My PhD problem in 10 lines

Sometimes things go wrong with the brain. Sometimes large sections of the brain fire together. These are seizures: epilepsy in adults, neonatal seizures in newborns.

When the brain does unusual things, we want to know.

In the immediate term, this allows us to medicate to stop brain damage. In the longer term, this allows us to plan interventions and adapt care for the patient.

To know that seizures are happening, we can use EEG. This measures voltages on the scalp; the sum total of the electrical activity of the brain below each electrode is picked up.

EEG is imperfect and crude. Mostly, it only picks up big changes in brain behaviour; and these need to be trawled through by highly trained neurophysiologists.

Because they’re highly trained, they’re expensive. We usually can’t get neurophysiologist to monitor patients 24/7.

Instead we train computers to do their job.

This is machine learning. We feed the computer examples of seizures and non-seizures, and the computer learns the difference.

Then we run it on live brain signals. When it sees signals that look like seizures it alerts the nurses.

But computers get things wrong. Especially, they get things wrong when other signals interfere with brain signals. Every time a patient moves, blinks, or tenses their jaw the machine gets confused.

The focus of my PhD then was to figure out ways to detect and remove these interferences, so we can monitor seizures more accurately.