Wearable AI Case Study

Real-time sleep staging from wearable EEG

A smartphone-based system for classifying sleep stages from single-channel EEG using deep learning and power spectral analysis.

IEEE BSN Paper↗︎arXiv Paper↗︎Code Repository↗︎

Wearable EEG sleep staging interface with hypnogram and signal visualization — Smartphone-based sleep staging pipeline using wearable EEG, real-time inference, and sleep-stage visualization.

Technical Layer

Tools & Methods

Wearable EEG
TensorFlow Lite
Deep learning
Sleep-EDF
Muse
Power spectral analysis
BioEssence integration

Period2018–2019

Signal1-channel EEG

PlatformSmartphone

Model1D Convolutional Neural Net

Outcome⋅ NeurIPS ML4H workshop paper ⋅ IEEE BSN conference oral presentation

My role

At the MIT Media Lab, I developed an end-to-end sleep-staging smartphone application using a deep-learning model trained on Sleep-EDF and adapted for wearable EEG. I built EEG and IMU data pipelines, implemented raw EEG visualization, and programmed EEG power spectral analysis across alpha, beta, gamma, delta, and theta bands. I also developed a connector module linking the sleep application with BioEssence, a wearable olfactory display designed to monitor cardio-respiratory information and support physiology-aware scent interventions for mental wellbeing.

The problem

Polysomnography is the gold standard for sleep staging, but it is resource-intensive, wired, and usually dependent on expert scoring in a sleep-lab setting. The project asked whether a lightweight wearable EEG and smartphone pipeline could classify sleep stages in real time.

The system

The system used a modified Muse EEG headband, Bluetooth data streaming, smartphone-based inference, and a TensorFlow Lite application to classify sleep stages from single-channel EEG.

The modelling approach

The deep-learning pipeline adapted a time-distributed convolutional neural network for five-stage sleep classification using the open Sleep-EDF dataset, with inference designed to run directly on a smartphone.

The signal-analysis layer

Alongside the model, I programmed power spectral analysis to inspect EEG power bands including alpha, beta, gamma, delta, and theta across sleep stages. This helped provide an interpretable signal layer beyond the neural-network output.

Why this matters

For digital health and neurotechnology, this project connects wearable biosensing, on-device machine learning, real-time physiological inference, and intervention-ready sleep interfaces.

Overview of real-time smartphone-based sleep staging using wearable EEG — Overview of the real-time sleep-staging pipeline: wearable EEG, smartphone inference, hypnogram output, and EEG power-band visualization.

Single-channel 1D CNN architecture for real-time sleep staging — Single-channel 1D CNN architecture used for sleep-stage classification from wearable EEG time-series data.

Giving a selected oral presentation at the UIC Dorin Forum - Chicago — Selected for oral presentation at the IEEE BSN 2019 conference, UIC Dorin Forum - Chicago

Primary outputs

IEEE BSN PaperReal-time Smartphone-based Sleep Staging using 1-Channel EEG · IEEE BSN 2019Open ↗︎arXiv PaperReal-Time Sleep Staging using Deep Learning on a Smartphone for a Wearable EEG · ML4H/NeurIPS 2018Open ↗︎Code RepositoryRealTimeSleepStaging GitHub repositoryOpen ↗︎

Featured in

Coverage

MIT Media LabProject overview and research contextRead ↗︎BioEssenceWearable olfactory display and physiological intervention platformRead ↗︎Seeker VideoThese Sleep Engineers Could Help You Hack Your DreamsWatch ↗︎NeurIPS ML4H 2018Machine Learning for Health workshop at NeurIPS 2018Read ↗︎

Continue Exploring

Explore the wider body of work.

Further projects span physiological computing, computational neuroscience, embodied systems, and interdisciplinary engineering.

View more projects Professional enquiries