On - Device Continual Learning for Seizure Detection on GAP9

Overview

Status: In progress

• Type: Master/Semester Thesis
• Professor: Prof. Dr. L. Benini
• Supervisors:
  • Thorir Mar Ingolfsson (IIS): thoriri@iis.ee.ethz.ch
  • Cristian Cioflan (IIS): cioflanc@iis.ee.ethz.ch
  • Dr. Andrea Cossettini (IIS -- EFCL): cossettini.andrea@iis.ee.ethz.ch

Short Description

The student will extend the existing work on Continual Learning for EEG-based seizure detection by optimizing and implementing the best methods on the GAP9 embedded device, with the goal of improving real-time, on-device seizure detection.

Introduction

Exploring neural activity in the human brain is an integral area of scientific research, offering crucial insights into various medical applications, including the detection and prediction of epileptic seizure activity. Electroencephalography (EEG), a non-invasive technique for recording brain activity, is a critical tool in this regard. Traditional EEG systems, while useful, often present challenges as they are large, cumbersome, and can be stigmatizing for the user.

In response to these limitations, the focus has shifted towards developing compact, energy-efficient wearable devices that enable long-term EEG monitoring. These devices need to be resilient to diverse noises or artifacts that can affect the EEG signal, such as external disturbances or patient movement.

Epilepsy, a central nervous system disorder, is characterized by abnormal brain activity resulting in seizures, unusual behavior, sensations, and occasionally, a loss of awareness. The standard practice for detecting and predicting epileptic seizure activity involves training systems on data collected within an epilepsy monitoring unit (EMU). Within the controlled conditions of an EMU, the data is typically devoid of artifacts. However, upon leaving the EMU, wearable EEG monitoring devices may encounter less ideal conditions, leading to EEG data that is frequently tainted with artifacts or varies significantly from the data used to train the initial seizure prediction model.

This poses an intriguing challenge as models trained with clean, reliable EEG data may underperform when confronted with variations in brain activity and data quality. While bringing the patient back to the EMU is a possibility, it's not always feasible and often inconvenient. Instead, a more promising approach involves continually fine-tuning the deployed models on the edge monitoring device they are installed on, leveraging patient feedback. This method allows the model to adapt to changes in brain activity and data quality, thereby ensuring more accurate and reliable predictions of epileptic seizure activity.

The procedure of perpetually refining a pre-trained neural network is referred to as Continual or Incremental Learning (CL). In this context, we distinguish two categories of CL: Class CL, where the model recognizes an increasing number of classes over time, and Domain CL, which maintains a fixed number of classes but acknowledges that the environment where samples are collected changes over time. If data regarding the target domain (e.g., user ID) is available for the system, we establish a third category: Task CL. Given the variability in the EEG data acquisition, this project targets the Domain CL scenario. In such a setting, the proposed system must adapt to the characteristics of its current user without forgetting the properties of prior users who might reuse the system. This necessitates mitigating the "catastrophic forgetting" phenomenon in our model.

Recently, several methodologies for Class Incremental Learning have been proposed (FACIL), while fewer techniques have been employed for Domain Incremental Learning (https://github.com/GMvandeVen/continual-learning, https://www.nature.com/articles/s42256-022-00568-3, https://arxiv.org/abs/2204.08817). The objective of this project is to propose and evaluate Domain CL methods to accommodate intra- and inter-patient variability in EEG data acquisition. While doing so, the proposed system must adhere to TinyML constraints, meaning that the adaptation process must be computationally efficient, and memory and storage requirements must be minimized. This project will contribute significantly to the development of an adaptable, efficient, and reliable wearable system for epileptic seizure detection and prediction.

This project aims to implement and optimize Continual Learning methods on the GAP9 embedded device, which will help enhance the efficiency and reliability of on-device seizure detection systems. The previous work focused on mitigating challenges in ML methods for seizure detection by using Continual Learning techniques, and this project seeks to build upon those findings and further improve upon them.

Character

• 20% Literature research
• 30% Evaluation
• 50% Software implementation and optimization

Prerequisites

• Familiarity with Python.
• Knowledge of Deep Learning and Continual Learning techniques.
• Experience with Embedded Systems, ideally with GAP8 or GAP9 devices.
• Proficiency in a deep learning framework like PyTorch or TensorFlow.

Project Goals

The main tasks of this project are:

• Task 1 - Literature Review (2 Weeks)

  Conduct a comprehensive literature review of the latest advancements in the field of on-device machine learning, specifically focusing on continual learning for medical applications.

• Task 2 - Continual Learning Model Development (2 Weeks)

  Develop a machine learning model for seizure detection using the optimal continual learning techniques identified from the previous work and new literature review.

• Task 3 - Model Testing and Refinement (2 Weeks)

  Conduct initial testing and validation of the developed model using seizure datasets. Analyze the results, identify areas of improvement, and refine the model accordingly.

• Task 4 - Implementation on GAP9 (2 Weeks)

  Begin the adaptation process to implement the model on the GAP9 embedded device. This includes work to ensure the model is optimized for low-power consumption and real-time operation.

• Task 5 - On-Device Model Testing (2 Weeks)

  Conduct testing of the model on the GAP9 device. This will be an iterative process, involving debugging, refinement, and repeated testing.

• Task 6 - Performance Evaluation (2 Weeks)

  Perform a comprehensive evaluation of the model's performance in terms of seizure detection accuracy, power consumption, and real-time operation. Compare the model's performance to the previous work.

• Task 7 - Report Writing and Presentation Preparation (2 Weeks)

  Summarize the project's work into a detailed report. The report should include the design process, challenges faced, solutions devised, performance metrics, and recommendations for future work. Prepare a presentation summarizing the project and its findings.

Project Organization

Weekly Meetings

The student shall meet with the advisor(s) every week in order to discuss any issues/problems that may have persisted during the previous week and with a suggestion of next steps. These meetings are meant to provide a guaranteed time slot for mutual exchange of information on how to proceed, clear out any questions from either side and to ensure the student’s progress.

Report

Documentation is an important and often overlooked aspect of engineering. One final report has to be completed within this project. Any form of word processing software is allowed for writing the reports, nevertheless the use of LaTeX with Tgif (See: http://bourbon.usc.edu:8001/tgif/index.html and http://www.dz.ee.ethz.ch/en/information/how-to/drawing-schematics.html) or any other vector drawing software (for block diagrams) is strongly encouraged by the IIS staff.

Final Report

A digital copy of the report, the presentation, the developed software, build script/project files, drawings/illustrations, acquired data, etc. needs to be handed in at the end of the project. Note that this task description is part of your report and has to be attached to your final report.

Presentation

At the end of the project, the outcome of the thesis will be presented in a 15-minutes talk and 5 minutes of discussion in front of interested people of the Integrated Systems Laboratory. The presentation is open to the public, so you are welcome to invite interested friends. The exact date will be determined towards the end of the work.