Sleep Spindle Detection using CNN

A Master's thesis project by Subahini Nadarajh (University of Basel)

---

Project Overview

This project focuses on detecting sleep spindles in EEG recordings using a Convolutional Neural Network (CNN). The goal is to build a model that learns to identify spindles directly from raw EEG signals — without relying on pre-existing models for labeling.

The project uses data stored on the DBE lakeFS server and processes it into a clean format for model training and evaluation. The work includes signal preprocessing, label alignment, and training a CNN with temporal windows of EEG data.

Getting Started

1. Clone the Repository

git clone https://github.com/yourusername/spindle-project.git
cd spindle-project

2. Create a Virtual Environment

python3 -m venv .venv
source .venv/bin/activate

3.Install Dependencies

pip install -r requirements.txt

---

Configuration

Edit the file config/pipeline.yaml to point to the EEG file and label file you want to work with.

---

Downloading Data

Download EEG from lakeFS:

python code/download_raw.py

Download Spindle Labels:

python code/download_labels.py

---

Preprocessing

Segments the EEG into overlapping windows and labels each based on spindle presence (from JSON):

python code/preprocess.py

You’ll get:

• data/X_windows.npy: preprocessed EEG windows (e.g., shape [N, 6, 512])
• data/y_labels.npy: corresponding binary labels (0 = no spindle, 1 = spindle)

---

Training the CNN

Train the model using:

python code/train_cnn.py

The model uses a lightweight 1D CNN that:

• Takes EEG from 6 channels
• Detects temporal patterns
• Outputs a probability of spindle presence

Author

Subahini Nadarajh
subahini.nadarajh@stud.unibas.ch

Examiner & Supervisor: Prof. Dr. Volker Roth Supervisor: Florentin Bieder Project Collaborators: Prof. Alex Datta and Dr. Martina Studer (UKBB)

Faculty of Science, University of Basel Department of Mathematics and Computer Science

In collaboration with the Department of Biomedical Engineering CIAN Group – Center for medical Image Analysis & Navigation and University Children’s Hospital Basel (UKBB)

---

Acknowledgements

• Data courtesy of the DBE sleep spindle research team
• lakeFS for large-scale data versioning
• MNE and PyTorch libraries