The Work
Carter built Hypnos, a foundation model for overnight physiological signals (EEG, EOG, ECG, respiratory) that applies next-token prediction, the technique behind LLM pretraining, to sensor data. Predicting the next second of signal yields an embedding space where physiological states such as sleep stages become linearly separable without labels. This approach, which uses minimal sensor-specific design biases, can accelerate the development of diagnostic algorithms using novel sensor hardware.
What Worked
Hypnos outperformed existing approaches across a range of downstream tasks, e.g. beating SleepFM by 15% AUPRC and reaching expert-level sleep staging with 100x less the labelled data. The model was open-sourced and had over 1,000 downloads in its first month.
What Was Learned
A collaborative AI co-scientist approach significantly accelerated multiple aspects of the project. However, a fully autonomous research loop was susceptible to wandering into ineffective dead-ends; humans still have better research taste.
About
Jonathan holds a DPhil in Machine Learning from Oxford, where he was supervised by Prof. Lord (Lionel) Tarassenko. During his DPhil, he designed, built, and open-sourced world-leading deep learning models for quantitative sleep monitoring from wearables. These are currently being evaluated in clinical studies for the longitudinal monitoring of linked health conditions including insomnia and depression. Alongside his DPhil, he has spent several years working in industry as an AI researcher/engineer at HealthTech companies LIO and Sanome. There he helped to develop several AI-enabled medical devices which have been deployed in healthcare systems in the UK and US. (Personal website: https://joml.io)
