NormWear

NormWear is a multimodal foundation model for wearable physiological sensing, designed to extract generalizable representations from the heterogeneous time-series signals produced by consumer and clinical wearables. Wearable devices capture diverse modalities — photoplethysmography (PPG), electrocardiography (ECG), electroencephalography (EEG), galvanic skin response (GSR), and inertial measurement (IMU) — but real-world data is plagued by variability in sensor placement, sampling rate, channel availability, and device configuration. Most prior models were trained narrowly for a single signal type or task, limiting transfer. NormWear targets this fragmentation by learning a shared representation space that is compatible with arbitrary combinations of sensors and channels.

Developed by Yunfei Luo, Yuliang Chen, Asif Salekin, and Tauhidur Rahman at the MOSAIC mobile-sensing lab at the University of California, San Diego, NormWear was first released as a preprint in December 2024 and has since been accepted (in press, 2026) at ACM Transactions on Computing for Healthcare. It is presented as the first general-purpose foundation model spanning this breadth of wearable modalities, positioning it alongside emerging biosignal foundation models while emphasizing cross-sensor generality rather than single-modality specialization.

The central technical contribution is a channel-aware attention mechanism paired with a shared liaison [CLS] token, which lets the model reason both within individual sensor channels and across multiple sensors simultaneously, regardless of how many channels a given downstream dataset provides.

#Key Features

• Cross-modal generality: A single pretrained encoder handles PPG, ECG, EEG, GSR, and IMU signals, accommodating variable channel counts and sensor configurations without per-modality retraining.
• Channel-aware attention: A shared liaison [CLS] token mediates attention within each sensor channel and across channels, capturing both intra-sensor temporal structure and inter-sensor relationships.
• Self-supervised pretraining: The model is pretrained on large-scale public physiological datasets using a masked-reconstruction objective with configurable time/frequency masking schemes, requiring no task labels.
• Time-frequency representation: Continuous Wavelet Transform (CWT) scalograms feed the encoder, exposing multi-scale frequency structure that is informative for periodic biosignals like PPG and ECG.
• Flexible evaluation regimes: NormWear supports zero-shot inference via signal-text alignment, plus partial-shot and full-shot fine-tuning for both classification and regression tasks.

#Technical Details

NormWear is a transformer-based encoder (with a decoder used during masked pretraining) of roughly 0.2B parameters. Inputs are converted to CWT scalograms and tokenized into patches, producing 768-dimensional patch embeddings with on the order of 365 patches per channel. Pretraining uses masked signal reconstruction as the self-supervised objective, with random or structured masking applied across time and frequency dimensions. The channel-aware attention layers operate over the per-channel patch sequences while the shared [CLS] token aggregates information across channels, enabling the architecture to scale to whatever sensor set a downstream dataset exposes. The model was benchmarked across 11 public wearable sensing datasets spanning 18 downstream applications in four domains — mental health, body-state inference, vital-sign estimation, and disease-risk evaluation — under zero-shot, partial-shot, and full-shot protocols, where it reported consistent improvements over state-of-the-art baselines.

#Applications

NormWear is intended for researchers and developers building health-inference pipelines on wearable data. Its representations support mental-health assessment, body-state inference (such as activity or stress states), vital-sign estimation (for example heart-rate or respiration-related targets), and disease-risk evaluation. Because the model accepts heterogeneous sensor inputs and supports zero-shot text-aligned classification, it lowers the barrier for teams with small or label-scarce datasets to bootstrap a task without training from scratch, and it provides a common backbone for benchmarking new wearable health applications.

#Impact

NormWear advances the case for general-purpose foundation models in the biosignal domain, where most prior work was siloed by signal type. By demonstrating that one channel-aware encoder can transfer across five modalities and 18 tasks, it offers a reusable starting point for the mobile-health community. The code is released under Apache-2.0 with pretrained weights distributed via Hugging Face and a GitHub release, supporting reproduction and extension. As a recent release its long-term adoption is still emerging, and reported gains depend on the specific dataset and evaluation regime; broader independent validation across devices and populations will determine how well its generality holds in deployment.