saeub / cogpred

CogPred

Experiments in predicting hearing loss (based on pure tone average (PTA)) from EEG recordings.

Setup

• pip install -r requirements.txt

Directory structure

• src/: source code
• data/
  • raw/EEG/${subject_id}/: raw EEG data (1 folder per subject)
    • EEG1/${subject_id}_Termin1_CogTAiL.bdf: raw EEG data file
  • preproccessed
    • eeg: preprocessed EEG data (PyTorch)
    • measurements: auditory and cognitive measurements
      • PTA.csv: auditory threshold measurements
      • cognitive_measures.csv: cognitive measurements

EEG preprocessing

We only perform minimal automatic preprocessing (cropping/aligning, bandpass filtering, and optional time-frequency analysis).

1. Make sure that all the raw data files are linked/mounted in the directory structure described above.
2. Run python src/preprocess.py with subject IDs as arguments. To preprocess all: ls data/raw/EEG/ | xargs python src/preprocess.py
   Use --tfr to generate the time-frequency representation instead of the raw time-domain signal. This adds a frequency dimension and reduces time resolution).

Models

The model architecture is a CNN with three convolution and max-pooling layers (two in the case of TFR input, due to the already reduced dimensionality), batch normalization and optional dropout, followed by two dense layers. See models.py for details.

Training and evaluation

python src/main.py will run ten-fold cross-validation with default hyperparameters. python src/main.py --help shows configurable hyperparameters.

Results

In the table below, accuracy, F1, and area under the ROC curve (AUC) are the means of those values across all ten test folds. The models were trained for a maximum of 40 epochs, and the checkpoint with the best validation AUC score was used for evaluation.

Preprocessing	Channel grouping	Dropout	Accuracy	F1	AUC
raw	no	0.0	0.615	0.641	0.508
raw	no	5.0	0.585	0.640	0.466
raw	yes	0.0	0.508	0.470	0.577
raw	yes	5.0	0.531	0.485	0.536
tfr	no	0.0	0.654	0.673	0.485
tfr	no	5.0	0.592	0.515	0.457
tfr	yes	0.0	0.562	0.504	0.550
tfr	yes	5.0	0.431	0.255	0.493

Other things I tried

• Regression instead of classification (didn't learn anything)
• Not upsampling the minority class (didn't learn anything or very slowly)
• Predicting n-back performance instead of PTA (wasn't any easier)