What it was fed: a trillion minutes of body data

First, let's look at just how much data it consumed. This dataset comes from 5 million consenting users, spanning September 2024 to September 2025, covering over 100 countries including all 50 U.S. states.

This is by far the largest and most diverse batch of wearable data ever used to train a model. It brings together 34 per-minute aggregated features from five kinds of sensors.

Two of these abbreviations are worth explaining. PPG (photoplethysmography) is how a watch shines light against your skin and infers heartbeat and blood oxygen from how blood vessels absorb the changing light — this is also the main principle behind how smartwatches measure heart rate. EDA (electrodermal activity) measures conductance on the skin's surface; when a person gets nervous and sweats, conductance changes, so it's often used as an indirect signal of stress and emotion.

Watches drop out — how the model learns not to sweat it

One of SensorFM's most crucial design choices is how it handles "incomplete data." Watches get taken off to charge, go into power-saving mode, and sensors switch on and off. For wearable data, incompleteness is the norm, not the exception.

Trained this way, SensorFM is inherently "missingness-aware": it not only tolerates incomplete data, it can actually use the incompleteness to learn.

SensorFM is pretrained on over a trillion minutes of multimodal sensor data through "missingness-aware masked reconstruction." Source: Google Research

Scale it up, and performance keeps climbing — with no ceiling in sight

What happens when you scale up data and model size together? Google ran a set of experiments spanning four orders of magnitude: data volume from 2 million to 2 billion sensor-hours, model size from 100K to 100M parameters.

Pretraining loss falls steadily as data and model size grow, winning 33 of 35 tasks. More importantly: the performance curve shows no sign of saturating — meaning more data and more parameters will very likely make it even stronger.

Two metrics worth explaining here. AUC measures how accurate a classification judgment is (0.5 is equivalent to a coin flip, 1 is a perfect judgment); Pearson correlation measures how closely predicted values track real values (ranging from -1 to 1, with values closer to 1 meaning tighter tracking).

Scaling data volume and model capacity together by orders of magnitude, both pretraining and downstream task performance keep climbing, with no sign of saturation. Source: Google Research

Freeze the model, attach a linear head, and it's already good enough

There's an easy way to test whether a model's learned representation is actually useful: freeze the model itself with not a single parameter changed, attach a simple linear head (called a linear probe) onto its output physiological representation, and see if it can directly get the job done.

Tested on 35 discriminative health tasks (data from 3 IRB-approved prospective studies, 13,985 participants total), spanning six major categories of health signals.

34 of 35 tasks beat the traditional baseline (each dot is one task; the hollow one is the sole task not won):

Training only that lightweight linear head beat the "feature-engineering-plus-supervised-learning" traditional baseline on 34 tasks. Two more details worth noting: the larger the model, the better it implicitly captures demographic features like age and gender on its own (so the added benefit of separately feeding in that information keeps shrinking); and for conditions like depression and anxiety, whose signals are notoriously faint and hard to measure through sensors, it also picks up on them — beating the baseline with only a small number of labeled samples.

This means the same physiological representation can transfer across dozens of health prediction tasks, dramatically compressing the labeled data and development time needed to adapt to a new task.

Adding just a linear probe on top of the frozen SensorFM representation, relative performance gains over baseline across various discriminative health tasks. Source: Google Research

A room full of AI agents, racing to build it a prediction head

Turning a physiological representation into a predictor for a specific task traditionally requires manual feature engineering, architecture selection, and hyperparameter tuning — tedious work. Google simply had AI do this itself.

Two more patterns: solution quality improved monotonically as the search progressed; and switching to a stronger AI model (like a newer version of Gemini) produced better solutions.

AI agents in a "classroom" repeatedly write, test, and improve code to build prediction heads for SensorFM representations, with solutions steadily improving over time. Source: Google Research

Doctors couldn't tell AI's guess from real checkup results

Finally, Google put SensorFM into a "Personal Health Agent" and ran a tightly designed experiment: having it generate health summaries, then comparing quality across three input conditions.

How it worked: a panel of clinicians, under blind conditions, scored the summaries across five dimensions — context, relevance, plausibility, personalization, and potential harm — producing 1,860 expert ratings in total, consuming over 40 hours of expert evaluation time.

Two conclusions: adding SensorFM predictions raised scores above baseline on all five dimensions; and more importantly, clinician scores for summaries using SensorFM predictions versus real measurements showed no statistically significant difference. In other words, in this health-assistant scenario, the model's inferences are almost as useful as actual medical measurements.

Clinician ratings of generated responses improve after adding SensorFM as a tool for the personal health agent. Source: Google Research

What's still missing: privacy, clinical validation, and the correlation trap

Beyond these results, a few things need to be said clearly.

On July 9, 2026, Google Research released SensorFM, using over a trillion minutes of wearable data from 5 million people to train a general-purpose base that can make sense of the body's signals (corresponding paper arXiv:2605.22759).

Two key results: first, freezing the base and training only this one layer still beat the old approach on 34 of 35 health tasks; second, in a blind clinician evaluation (where doctors scoring the summaries didn't know what was behind each one), health summaries generated from its inferences scored no statistically significant difference from ones generated using real medical measurements. And as data and model size keep scaling up, performance keeps climbing with no ceiling in sight. But what makes it so reliable? The key is how it deals with "incomplete data."

Watches get taken off to charge, go into power-saving mode, and sensors switch on and off — for wearable data, incompleteness is the norm, not the exception. SensorFM's signature design, AIM (Adaptive Inheritance Masking), is built specifically to handle this.

The old approach either guesses to fill in the gaps (risking wrong guesses that introduce bias) or simply discards the incomplete segments (wasting data); AIM takes a third path, treating data genuinely dropped by the device and data deliberately masked during training as the same thing, so the model is inherently used to "incomplete information," treating "filling in the missing parts" itself as the skill to learn.

Ordinary people don't have a feel for "task win rate" — converting it to "how much effort it takes to onboard a new health task" is more intuitive.

One reminder: all figures above come from internal testing in Google's official blog post and its paper, and have not yet been independently reproduced by a third party.