← All projects
NeuroscienceArchived2024

Feeling the Frequency

What a cheap EEG headset and a smartwatch can decode

Classifying what someone is doing from a single-electrode EEG headset and a smartwatch, with one model per person.

Raw EEG and motion signals for subject 01 during four activities
85.5%mean validation accuracy, chance is 25%
31subjects, one model each
94.2%best subject

What it does

The subjects watched a happy video, typed answers about it, watched a sad video, and typed answers about that. I trained a separate classifier per person to tell those four apart, with 70 Random Forest configurations swept per subject. Saved models scored about 85% to 94% on a held-out test split for 9 subjects.

Validation accuracy across 31 subjects.
Validation accuracy across 31 subjects.
Subject 01: most errors are between the two typing activities.
Subject 01: most errors are between the two typing activities.
Depth drives accuracy; tree count barely matters.
Depth drives accuracy; tree count barely matters.

What I'd fix before trusting it

  • Shuffled split. Train and test rows come from the same minutes of recording, which inflates accuracy.
  • Motion vs EEG. Six of the seven features are wrist motion, and typing vs watching is easy to read from a smartwatch. There's no EEG-only run, so I can't say how much the brain signal contributes.
  • Per-subject only. Nothing tests transfer to a new person.
  • The poster's number. The poster reports 89% and frames this as decoding emotion. I can't reproduce that figure from the saved outputs, and the framing goes further than the data does. The numbers on this page are the ones I can back up.