What is it about?
NormWear is a novel foundation model specifically designed for multivariate wearable sensing of physiological signals, addressing the challenges of diverse and heterogeneous sensor data (e.g., ECG, PPG, EEG, GSR, IMU). It introduces a unified architecture that learns generalizable representations across different sensing configurations and health‑related tasks by pretraining on millions of wearable signal segments and leveraging advanced tokenization and cross‑sensor attention mechanisms.
Featured Image
Photo by Shawn Rain on Unsplash
Why is it important?
Wearable devices generate massive, varied physiological time series, yet existing models often struggle to generalize across modalities or sensor setups. NormWear fills this gap by enabling robust performance across 18 diverse downstream tasks, from mental health inference to vital sign estimation, and supports zero‑shot adaptation to unseen applications. This capability makes it a significant step toward scalable, real‑world digital health analytics.
Perspectives
By open‑sourcing the model and demonstrating broad applicability, NormWear aims to empower both the research and developer communities to build more flexible and generalizable wearable‑based solutions. Its design supports innovation in health monitoring, cross‑device analytics, and efficient transfer learning, potentially accelerating progress in digital health, personalized medicine, and ubiquitous sensing.
Yunfei Luo
University of California San Diego
NormWear is a foundation model for multimodal physiological timeseries data which is input (sensor modality and type) agnostic and can solve a wide variety of health related task with text prompt (i.e., questions). The wavelet based preprocessing, CLS token-based attention and the representation alignment between physiological signal and text prompt is some of the secret sauce behind the success of this modeling approach. The paper, code, data and associated documentations are open source and we hope that this model will help to push this area forward.
Tauhidur Rahman
University of California San Diego
Read the Original
This page is a summary of: Toward Foundation Model for Multivariate Wearable Sensing of Physiological Signals, ACM Transactions on Computing for Healthcare, March 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3803808.
You can read the full text:
Resources
Contributors
The following have contributed to this page
