SleepLong: Towards Generating Long-Sequence Sleep Heart Rate Signals with Conditional Diffusion

Sleep-stage classification is a critical step in assessing sleep quality. Wearable sleep trackers offer a promising solution for long-term monitoring outside traditional clinical settings. Most wearable sleep trackers are heart-rate-based, but their effectiveness is limited by shortage of good-quality publicly available data. To address this, diffusion models offer a privacy-aware approach to generate data for augmentation and to train classification models. Existing generation methods typically focus on individual sleep stages in isolation, without modeling the dependencies and continuity across stages. This paper explores a spectrogram-based diffusion model to generate a long range sleep heart-rate sequence conditioned on sleep-stage labels (hypnogram), as opposed to generating the individual stages in isolation. We verify the effectiveness of the approach in sleep-stage classification tasks using two publicly available datasets, HMC and DREAMT.