Meta Learning for EEG Classification in Schizophrenia

This repository contains code for the feature extraction process and the experiments necessary to replicate the results from the workshop paper Meta-Learning on Spectral Images of Electroencephalogram of Schizophrenics, previously accepted in the MeL4CV workshop at AAAI 2021.

The workshop paper explores the possibility of using popular meta learning models (MAML and Prototypical networks) to classify stacked spectrograms. These spectrograms are representations of electroencephalogram (EEG) readings which were converted from continuous time-series to sets of images. Specifically, two EEG datasets were used in the experiments; Dataset-1 was split into 20 second slices and Dataset-2 was split into 5-second slices. Samples were acquired with an overlap rate of 20% to provide more exhaustive and substantial input.

Data Processing Pipeline

A notebook for converting Dataset-2 EEG values to spectrograms, and a set of sample spectrogram images can be found in the files below: /Spectrogram Generators/Spectrogram_Generator_for_Dataset-2_(bio_msu_ru)_All_Channels_fractional_noverlap_SML.ipynb
/Data/Extracted/Sample_Spectrograms

For both EEG datasets, MAML and prototypical networks were used to train and test the model. In an additional set of experiments with MAML, pre-training was performed with miniImageNet, followed by further training with Dataset-2 and testing with Dataset-1 (and the reverse).

/MLModels/Meta Learning Models/MAML_Pytorch_with_Dataset-1.ipynb
/MLModels/Meta Learning Models/MAML_Pytorch_with_Dataset-2.ipynb
/MLModels/Meta Learning Models/MAML_Pytorch_with_further_training_Test_Dataset-1.ipynb
/MLModels/Meta Learning Models/MAML_Pytorch_with_further_training_Test_Dataset-2.ipynb
/MLModels/Meta Learning Models/Prototypical_Networks_with_Dataset-1.ipynb
/MLModels/Meta Learning Models/Prototypical_Networks_with_Dataset-2.ipynb

Results

Network Name	Val Loss	Val Acc	Test Acc	Test AUC	Test F1 Macro
Dataset-1
CNN	1.6223	0.5000	0.6646	0.8128	0.6468
CNN + Fine Tuning	0.6348	0.5455	0.5021	0.5014	0.3579
Prototypical Networks + CNN	0.0112	0.9977	0.9768	0.9772	0.9768
MAML + CNN	0.5000	0.8945	0.8420	0.8426	0.8419
MAML + CNN with Further Training	0.2184	0.9342	0.9683	0.9342	0.9341
Dataset-2
CNN	0.0000	1.0000	1.0000	1.0000	1.0000
CNN + Fine Tuning	0.5805	0.5014	0.6814	0.6892	0.6560
Prototypical Networks + CNN	0.5980	0.8333	0.8587	0.8580	0.8584
MAML + CNN	1.317	0.7749	0.8133	0.8122	0.8097
MAML + CNN with Further Training	0.1888	0.9541	0.9489	0.9489	0.9489

This repository contains additional feature extraction methods and ML experiments which were not included in the workshop paper. The notebooks mentioned above are intended to be executed on Google Colab.

Datasets

Dataset-1

Olejarczyk, E.; and Jernajczyk, W. 2017. EEG in Schizophrenia.
doi:10.18150/repod.0107441. URL https://doi.org/10.18150/repod.0107441.

Dataset-2

Gorbachevskaya, N. N.; and Borisov., S. V. 2015. EEG of healthy adolescents and
adolescents with symptoms of schizophrenia. EEG Database - Schizophrenia
URL http://brain.bio.msu.ru/eeg_schizophrenia.htm.

Mini-ImageNet

Vinyals, O.; Blundell, C.; Lillicrap, T.; Wierstra, D.; et al. 2016. Matching networks
for one shot learning.
In Advances in neural information processing systems, 3630–3638.