Abstract

Methods: We propose a multimodal framework that combines encoders for PPG, inertial measurements, and temperature signals with a learnable state-space model for inter-beat inference. The state-space dynamics adapt to non-linear changes and PAT-related shifts. A trust gate uses predicted uncertainty to down-weight corrupted intervals.

Results: Using a single model configuration across three public datasets (DaLiA, WildPPG, BIDMC), our method consistently improves inter-beat interval accuracy and HRV indices compared to prior work. For SDNN, we reduce error by up to 80\% relative to traditional peak detection, while improving agreement with ECG-derived references.

Conclusion: Uncertainty-aware multimodal observations with an adaptive state-space model (SSM) yield robust HRV estimation under real-world artifacts.

Significance: Our method enables robust HRV monitoring in realistic settings from common wearable sensors and provides strong baselines and results to support research and future applications.

Code: github.com/eth-siplab/Continuous_HRV_from_PPG