A translator gets distracted by every rustle in the hall, and a one-hour meeting stretches into two. That's how the standard method of locating an epileptic focus works: it waits for loud phrases — the moments when a seizure flares up in the brain. But epilepsy is more like a conversation where the important words aren't shouted; they hide in the pauses between remarks. The disease's source often lurks in the quiet of normal brain activity.
The EpiiSLM system takes a different approach. It was trained on 104,990 minutes of recordings from electrodes implanted in the brain (stereo-EEG), collected at the Montreal Neurological Institute. The first part of the system, a signal model, listens to the rhythms of quiet sleep and spots deviations from a healthy baseline. The second, a language model, pieces together these hints along with MRI data and medical history. Together, they give the doctor a clear picture: exactly where the seizure source is hiding.
The result is striking: just one night of interictal sleep is enough for the system to point to the focus with near-perfect accuracy at the level of a brain lobe. On external data from another clinic, accuracy remained high—0.857. Behind these numbers lies a real opportunity to cut the time a person spends with wires in their head, waiting for a seizure, and to boost the chance of being free from seizures for good.
During sleep, the brain isn't resting: it's processing memories. In people with epilepsy, these nighttime rhythms hide clues about the problem area, and now we've learned to read them.
In the TV series 'Perception,' a neuroscientist helps the FBI solve crimes by analyzing brain activity. EpiiSLM does something similar: it tracks down a hidden source of chaos by following barely visible traces.