In my experience that you can use the transfer learning by using the model pre-trained for most famous are VGG-16 VGG-19 or ResNet-34 Resnet-50 ......ect in python and we can develop also architecture of CNN for Classification your datasets. I hope that helpful .

Dear Fazla,

I hope these sources to help you.

Feature classification of EEG signal with binary heuristic optimization algorithms

Conference Paper Feature classification of EEG signal with binary heuristic o...

for python code:

A Collection of Convolutional Neural Network (CNN) models for EEG signal processing and classification, written in Keras and Tensorflow.

https://github.com/vlawhern/arl-eegmodels

EntroPy is a Python 3 package providing several time-efficient algorithms for computing the complexity of one-dimensional time-series. It can be used for example to extract features from EEG signals.

https://github.com/raphaelvallat/entropy

check deep learning fundamentals for answer

Hello Fazla Rabbi Mashrur

Kindly have a look into https://machinelearningmastery.com/feature-selection-with-real-and-categorical-data/ this may be an answer to your question or near to it.

regards!

Suraj

In my experience that you can use the transfer learning by using the model pre-trained for most famous are VGG-16 VGG-19 or ResNet-34 Resnet-50 ......ect in python and we can develop also architecture of CNN for Classification your datasets. I hope that helpful .

Transfer learning on famous pre-trained models is not the best approach because you have small dataset. You could design NN architecture from scratch. Don't use a large number of layers and units. You can extract EEG features using more appropriate from the state of the art approaches and then use it as NN input.

Another solution is to get a pre trained network on other EEG problem and use it as feature extractor and then classificate features with SVM.

To solve dataset unbalance try to set mini random batches with equal numbers of both classes.

Thanks a lot Mirko Agarla Barodi Anass Suraj Bhagat Bhanu PRAKASH Kolla Ibrahim mohamed Gad

Fazla Rabbi Mashrur

1. You can use mini batch balancing scheme for handle the data imbalancing issue ,even the domain invariences and if you want to oversample the minority class you can use VAE network .

2.Automated feature selection would be a good choice .. and for that 1d CNN +BiLSTM architecture with a signal processing level strategy can help the model to get both the spectral and temporal features . You can look the DeepSleepNet architecture .