Google’s SensorFM turns messy wearable sensor data into a general-purpose health intelligence layer

2026-07-13 · Source: The Decoder · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Medical Devices & Health Technology · Depth: Advanced, medium

Summary

Google Research has introduced SensorFM, a foundation model designed to create a general representation of human physiology and behavior from wearable sensor data. Trained on over one trillion minutes of unlabeled multimodal data from five million Fitbit and Pixel Watch users across 100 countries and 20 device models, SensorFM processes 34 features from five sensor types. The model, with variants ranging from 100,000 to 100 million parameters, demonstrated improved performance with increased data and size. It outperformed supervised baselines using hand-crafted features on 34 out of 35 health and behavioral tasks in tests involving 13,985 participants. Furthermore, health summaries augmented with SensorFM's predictions were rated significantly higher by four clinicians across five dimensions compared to a baseline, indicating its potential to enhance AI health agents. SensorFM is currently a research model.

Key takeaway

For AI Scientists and Machine Learning Engineers developing health applications, SensorFM demonstrates the power of foundation models for wearable data. You should consider self-supervised pretraining on large, unlabeled datasets to overcome the limitations of siloed, task-specific models. This approach can significantly improve performance and label efficiency for diverse health and behavioral predictions, potentially enhancing personalized AI health assistants.

Key insights

SensorFM uses massive unlabeled wearable data to create a general foundation model for diverse health and behavioral tasks.

Principles

• Scaling model size and data improves performance.
• Self-supervised pretraining enhances label efficiency.
• Foundation models can unify siloed health features.

Method

SensorFM is trained self-supervised using "Adaptive and Inherited Masking" (AIM) to reconstruct masked data segments, learning to handle data gaps.

In practice

• Integrate SensorFM predictions into health agents for richer context.
• Use SensorFM's representations for new tasks with few labels.

Topics

• SensorFM
• Foundation Models
• Wearable Data
• Health Intelligence
• Self-supervised Learning
• AI Health Agents

Best for: AI Scientist, Machine Learning Engineer, Research Scientist

Related on AIssential

• Towards a General Intelligence and Interface for Wearable Health Data — Large-scale pretraining on wearable data creates a foundation model for diverse health insights.
• Physical activities enable scalable foundation modelling for broad-spectrum health prediction — StepFM uses low-dimensional step data to create a scalable, privacy-preserving foundation model for broad-spectrum health prediction.
• A foundation model for electrodermal activity data — UME is the first open-source foundation model for EDA, outperforming baselines with 20x less computation.
• Two Stanford grads raise $11M to build a noninvasive wearable for hormone tracking — Clair Health offers a noninvasive wearable and AI for continuous, multi-biomarker hormone tracking to empower women's health understanding.
• SL-S4Wave: Self-Supervised Learning of Physiological Waveforms with Structured State Space Models — SL-S4Wave uses tailored S4 models and contrastive learning for robust, label-efficient physiological waveform analysis.
• Self-supervision drives representational convergence in medical foundation models more than clinical supervision — Self-supervised objectives, not scale or clinical supervision, primarily drive representational convergence in medical foundation models.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by The Decoder.