9 Oct · BCV

Meetings

Link to page

Split data into 10 parts based on time (timestamps) and apply the train-test-val split for each part.

Shuffle training set.

EEGNet / DeepConvNet / ShallawNet / LSTM / DGCNN / TCN/ ATCNET / EEGTransformer

one session / cross sessions

no sampling / class weights / undersampling / oversampling / kfold cross validation (normal split)

EEGNet - one session:

EEG Data Shape: (80300, 2000)

Training label distribution:

forward

23864

stop

17672

reverse

8534

turn_right

868

turn_left

452

There are no rows in this table

⁠

Validation set label distribution:

forward

5572

stop

3452

reverse

3707

turn_right

turn_left

There are no rows in this table

⁠

Testing set label distribution:

forward

6770

stop

4013

reverse

4694

turn_right

403

turn_left

180

There are no rows in this table

⁠

No sampling:

⁠

⁠

⁠

Class weights:

⁠

⁠

⁠

Undersampling:

⁠

⁠

⁠

Oversampling:

⁠

⁠

⁠

kfold cross validation:

Undersampling

⁠

⁠

Final Average Training Accuracy: 57.27% Final Average Testing Accuracy: 36.56%

Oversampling

⁠

⁠

Final Average Training Accuracy: 77.88% Final Average Testing Accuracy: 36.58%

EEGNet - cross sessions:

1 - EEG Data Shape: (102838, 2000) 2 - EEG Data Shape: (93569, 2000) 3 - EEG Data Shape: (83650, 2000) 4 - EEG Data Shape: (80300, 2000) 5 - EEG Data Shape: (91344, 2000) 6 - EEG Data Shape: (104990, 2000) 7 - EEG Data Shape: (94723, 2000) 8 - EEG Data Shape: (92417, 2000)

Combined data shape: (743831, 2000)

Training set shape: (476041, 2000) Validation set shape: (119020, 2000) Testing set shape: (148770, 2000)

Training label distribution:

forward

230748

stop

128787

reverse

103378

turn_right

7646

turn_left

5482

There are no rows in this table

⁠

Validation set label distribution:

forward

54348

stop

34527

reverse

27077

turn_right

1564

turn_left

1504

There are no rows in this table

⁠

Testing set label distribution:

forward

75130

stop

44905

reverse

27077

turn_right

2420

turn_left

1760

There are no rows in this table

⁠

No sampling:

⁠

⁠

⁠

Class weights:

⁠

⁠

⁠

Undesampling:

⁠

⁠

⁠

Oversampling:

⁠

⁠

⁠

kfold cross validation:

Undersampling

⁠

⁠

Oversampling

Final Average Training Accuracy: 46.95% Final Average Testing Accuracy: 8.24%

Quality of EEG data:

Preprocessing Techniques Applied:

Power Spectral Density (PSD): Plots were generated before and after preprocessing to visualize changes in the signal.

Artifact Removal: Independent Component Analysis (ICA) was used to identify and remove artifacts. EOG artifact detection was skipped due to the absence of suitable EOG channels.

eog_inds, _ = ica.find_bads_eog(raw) # Detects eye-blink-related components

ecg_inds, _ = ica.find_bads_ecg(raw) # Detects heartbeat-related components

Noise Reduction: Notch filters were applied to remove powerline noise at 50 Hz, followed by a bandpass filter (1-40 Hz) to focus on EEG-relevant frequencies.

Normalization: Z-score normalization was applied to standardize the signal amplitude across channels.

⁠

⁠

⁠

⁠

⁠

⁠

Quality Metrics:

Signal-to-Noise Ratio (SNR):

Before Preprocessing: SNR was 0.00 dB, indicating low signal quality.

After Preprocessing: SNR improved to 40.42 dB, showing a significant enhancement in signal quality.

Artifact Rejection Rate: No artifacts were detected in the final data set (0.00%).

Noise Reduction: The noise power was consistent before and after preprocessing, with a rate of 0.00%, indicating that the applied filters effectively isolated the EEG signal without introducing new noise.

Total Power Analysis: Both the total power and noise power before and after preprocessing remained at 171,614.16 and 15.57, respectively.

Want to print your doc?
This is not the way.

Try clicking the ⋯ next to your doc name or using a keyboard shortcut (

CtrlP

) instead.