Page 170 - AI for Good Innovate for Impact

AI for Good Innovate for Impact

(continued)

Item Details
Model Training and Fine-Tun- A Support Vector Classifier (SVC) model is trained using the
ing preprocessed and noise-filtered dataset. The model employs
a Radial Basis Function (RBF) kernel, which is well-suited for
handling the complex patterns present in EEG signals. Initially, a
basic SVC model is trained with default parameters. To improve
performance, hyperparameter optimization is performed using
techniques like GridSearchCV. Parameters such as the regulariza-
tion term (C) and the gamma parameter are fine-tuned through
5-fold cross-validation.
Testbeds or Pilot Deploy- We curated models using data from 10 different people, and test-
ments ing has shown very promising results. We also have Memoranda
of Understanding (MoUs) with hospitals to further test and refine
our system.
Code repositories Not Available

2 Use Case Description

2�1 Description

Paralysis victims and senior citizens have always been at the mercy of someone else to help
them carry out their day-to-day activities in life. This makes their life more miserable, knowing
they will never be independent again. We aim to address this issue through the implementation
of our brain-controlled wheelchair. Our devices aim to solve all their mobility concerns to make
them feel more at ease and regain control of their lives. Our product is a brain wave-controlled
wheelchair, which makes use of EEG technology to obtain readings of the brain signals. All of
these signals are later preprocessed and then classified using machine learning and fed to the
wheelchair, which is then set to perform the various functions as and when they are required.
Our product poses an innovative and significantly affordable alternative to the options available
in the market for the locomotion of paralysis patients. Upon widespread implementation, it will
make the lives of the users much more comfortable. The data is collected from 10 individuals
who are made to wear the sensors according to the international 10-20 system. The sensors
are placed on the motor cortex to detect any possible movements. Each individual is later
asked to behave differently to obtain different kinds of data. At first, they are asked to remain
in a calm state – think of only pleasant thoughts and nothing too intense. Later, they are made
to maintain a position where they are told to primarily focus on clenching their jaw, and finally,
the next set of data is to determine active movements, so they are told to put as much stress as
possible into thoughts about limb movement without actually moving them. This is to simulate
the conditions of the paralysed people and to train the model based on the respective data.

Each state is maintained for about 20 minutes when the data is continuously collected and
then labelled according to what the particular state was. In all, an individual is observed for 60
minutes to obtain 1.8 million data points. When the data from all the participants is combined,
the total obtained is 18 million data points. All of these points are labelled according to the user
they were collected and the particular state. The data collected is stored as comma-separated
value files stored locally on a system. This data is further used in training the model to obtain

134