Loading saved reports…
Protocol
Fixed before modelling: 5 outer folds over participants (10 test / 40 train each),
0.5–3.5 s cue-locked window, CSP+LDA with inner grid on training participants only.
Full text:
docs/evaluation_protocol.md.
Frozen outer folds
Evaluation
Per-fold test accuracy (unseen participants)
CSP+LDA vs reduced EEGNet (EEGNet-R)
EEGNet-R is the pre-registered reduced variant (dropout 0.25, max 15 epochs). The original full EEGNet protocol was timed but not run. Full EEGNet was never executed.
Participants
Each row is one test participant in their outer fold. Accuracy below 50% is highlighted. The lower bound comes from a one-sided binomial test (greater than chance); the stored upper end is always 100% for that one-sided interval, not a two-sided CI. Holm adjustment is across all 50 participants.
Signal metadata
Channel labels (64 scalp EEG)
Limits
- Engineering results for one CSP+LDA pipeline on 50 participants of one public subset. Not clinical, assistive, or neuroscience conclusions.
- Group mean above chance does not mean every person decodes well. Median participant accuracy is about 58%; many individuals cannot be separated from guessing at this trial count.
- Test trials were scored without extra cleaning; training-only rejection removed 10–29 trials per fold.
- Only five outer folds — fold SD is a rough spread measure, not a precise uncertainty on the mean.
- Neural comparison is EEGNet-R only; do not generalise to EEGNet or deep models in general.