Seizures are rare in EEG recordings, which makes them hard to learn: fewer than one in ten clinical segments is an actual seizure, and labeling all that data is costly. DiffEEG sidesteps both problems. It's a compact model, around 9.6 million parameters, pretrained on 1.3 million unlabeled segments so it learns the general grammar of brain signals before seeing any labels, much like a language model reading raw text first.

The fine-tuning is the clever part. Instead of chasing overall accuracy, which just rewards ignoring rare seizures, it uses reinforcement learning to push up F1 directly. Under strict patient-by-patient testing across 279 patients, it reached 81% accuracy for yes-or-no detection and kept usable seizure recall despite the imbalance. Sorting events into four subtypes was harder, near 61%.

These figures come from the abstract, so read the paper for how it holds up in a real clinic.