Reading an EEG is still how epilepsy gets diagnosed and seizures get identified, and deep learning has genuinely improved automated interpretation. Getting those systems from a research paper into a hospital is where progress stalls, and this review argues the obstacles are architectural rather than incidental.

Three problems recur. The models are opaque, and a neurologist asked to act on a seizure flag reasonably wants to know what in the trace produced it. They need large volumes of balanced annotated data, which is exactly what clinical practice does not generate, since seizures are rare by definition and labelling them takes expert time. And they are trained under conditions tidier than any real ward.

The framing as an evolution from handcrafted features toward functional edge learning is the useful thread. It reads the field's history as a sequence of answers to the same question, which is how much structure to build in versus how much to learn, and where each answer left clinical deployment.