EEG has been shown to be an interesting tool for a number of intensive care applications. However its potential remains partially untapped because 'human interpretation' is not always feasible long continuous records. This talk will present some of the ongoing approaches that we are developing as part of my PhD work for automated analysis of intensive care EEG using machine learning.
First I will present a simple proof of concept on the use of deep networks on raw EEG samples for a different application: supervised sleep scoring using convolutional networks. Then I will delve into the more challenging problem of how to extract meaningful representations out of multivariate time series using unsupervised approaches. I will introduce why it is interesting to be able to represent signals without the aid of supervision information, and how such representational models can serve a clinical application system. Different models and associated architectures for doing this in practice will be showcased and compared: variational autoencoders, unsupervised approaches using joint clustering and feature learning, and generative adversarial models.
Wednesday 30 November 2016, 16:30 - 17:30 h
Building Carré - room CR 3.718