Handwriting decoding as a challenging motor task for EEG Foundation Models

A new EEG handwriting-decoding dataset was built to reduce confounds from earlier studies.

On a 4-letter classification task, knowing movement onset and improving test-time signal quality significantly improved accuracy.

Current EEG foundation models did not outperform smaller task-specific models on this harder benchmark.

The results suggest some foundation models may look strong on motor imagery tasks but still struggle with real-world decoding.