Skip to main content
Frozen linear probes, 5-fold subject-level cross-validation. Balanced accuracy (%). The first column is EPI-250k, our base foundation model — not publicly released. It is the upper bound on what the MRL distillation can preserve. The remaining columns are the MRL model at each truncation dimension, which is what pip install neuroencoder gives you.

Private clinical tasks

40,909 annotated 30-second epochs from the Swiss Epilepsy Center.
TaskEPI-250k7683841924816
Seizure / Wake93.493.192.792.591.584.1
Sleep (5-class)85.177.077.476.976.573.2
Artifact / Wake90.290.590.390.590.765.9
Seizure / Sleep88.885.284.984.082.179.4
Spike / Seizure81.576.275.974.771.065.5
Spike / Wake97.094.894.794.692.987.2
Artifact / Spike78.876.075.675.374.470.4
Category (6-cls)36.333.633.332.831.727.4
Clinical Sub (7-cls)42.731.431.431.427.023.7
All Sublabels (49-cls)22.114.814.413.712.310.6

Public benchmarks

10 standard public EEG datasets, evaluated under identical conditions.
TaskEPI-250k7683841924816
TUAB73.172.472.572.972.270.4
TUEV54.545.947.246.742.832.1
TUAR45.243.042.942.239.536.5
TUSL73.371.575.177.171.369.7
Mumtaz82.180.781.882.683.283.1
Schizo71.170.169.469.569.466.7
MentArith60.960.259.958.655.652.2
ADFTD43.240.040.041.038.635.9
PhysioMI30.328.328.427.327.725.2
Parkinsons62.958.958.658.255.953.2
The numeric column headers (768, 384, …) are the MRL truncation dimensions.

Dimension retention

Mean delta vs the EPI-250k base model, across all 20 tasks.
MRL dimMean delta
768-3.4 pp
384-3.3 pp
192-3.5 pp
48-5.3 pp
16-10.0 pp
Binary tasks retain accuracy best. Fine-grained multi-class tasks (TUEV, sublabels) and tasks with large domain shift from pre-training data (Parkinsons, MI) degrade more sharply.