Google SensorFM — The AI Foundation Model For Wearables.

Google Research has unveiled SensorFM, a foundation model for wearables that’s been trained on a dataset comprising 1 trillion minutes of data collected by Fitbit and Pixel Watch devices worn by 5 million people in over 100 countries. SensorFM comes in several sizes between 100,000 and 100 million parameters.

SensorFM beat bespoke machine learning algorithms at 34 out of 35 health-related predictive tasks such as estimating heart rate and skin conductance, as well as behavioral metrics. SensorFM was evaluated on data from 13,985 participants in three separate studies.

It’s based on the idea that rather than having dozens of algorithms dedicated to each specific metric, there should be a single model that can capture the entirety of the physiological system. The reconstruction error of the largest version of SensorFM is 31% lower than that of the smallest.

Google has already applied SensorFM to its concept of an AI-powered personal doctor, where the use of the model’s predictions in a medical summary achieved “statistically significant” gains over the baseline in all five categories (context, personalization, reasoning, relevance, safety). The company hasn’t announced any plans to turn SensorFM into a commercial product; it’s strictly a research project.