CNN+LSTM for EEG classification, training accuracy not getting better than 50 in 2 classes?

Hello everyone

I am trying to use CNN LSTM networks for classifying healthy EEG data from disease one, I am using a signal with 1266 length and 3 channels for this purpose. but my training accuracy is always 50 and doesn't get better. here is my code and training progress plot.thanks in advance for helps.

numFeatures = 3;

numHiddenUnits1 = 100;

numHiddenUnits2 = 150;

numHiddenUnits3 = 200;

numClasses = 2;

numFeatures = 3;

numClasses = 2;

filterSize = 3;

numFilters = 8;

layers = [ ...

sequenceInputLayer(3,Normalization="zerocenter")

% sequenceFoldingLayer('Name','fold')

convolution1dLayer(filterSize,numFilters,Padding="same")

reluLayer

convolution1dLayer(filterSize,2*numFilters,Padding="same")

reluLayer

convolution1dLayer(filterSize,4*numFilters,Padding="same")

reluLayer

convolution1dLayer(filterSize,8*numFilters,Padding="same")

reluLayer

convolution1dLayer(filterSize,8*numFilters,Padding="same")

reluLayer

globalAveragePooling1dLayer

% sequenceUnfoldingLayer('Name','unfold')

flattenLayer('Name','flatten')

bilstmLayer(numHiddenUnits1,'OutputMode','sequence')

layerNormalizationLayer

% dropoutLayer(0.6)

bilstmLayer(numHiddenUnits2,'OutputMode','sequence')

layerNormalizationLayer

% dropoutLayer(0.6)

bilstmLayer(numHiddenUnits2,'OutputMode','sequence')

layerNormalizationLayer

bilstmLayer(numHiddenUnits3,'OutputMode','last')

layerNormalizationLayer

dropoutLayer(0.6)

fullyConnectedLayer(numClasses)

softmaxLayer

classificationLayer];

lgraph = layerGraph(layers);

maxEpochs = 250;

miniBatchSize = 30;

options = trainingOptions('adam', ...

'ExecutionEnvironment','gpu', ...

'MaxEpochs',maxEpochs, ...

'InitialLearnRate',0.01,...

'LearnRateSchedule','piecewise', ...

'LearnRateDropFactor',0.1, ...

'LearnRateDropPeriod',30, ...

'MiniBatchSize',miniBatchSize, ...

'GradientThreshold',1, ...

'ValidationData',{zschannel2delay6test,ytest2},...

'shuffle','every-epoch',...

'ValidationFrequency',50,...

'ValidationPatience',Inf,...

'Verbose',false, ...

'Plots','training-progress');

net = trainNetwork(zschannel2delay6,ytrain2,lgraph,options);

[Farnaz Garehdaghi]

and I have 440 samples for train and 160 samples for test. My train data is [440 1] cell and every row is [3 1266] and my labels matrix is a categorical [440 1] matrix.

Answer 1

[Gayathri]

Hi @Farnaz Garehdaghi,

There are some points which you can try to improve the accuracy of the model.

• A learning rate of 0.01 might be too high. Try reducing it to 0.001 or even lower to see if the network starts to learn better.
• Your network might be too complex for the given task. Try reducing the number of layers or hidden units in the LSTM layers.
• Consider augmenting your data to provide more training examples.
• Try training the model on a small subset of the data to see if it can overfit. If it can, the model architecture might be appropriate, and the issue could be with hyperparameters or the data.
• Experiment with different batch sizes. Smaller batch sizes can sometimes help the model converge better.

Hope this helps!