Adaptive Chirplet Transform and Deep Learning Algorithms for EEG-based Sleep Stage Detection
27th Annual Mersivity / WaterHCI Symposium, Summer 2025
Toronto and online, June 20 to August 25, 2025 · Published in the symposium proceedings (open access on Zenodo)
Eyes closed, headband on: the kind of consumer-EEG recording this pipeline is built to stage.
Sleep staging from EEG is usually done with Fourier or wavelet features. This thesis-style paper reviews the deep-learning side of the field and builds a pipeline that feeds the Adaptive Chirplet Transform, a time-frequency method Mann designed for signals whose frequency drifts, into those networks. It is a foundation for the club's later sleep work.
Abstract
Electroencephalography (EEG) is a powerful tool for analyzing brain activity, particularly in the context of sleep stage detection. However, traditional signal processing methods like the Fourier and Wavelet Transforms often struggle to capture the non-stationary and time-varying nature of EEG signals. This thesis explores the Adaptive Chirplet Transform (ACT) as a novel feature extraction technique for EEG-based sleep stage classification using deep learning. The ACT offers a dynamic time-frequency representation by matching signal segments to Gaussian chirplets. A GPU-accelerated implementation of the ACT was developed and applied to the Bitbrain Open Access Sleep Dataset, processing over 10,000 epochs from three nights of sleep. Feature vectors extracted by the ACT were reshaped into 5×5 matrices and used as input to a hybrid Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU) model. The model achieved a training accuracy of 91.5% and a test accuracy of 57.9%, demonstrating the ACT's potential to encode relevant features with a 153.6:1 compression ratio. The findings suggest that, with further optimization and scaling, the ACT could become the next state of the art signal processing and feature extraction method for EEG.
Authors
João Pedro Bicalho Andrade, Steve Mann
Links
- Paper in the Mersivity / WaterHCI Summer-2025 Symposium Proceedings (Zenodo, open access)
- Related project: State-of-Sleep in HBOT
- Related project: Asymmetric Chirplets
- Related project: Chirplets in the Snow