Training a Convolutional Neural Net to Detect Epileptic Spikes in EEG (Industry project)

In this project, we will train a convolutional neural net (CNN) to detect epileptic spikes in the EEG. For training data, we will use the Temple University Hospital EEG Events Corpus (https://isip.piconepress.com/projects/tuh_eeg/html/downloads.shtml), a collection of EEGs from epileptic patients, manually annotated with spikes and other epilepsy-related events. We will organize the data with MongoDB, write parsers for the EEG and annotation files, and use PyTorch to build out a training pipeline which feeds the raw EEG signals to a CNN (similar to image processing, but 1-dimensional). Finally, we will train our CNN and evaluate its performance.

Objectives

  1. A trained ML model that detects epileptic spikes in an EEG signal
  2. An evaluation of the performance of the ML model
  3. A customizable pipeline for prototyping epileptic spike detection models

Motivations

Electroencephalography (EEG) is a powerful brain imaging method that provides a window into native electrical activity in the brain (brain waves) with millisecond precision. Today, EEG has applications in several domains, including sleep monitoring, disease diagnosis, evaluating brain trauma, attention, learning, and memory research, neuromarketing, and even brain computer interfaces. Much of EEG analysis and interpretation comes down to classifying sections of the EEG signal. For example, separating the noisy (artifactual) sections of the signal from the clean sections is critical for most analyses. Manually annotating EEGs is time-intensive work, so many EEG researchers have turned to machine learning approaches to automate these tasks. One important EEG annotation task is the detection of epileptic spikes in the EEGs of people with epilepsy. Clinicians use these spikes to detect and predict seizure onset and to locate the region of the brain from which epileptic seizures arise. 

Qualifications

Minimum Qualifications:

Python, Git/Github, command line interface

Preferred Qualifications:

MongoDB, machine learning, statistics


Details

Project Partner:

Roma Shusterman

NDA/IPA:

No Agreement Required

Number Groups:

1

Project Status:

Accepting Applicants

Card Image Capstone