Z. Bao, Y. Benezeth, F. Yang, Y. Zhang, H. Wang, C. Li, "TF-MambaNet: A Temporal and Frequency Fused Bidirectional Mamba Architecture for PPG Foundation Model", IEEE ICASSP 2026 (pdf)
Photoplethysmography (PPG) is the most widely used non-invasive physiological signal monitoring technology. However, most existing deep learning models based on PPG are designed for specific tasks and lack cross-scenario generalization capability.
This paper proposes TF-MambaNet, a foundation model for PPG signals. The model is pre-trained on 50,000 hours of real-world PPG data collected from 100 users in real-world scenarios. It adopts a lightweight bidirectional Mamba architecture that integrates temporal and frequency domain information, combined with a novel multi-scale hybrid loss function.
This design enables the model to capture global trends, local patterns, and amplitude dynamics of blood volume pulse (BVP) signals. TF-MambaNet obtains competitive results while maintaining an extremely small parameter size (1.5M parameters), making it directly applicable as a feature extractor for personalized health monitoring.
You can download the dataset here (available very soon) and code here.
If you use this dataset, please cite this paper:
Zhanpeng Bao, Yannick Benezeth, Fan Yang, Yonglin Zhang, Haibin Wang, Chao Li
TF-MambaNet: A Temporal and Frequency Fused Bidirectional Mamba Architecture for PPG Foundation Model
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2026.