Google's SensorFM matches ground-truth health labels across 34 of 35 wearable tasks
Google Research's SensorFM, a foundation model pre-trained on over a trillion minutes of wearable sensor data from 5 million consented people, beat supervised baselines on 34 of 35 health-prediction tasks across cardiovascular, metabolic, and sleep. The result that lands: blinded clinicians scoring 1,860 summaries found no statistically significant difference between grounding a Personal Health Agent in SensorFM's inferences and in real ground-truth measurements, putting one generalist physiology model in place of the bespoke, single-outcome pipelines wearable health has relied on.
Source: research.google ↗
there was no statistically significant difference between grounding the agent in SensorFM predictions versus actual ground-truth measurements
Why this matters
- → One foundation model replaces dozens of bespoke wearable health pipelines.
- → Clinicians found SensorFM inferences as reliable as actual ground-truth measurements.
- → Trained on trillion minutes of real-world fragmented sensor data, not curated labels.