In 2008, a BCI Competition was held on EEG datasets to find the best ML and statistical algorithms to differentiate different classes of neural data. The BCI Competition IV 2b is a motor imagery dataset with eye artifact data, making it a very realistic dataset. The subjects are prompted to imagine left vs right hand movement and the EEG + EOG signals for each trial are collected.
The first three notebooks 00_Starter Notebook, 01_Feature Extraction, and 02_Feature Modeling are preliminary notebooks that visualize the data and classify based on common feature extraction techniques (PSD, first & second order power ratio, alpha instantaneous amplitude median, etc) and statistical models (SVM, Logistic Regression, etc). The fourth notebook 03_Feature Extraction v2 implemented the extension of Filter-Bank Common Spatial Pattern (FICSP) based on Filter bank common spatial pattern algorithm on BCI competition IV Datasets 2a and 2b by Kai Keng Ang to extract features. In the final notebook 04_Feature Modeling_v2, models such as custom CNN, DNN, and other statistical models are tuned by running grid search on hyperparameters, and the best performing model is used to submit the prediction to Kaggles.
The best model selected based on F1 score is the ensemble model that runs a Logistic Regression classifier on top of the outputs of Logistic Regression, SVM, AdaBoost, and Random Forest, achieving 89% F1 score on the unseen Kaggles Dataset, being the most accurate model of the competition. (UCSD Neural Data Challenge)