EEG Workload Classifier
Logistic-regression and CNN baselines for classifying mental workload levels from raw EEG signals.
Why it mattersA reproducible biosignal-classification pipeline - frequency-domain feature engineering, dimensionality reduction, and apples-to-apples evaluation across linear and deep baselines.
What it does
Mental-workload classification from raw EEG signals via two complementary models: a lightweight, interpretable logistic regression and a deeper CNN. Both share the same feature pipeline and cross-validation harness so results are directly comparable.
Where it applies
- Biosignal research where reproducibility matters more than a single big model - the linear baseline is there to show the deep model is doing actual work, not just overfitting.
- A teaching example for frequency-domain feature engineering on physiological signals: Fourier transforms, band-power features, and dimensionality reduction.
- The pattern transfers to other multi-channel time-series classification problems - swap EEG for IMU, ECG, or EMG and the same scaffolding applies.
How it works (high level)
Raw .mat EEG files are preprocessed and split with K-fold and stratified K-fold cross-validation. Frequency-domain features (FFT-based band power per channel) feed the logistic-regression baseline; the same windows feed a CNN that learns its own filters. Outputs include per-fold accuracy, confusion matrices, and training curves, so the comparison is auditable.
Outcome
An apples-to-apples comparison of a linear baseline and a deep model on the same biosignal task, with a reusable feature-engineering and evaluation harness.
Stack
Python · Keras · scikit-learn · SciPy · NumPy.