The evaluation
Inside one dataset, then across the gap
Internal cross-validation — ds007526
133 participants (106 PD / 27 HC), participant-grouped 5×5-fold CV. Brackets are 95% participant-level bootstrap intervals.
| model | balanced accuracy | ROC-AUC |
|---|---|---|
| majority | 0.500 [0.50, 0.50] | 0.500 [0.50, 0.50] |
| demographics | 0.645 [0.54, 0.75] | 0.688 [0.57, 0.79] |
| logreg | 0.563 [0.47, 0.67] | 0.718 [0.61, 0.82] |
| svm_linear | 0.558 [0.46, 0.66] | 0.662 [0.55, 0.77] |
| random_forest | 0.591 [0.50, 0.69] | 0.759 [0.66, 0.85] |
A demographics-only model (age + sex) rivals or beats every EEG model on balanced accuracy — much of the within-cohort separation is demographic, not neural.
Frozen external transfer — ds007526 → ds002778
Fit on 133 development participants, evaluated once on 30 external participants. The external cohort was never used for tuning. "gap" is internal minus external balanced accuracy.
| model | external balanced acc | internal | gap |
|---|---|---|---|
| majority | 0.500 [0.50, 0.50] | 0.500 | +0.000 |
| demographics | 0.500 [0.33, 0.68] | 0.645 | +0.145 |
| logreg | 0.667 [0.50, 0.83] | 0.563 | -0.104 |
| svm_linear | 0.633 [0.46, 0.80] | 0.558 | -0.075 |
| random_forest | 0.533 [0.40, 0.67] | 0.591 | +0.058 |
The demographic shortcut collapses to chance on the balanced external cohort; linear EEG models transfer more gracefully, but with only 30 external participants the intervals are wide — encouraging, not conclusive.