BROWSE
Detail
Conference Paper
A Multimodal Framework for Variable-Length Sleep Staging in Wearable Applications
Cited - time in 
Share 
Share
- Authors
- Hyeonjeong Lee, Minseong Kim, Hwi-Gang Kim, Borum Nam, Kwang-Yong Kim
- Issue Date
- 2026-07
- Citation
- International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) 2026, pp.5574-5579
- Publisher
- IEEE
- Language
- English
- Type
- Conference Paper
- Abstract
- Sleep staging using photoplethysmography (PPG) enables an unobtrusive solution for continuous sleep monitoring in daily life. However, existing PPG-based models typically rely on fixed, overnight-long inputs to capture temporal context, which limits their flexibility for short-term recordings and increases computational overhead. Moreover, the optimal context length required for different wearable-derived modalities remains underexplored. In this work, we propose Flexi-VNet (Flexible-length Vital-sign Network), a multimodal framework to support variable-length sequences. We incorporated a redesigned architecture of SleepPPG-Net with attention-based temporal aggregation and positional encodings to ensure stable context learning across diverse scales. To better capture both transient dynamics and long-term physiological trends, hybrid modality-specific encoders were employed that jointly integrate waveforms and statistical features for HR and SpO2. Evaluated on the MESA dataset, our results reveal that while PPG achieves reliable performance with only a few hours of data, HR and SpO2 require longer contexts due to their slower physiological changes. Notably, the multimodal fusion significantly outperforms PPG-only models, particularly in deep sleep classification. We also demonstrate that the hybrid HR-SpO2 configuration provides a promising alternative for independent sleep staging when PPG signals are unavailable. By validating feasibility across various time scales, this work provides an actionable foundation for wearable sleep monitoring systems.
- Keyword
- sleep stage classification, photoplethysmography, deep learning, heart rate, oxygen saturation