EEG Harmful Brain Activity Classification - With 1D, 2D ConvNets

Index

• Description
• Setup
• Research Notes & Thought Process

Description

This repository contains utilities, custom DataSet and DataLoader classes, notebooks, and other tools developed for harmful brain activity classification using data from the HMS Kaggle competition. Currently, the model achieves state-of-the-art performance, with a KLDivLoss per example of 0.27.

I am also logging the lessons I've learned throughout this project.

What's Next?

For the 2D model, I’m exploring other signal processing augmentation techniques, such as focusing on the middle 10 seconds of the spectrogram, inspired by this Kaggle post.

I’m also exploring the theory behind time series data classification using 1D convolutions and plan to develop a 1D ConvNet.

Finally, I plan to deploy this on my personal website to gain experience designing user-friendly UI/UX for AI interfaces (coming soon).

Setup

This project uses conda environments.

Clone the repo:

git clone https://github.com/sumkawa/icu-lstm.git
cd icu-lstm

Install dependencies, required packages, and setup conda development environment:

make install

Get data from kaggle competition:

kaggle competitions download -c hms-harmful-brain-activity-classification -p ./data

Activate the conda environment:

conda activate icu_classifier_env

Converting .parquet to .npy, visualize initial data

Give proper user permissions:

chmod +x visualize.py parq_to_npy.py

Seed database by converting parquet to npy for more efficient data loading and model compatibility (this step is only necessary if training with efficient net - mixnet implementations convert raw EEG data on the fly):

python parq_to_npy.py

(Optional) Visualize data with matplotlib:

python visualize.py