Page 241 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

232 CHAPTER 11 Deep Learning Approaches to Electrophysiological

The potential of HD-EEG has been widely proven in the identification of the
epileptogenic onset zone through electrical source imaging [27e29] but it is mostly
unexplored in many other fields of application, like dementia. However, it requires a
high computational effort and generate huge amounts of data, particularly in
long-time monitoring. It is clear that DL approaches can be of great help to manage
this kind of data.

4.4 MAGNETOENCEPHALOGRAPHY

Magnetoencephalography (MEG) is a functional neuroimaging technique for map-
ping brain activity by recording magnetic fields produced by electrical currents
occurring naturally in the brain, using very sensitive magnetometers. Arrays of
superconducting quantum interference devices (SQUIDs) are currently the most
common magnetometer.
Although EEG and MEG signals originate from the same neurophysiological
processes, many important differences can be highlighted. Magnetic fields are less
distorted than electric fields by the different conductivity properties of the head
tissues (the skull is insulating whereas the scalp is conducting), resulting in a mild
sensitivity to volume conduction effects and in a greater spatial resolution. This has
relevant implications for connectivity analyses and source modeling. Furthermore,
MEG measurements are absolute as they are independent on the reference choice.
However, EEG is far more affordable, manageable, and cheap than MEG, which
caused its widespread availability, as compared to MEG technology, both in research
and clinical practice.

5. DEEP LEARNING MODELS FOR EEG SIGNAL PROCESSING
In recent years, DL architectures have been applied for the analysis of EEG
recordings in cognitive neuroscience and neurology. In this research area, DL
models have been developed to learn discriminating features from EEG signals
recorded from patients with neurological disorders.

5.1 STACKED AUTOENCODERS
DL methodologies are of growing interest to process complex signals like EEG or
MEG, both in disease diagnosis and Brain Computer Interface (BCI) systems. These
kinds of signals represent practical examples of noisy and nonstationary multivariate
time-series, being acquired simultaneously on multiple channels. Typically, EEG is
acquired during a long time for diagnosis purposes and it presents some artifactual
and noise activity that may reduce its reliability and visual interpretability. The DL
approach to EEG/MEG signal processing can be proposed in two basic ways (see
Fig. 11.11): (1) in series, in which a feature engineering step yields a high number
of features that are then combined and reduced by a DL network before using them
for classification; (2) in parallel, where the data-driven features and the engineered