Machine Learning-aided sleep modulation

Sleep is structured by specific brain rhythms, most notably slow waves, which support memory consolidation, neural recovery, and overall mental health. Closed-loop auditory stimulation (CLAS) is an established, non-invasive way to interact with these rhythms during deep sleep, but current systems rely on a single EEG channel and treat all slow waves as equivalent - even though slow waves come in functionally distinct subtypes.

In this project, we use machine learning to make sleep modulation more precise. We develop deep-learning models that analyse high-density EEG (hdEEG) in real time, infer predictive features, and classify slow-wave subtypes from short EEG windows. These models drive a closed-loop system that combines real-time hdEEG analysis with auditory stimulation to selectively target specific slow-wave types, rather than slow-wave activity in general.

The project follows three lines of work that build on each other:

• Post-hoc analysis: characterising slow-wave subtypes in existing hdEEG sleep recordings and identifying the EEG features most informative for real-time classification.
• Modeling: training and validating the machine-learning pipeline on simulated and recorded data, with a focus on robustness, real-time speed, and source-level interpretability.
• Integration & clinical translation: combining the models with closed-loop hardware and pilot testing in adults, leading toward a clinical study in children with ADHD.

By combining methodological advances in machine learning with translational sleep research, this project aims to make non-invasive sleep modulation more selective, better interpretable, and ultimately more useful in clinical settings.

Group Leader: Prof. Dr. Reto Huber
Project Leaders: Dr. Samuel Wehrli & Dr. Sara Fattinger
PhD Student: Julian Amacker

Institutions: ZHAW School of Life Sciences and Facility Management, Institute of Computational Life Sciences (ICLS); Universitäts-Kinderspital Zürich, Abteilung Entwicklungspädiatrie

Projektpartner: TI Solutions AG; Ente Ospedaliero Cantonale (EOC) / Istituto di Neuroscienze Cliniche della Svizzera Italiana; Universität Zürich; Universitätsspital Zürich

• DIZH-Innovationsprogramm — ConCLAS (Digitalisierungsinitiative der Zürcher Hochschulen, seit 09/2024)