Page 245 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

236 CHAPTER 11 Deep Learning Approaches to Electrophysiological

5.3 DEEP CONVOLUTIONAL NEURAL NETWORKS
A second approach to DL-EEG has been introduced in Ref. [32]. It is based on a
customized CNN for discriminating EEG recordings of subjects with AD, Mild
Cognitive Impairment (MCI), an early form of dementia, and HC. Fig. 11.13 shows
the flowchart of the method. The EEG recordings have been transformed in the
time-frequency domain, on an epoch-by-epoch basis. Then, a grand total of 228
time-frequency features have been extracted and used as input vector to the deep
network. The DL-model includes a convolutional layer followed by a sigmoidal
nonlinearity, a max pooling layer, an autoencoder, and a classification single hidden
layer MLP-NN. More than one convolutional-pooling stage can be used to generate
higher-level features. The combined DL-processor globally reduced the dimension-
ality of the input feature space from 228 to 10 latent features, providing very good
performance both in binary (83% accuracy) and 3-way classification (82% accuracy)
as reported in Table II of Ref. [32].

5.4 OTHER DL APPROACHES

Other researchers have faced the problem of DL-EEG pattern classification. In
particular, Zhao et al. [33] proposed a 3-stacked Restricted Boltzmann Machine
(RBM) structure for the classification of AD patients. They claimed reaching 92%
of accuracy.
Other papers focused on detection of epileptic seizures in EEG recordings. Wulsin
et al. [34] modeled a DBN model for classification and anomaly detection of epileptic
patterns. Mirowski et al. [35] developed a CNN for seizure prediction achieving
zero-false alarm on 95% patients analyzed; whereas, Turner et al. [36] proved the
effectiveness of DBN algorithm in seizure detection (F-measure up to 0.9).
Recently, DL networks have been also applied in the fast growing field of Brain
Computer Interface (BCI) as rapid serial visual presentation (RSVP) tasks [37],
steady-state visual evoked potential (SSVEP) classification [38], and P300 waves
detection [39].

6. FUTURE DIRECTIONS OF RESEARCH

The examples presented in this chapter have shown that DL can be a powerful tool to
assist the solution of difficult biomedical signal processing problems like discrimi-
nating brain states in order to differentiate brain pathological conditions or to
interpret tasks in BCI applications, like the classification of left and right hand in
motor imagery. The use ofEEG isthegoldstandardinboth cases.Itischeap and nonin-
vasive, and can be repeated easily being commonly well accepted by patients. Further-
more, the relevant data can be acquired through the increasingly popular wearable
devices that allow capturing continuously physiological and functional data in both
wellbeing and healthcare applications. This potentially rich information content can
be transmitted through Bluetooth and smartphone channels for remote monitoring.