Self-supervised Learning for Incomplete Multimodal Wearable Sensor Data

ICLR 2026 Conference Submission21182 Authors

Keywords: Health, Foundation Model, Masking, Wearables, Sensors, Self-supervised Learning, mHealth, Mobile Health, Biosensor, PPG, Acceleroemtry

TL;DR: AIM learns from incomplete wearable data via inherited masking, eliminating imputation needs while outperforming SOTA on classification, regression & generation.

Abstract: Foundation models, a cornerstone of recent advancements in machine learning, have predominantly thrived on complete and well-structured data. However, wearable sensor data frequently suffers from significant missingness, posing a substantial challenge for the training of generalist models in this domain. This paper introduces Adaptive and Inherited Masking (AIM), a novel self-supervised learning (SSL) approach that learns robust representations directly from incomplete data without requiring explicit imputation. Leveraging AIM, we develop AIM_FM, a foundation model pre-trained on 40 million hours of fragmented multimodal wearable sensor data. We find that with AIM this model exhibits improved scaling and performance across a diverse range of tasks as compared to current state-of-the-art wearable-sensor foundation models trained on imputed data. Critically, AIM_FM maintains high performance even under targeted missingness scenarios (e.g., absent sensors, contiguous missingness). We will release our metabolic study dataset with reproducible training+evaluation code.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 21182