I assume these models, and your end-product goal, is to make some kind of diagnosis, or flag anomalies? This is very interesting, and any more info on what your use-case or end-product would be great.
One of the main features of our software is epileptogenic spike detection. Epileptogenic spikes are little blips in the record and are indicative that the patient has epilepsy. Our models are now good enough at finding these spikes as human neurologists. There may be only a handful of these spikes in the record, so finding them automatically can be a huge time savings for the neurologist.
Our software also finds and eliminates various EEG artifacts. Blinking, for example, causes a huge artifact in the records. (It turns out that your eye is a dipole, and rolls upward every time you blink, which produces a huge signal.) The next major goal is seizure detection. It turns out there are a lot of things that can look a lot like a seizure in an EEG record. Even disconnecting the electrodes can look like a seizure.
The holy grail would be to give the program an EEG record and have it spit out the neurologist's report. We're not there yet, but we're working towards it. (By the way, we're hiring another ML engineer and software engineer. Email jobs@persyst.com if you're interested!)
I'm looking at using machine learning on time series data in a similar way, replacing EEG with 3D motion data. However, ML is not my forte and I am just cutting my teeth. I would really appreciate if you be able to share some of the 'buzzwords' or ML topics that I might look into as I explore this work?