Brain signal seminar

1. Semantic Classification of hearing words using ECoG
20185748, Ahmed Nafiz Ishtiaque
Bio-Medical Engineering
Introduction
Communication impairments can originate from neurodegenerative disorders that affect the motor
production and articulation of speech, such as amyotrophic lateral sclerosis, or from language
disorders that affect the cognitive production or comprehension of language, such as Broca’s aphasia,
Wernicke’s aphasia. One goal of characterizing neural activity during speech production and
comprehension is to develop neuro technological applications, brain-computer interface (BCI), to
restore communication to those affected by speech and language disorders.
Electrocorticography (ECoG), the measurement of electrical activity from cortex, has become a
highly promising neural signal acquisition modality for studying speech and language processing due
to its capability to provide high spatial and temporal resolution.
Motivation
ECoG-based BCI for communication have been developed by using simplistic linear mapping
between the sound and the ECoG signals. As a result, the performance is not nearly as robust as
needed for ECoG-based BCI.
Considering that human brain is semantic system to improve the performance. The sound
stimuli are eight monosyllabic Korean words; consist of two semantic groups (human body or
number). The event-related spectral perturbation (ERSP) is used to measure stimulus-related
brain activity. The decoding features are calculated by averaging ERSPs in each frequency
band and time window. The minimal-redundancy-maximal-relevance criterion is used to
reduce features dimensionality. The machine-learning algorithm is multi-layer support vector
machine for classification. The semantic-hierarchical model has the sub-model to classify
body group and number group prior to classification ECoG signal to one of 4 words.
Method
The study will compare two classification models to proof our hypothesis. The non-
hierarchical model directly classifies ECoG signals to one of 8 words without any sub-model.
4-fold cross validation is employed to calculate overall classification accuracy for both

2. semantic-hierarchical and non-hierarchical model. The number of ERSPs is 4 dataset in each
class.
Test data: 1/4 of ERSPs in each class
Training data: 3/4 of ERSPs in each class