.
the result
95% classification accuracy across all test cases, validated on precision, recall, and F1 rather than accuracy alone, because a screening tool's false-negative cost is not symmetric with its false-positive cost. Published at IEEE OTCON-2025 as a reproducible, non-invasive screening approach.
stack
- Python
- TensorFlow
- ResNet18
- Random Forest
- scikit-learn