Sleep analytics re-dreamed 💤

Deep Learning-based tooling from the Lamarr Institute for scientists, clinicians & innovators.

Explore the project ↓

Automated Sleep-Stage Tagging and Beyond

Polysomnography (PSG) is the gold standard for diagnosing sleep disorders, but its manual analysis is time-consuming and limited to a few core tasks. Current deep learning methods focus mainly on sleep stage tagging, leaving other key events such as apneas, arousals, and desaturations underexplored. In our ongoing work, we integrate leading models - MRASleepNet, TinySleepNet, SleepTransformer, USleep, SeqSleepNet, and AttentionSleep - into a unified benchmark suite. Trained on thousands of hours of PSG data, these models are evaluated not only for sleep stage tagging but also for broader clinical event detection, already outperforming Nox Medical's Noxturnal software on multiple tasks. Stay tuned - results coming soon.

Automated Sleep-Stage Tagging and Beyond

Knowledge Distillation for Wearables

Wearable-based sleep monitoring can overcome the burdens of lab-based polysomnography that is resource-intensive, uncomfortable for patients, and prone to the first-night effect, where the lab environment alters natural sleep patterns. However, wearable-based sleep monitoring often lacks the accuracy necessary for a real-world deployment. Our knowledge distillation approach transfers expertise from high-performing PSG models to wearable-based models, consistently improving performance across configurations — paving the first steps towards a breakthrough that could revolutionize in-home patient care.

Paper has been accepted at ECML-PKDD 2025 MedTime Workshop .

Knowledge Distillation for Wearables

Foundational Models for EEG

Our goal is to move state-of-the-art EEG analysis beyond large research centers and into smaller facilities that often lack the compute resources and training data to develop their own models. We are building foundational models for the medical domain, designed to encode time-domain electroencephalogram (EEG) data into rich, high-dimensional latent representations. EEG data play a vital role in neurology, from sleep medicine to the diagnosis and treatment of epilepsy. Trained on millions of EEG patterns from tens of thousands of patients, our models are robust, versatile, and can be adapted to a variety of tasks with minimal additional data or computational resources. This makes it possible to deliver faster, more accurate diagnoses and improved treatments even in settings with limited resources. Stay tuned - results coming soon.

EEG foundation model

LLM-Powered Dashboard

We developed an interactive dashboard that visualizes sleep lab data alongside model predictions and integrated a Large Language Model (LLM) for deeper insights. Using advanced prompting techniques, the LLM provides clear, context-aware explanations of classifications while our state of the art classification models provide the predictions — helping clinicians understand complex AI decisions and paving the way for more transparent, data-driven sleep diagnostics.

Coming soon ↗

Dashboard preview

Get in Touch

Interested? Then contact our researchers directly! For research we also provide direct access to our analysis models.

✉️ sleepwalker.lamarr.cs@tu-dortmund.de

Meet the team

Dr. rer. nat. Sebastian Buschjäger
Dr. rer. nat. Sebastian Buschjäger
Computer Scientist and AI expert
Dr. rer. medic. Sarah Dietz-Terjung
Dr. rer. medic. Sarah Dietz-Terjung
Biologist and medical physicist
MSc. Matthias Jakobs
MSc. Matthias Jakobs
Computer Scientist and AI expert