When does an EEG foundation model help on people it has never seen? Motor imagery vs. seizure detection

I compared a pretrained EEG foundation model (LaBraM, 5.8M parameters) with a compact convolutional network trained from scratch (EEGNet, 2–3k parameters) in two subject-independent settings, changing only the model and selecting every threshold and stopping point on validation subjects. For left- vs. right-hand motor imagery (50 PhysioNet subjects), full fine-tuning of LaBraM reached 79.1% ± 1.7% across four subject splits versus 54.9% for EEGNet, but only when the input matched LaBraM's 0.1–75 Hz pretraining band; with the field-standard 8–30 Hz filter every model was at chance. For seizure detection (BIDS Siena, 14 patients, 47 seizures, SzCORE event scoring), the advantage became a trade-off: LaBraM raised about 3× fewer false alarms (9.4 vs. 29.6 per 24 h) but detected about half as many seizures (0.34 vs. 0.64). Averaging the two models kept the low false-alarm rate while recovering most of the sensitivity (event F1 0.37 vs. 0.27 and 0.24). Distilling LaBraM into EEGNet did not preserve its low false-alarm rate. Code, per-split results and recomputation scripts: https://github.com/YihengLi-1/eeg-motor-imagery-decoding