Verified · Jul 10, 2026

Google 7/9 SensorFM: foundation model trained on 1T+ minutes of wearable sensor data from 5M people

2 sources

Google Research on 2026-07-09 published SensorFM (Xin Liu + Daniel McDuff): a foundation model for wearable health pre-trained on more than 1 trillion minutes of sensor signal from ~5 million consented people (100+ countries, all 50 US states, 20+ Fitbit and Pixel Watch models; data window 2024-09 → 2025-09); 5 sensor types + 34 one-minute aggregate features. Scaling curve spans 4 orders of magnitude. Largest variant (SensorFM-B): 31% lower reconstruction loss, +9% avg AUC, +21% Pearson, wins on 33 of 35 transfer tasks. Personal Health Agent evaluation: 31 participant profiles, 1,860 clinician ratings.

Why now

The 7/9 release lands alongside other July health-AI activity (OpenAI Health Intelligence 7/7, Anthropic Healthbench 7/2) — SensorFM is the first in this set to publish the **actual pre-training data-scale numbers.** Creators can ship today using 1T minutes / 5M people / 35 tasks as concrete numbers, no waiting a week.

Why it is worth publishing

Specific numbers (1T minutes, 5M people, 33/35 task wins) make for a clean information card. But this is medical-adjacent — keep a visible health disclaimer on screen at all times; never frame as 'AI health diagnosis.'

Evidence basis

Google Research official blog + specific data-scale numbers + July health-AI topic window. High-heat triple combo.

“Google just trained a model on 1 trillion minutes of body data — but it is not diagnosing anyone.”

Angle

Plug SensorFM into the July health-AI thread. Don't write a stand-alone 'wearable AI breakthrough' — this is transfer-benchmark results, not FDA approval, not clinical deployment, not medical-decision support. Lead with the health disclaimer on screen at 0:00.

Format

Short talking-head video

Demo idea

60-second talking-head: 0-15s — three specific numbers on screen (1T+ minutes, 5M people, 33/35 tasks won, with the line 'pre-training representation, not clinical deployment'); 15-50s — 'this is a transfer-benchmark result on 35 health prediction tasks, not an FDA approval, not a clinical diagnostic, not a treatment decision tool. The Personal Health Agent evaluation was a 31-participant clinician-rating study, not a clinical trial.' ; 50-60s — CTA ('subscribe, I am going to break down OpenAI Health Intelligence and Anthropic Healthbench next').

Platform notes

Medical-adjacent (high risk): never frame as 'AI medical diagnosis', 'health-screening tool', or equivalent clinical-action language. The 31-participant clinician-rating study is not a clinical trial. Any medical-content creator on this topic must surface a visible disclaimer: 'this content is not medical advice' — and, depending on jurisdiction, may need a regulatory-style disclosure statement.

Usable claims

Google Research's SensorFM was pre-trained on more than 1 trillion minutes of sensor data from ~5 million consented people (PPG, accelerometry, EDA, skin temperature, altimetry) and transferred to 35 health prediction tasks; SensorFM-B, the largest variant, won 33 of 35 transfer tasks and gained 9% average AUC plus 21% Pearson regression gain over the smallest variant.

Evidence pipeline

From the news

Google 7/9 SensorFM: a foundation model for wearable health pre-trained on 1T+ minutes from ~5M people

Breakdown

Walks the headline numbers Google Research published on 7/9 (Xin Liu and Daniel McDuff): 1T+ minutes of sensor data, 5M consented people, 33 of 35 transfer tasks won — these are transfer-benchmark numbers, not FDA approval, not clinical deployment. The article pairs each number with a single visible disclaimer so the medical-adjacent risk cannot be missed.

Sources

Risks

Frame as a research result (foundation model + transfer benchmarks). Never present as 'AI health diagnosis', 'medical screening tool', or equivalent. Avoid clinical-action language; emphasize that the Personal Health Agent evaluation was a 31-participant clinician-rating study, not a clinical trial.

Demo ideas

Make a 3-bullet information card: large text '1T+ minutes' / '5M people' / '33 of 35 tasks'; small text 'transfer benchmark — not a clinical device'.
Compare Google SensorFM (sensor data foundation model) vs OpenAI Health Intelligence (health LLM benchmark) vs Anthropic Healthbench (medical dialogue eval) — three 'July health-AI updates' on one slide, each with a 'data source + deployment stage + risk label' row.
For long-term health-AI coverage, build a 'data volume → task win rate → risk label' 3-column visualization across the three vendors. Show viewers in 10 seconds that health-AI progress is a gradient, not a single breakthrough.