This repo contains scripts for a Kaggle competition (1) aimed at predicting seizures from 10-minute clips of EEG voltage recordings. The data came from 2 human subjects and 5 dogs.

The base directory includes modules with functions used for generating features from the raw time series data, exploratory analysis, training and optimizing generic models, and applying models to test data.

The computed features (but not the raw data) are stored in the data directory. Scripts to train and optimize specific classification models, using the functions in train_model.py and optimize_model.py, can be found in scripts, and the submissions directory contains CSV files and log files for all submitted predictions.

One of the main challenges was normalizing the predicted probabilities across the 7 different subjects to maximize the area under the ROC curve (AUC) for all subjects. Despite trying isotonic regression and other methods to combine the predictions, the combined AUC was limited to around 0.6 while the estimated AUC for individual subjects was in the range 0.75 to 0.95.