A robust PPG foundation model using multimodal physiological supervision

ICLR 2026 Conference Submission19085 Authors

Keywords: Photoplethysmography (PPG), health, ubiquitous computing, foundation model, wearables, representation learning, multimodal, self-supervised learning, time series, physiology

Abstract: Photoplethysmography (PPG), a non-invasive measure of changes in blood volume, is widely used in both wearable devices and clinical settings. Although recent work has explored PPG foundation models using large-scale intensive care unit (ICU) datasets, these efforts often assume the need for clean and high-quality signals. In contrast, we argue that the inherent noise and variability in ICU datasets can be harnessed to build more robust and generalizable representations. To address this, we propose a PPG foundation model that leverages accompanying electrocardiogram and respiratory signals in ICU datasets to select contrastive samples during pretraining. Our approach allows the model to retain and learn from noisy PPG segments, improving robustness without requiring multimodal inputs at inference. Our model, pretrained on 3x fewer subjects than existing state-of-the-art approaches, achieves performance improvements of up to 36\% in classification and 42\% in regression on 14 out of 15 diverse downstream tasks, including stress and heart rate prediction. Our results demonstrate that multimodal supervision can leverage clinical data to enable the development of robust, unimodal foundation models for both clinical and consumer-level data.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 19085