Page 237 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

228 CHAPTER 11 Deep Learning Approaches to Electrophysiological

neuronal metabolic processes and leads to a loss of connections between nerve cells.
Three main characteristics can be commonly observed in AD patients’ EEG signals,
compared to healthy controls: the slowing effect (the power at low frequencies
increases whereas the power at high frequencies decreases), the reduction of
complexity and of synchrony between pairs of EEG signals [46,47]. Such effects
come with the functional disconnection caused by the death of neurons [16,17].
In Creutzfeldt-Jacob Disease (CJD), a progressive transmissible form of enceph-
alopathy, patients develop dementia and a peculiar spongiform degeneration of
cortical and subcortical gray matter [18]. EEG helps in diagnosing CJD, particularly
in the middle/late stages of the disease when periodic sharp wave complexes
(PSWC) become clearly visible [19]. PSWC may disappear at the later stages of
the disease when spongiform changes involve whole cerebral cortex [20,21].
Fig. 11.6 (top) shows the EEG traces of a patient affected by CJD, a patient
affected by AD and a Healthy Control (HC), recorded by the occipital channel
O 1 . The EEG was recorded in a comfortable eye-closed resting state. The signals
were band-pass filtered at 0.5e32 Hz and sampled at 256 Hz. The CJD EEG
(Fig. 11.6, top) exhibits the typical aforementioned PSWC sharp and wave
complexes. The AD EEG (Fig. 11.6, middle) shows the “slowing effect,” peculiar
of cerebral degeneration due to AD.
The AD slowing effect is evident in Fig. 11.6 (middle) and Fig. 11.7 (middle).
The dominant peak in alpha band (8e13 Hz), peculiar of healthy subjects (HC) in
eye-closed resting state, is indeed present in the Power Spectral Densities (PSD)
of the HC whereas it slowed and reduced in the AD patient, as well as in the CJD
patient. Delta band (0e4 Hz) looks prominent in both AD and CJD patients, as
rather expected from clinical considerations. In Dl approaches, this clinical behavior
can be automatically extracted and represented in suitable features.

4.3 HIGH-DENSITY ELECTROENCEPHALOGRAPHY
EEG signals have very good temporal resolution but poor spatial resolution, because
of both volume conduction effects and large interelectrode distance. Biophysical
analyses reported that some information of the scalp electrical potentials is lost
unless an intersensor distance of 1e2 cm is used [22e25]. With standard 10e20
system, the average intersensor distance is 7 cm. Achieving a 1e2 cm sampling
density would require 500 EEG channels distributed uniformly over the scalp.
High-Density-EEG (HD-EEG) offers the high temporal resolution, typical of
EEG, in conjunction with high spatial resolution (Fig. 11.8). HD-EEG with 256-
channels provides adequate approximate spatial sampling [23,24]. An example of
high-density EEG recording is shown in Fig. 11.9.
Fig. 11.10 depicts the 3D scalp topography of the power of the EEG recording
shown in Fig. 11.9. The signal power was estimated for every EEG epoch, and
then averaged over the time giving the average power value per channel. The
obtained values were then mapped over the scalp, where the coloration of intersensor
areas was estimated by interpolation [26].