Most wearable health projects try to predict one outcome at a time. That works until you need to track dozens of signals like heart rate variability, sleep stages, stress, glucose risk, activity levels and more. Each new endpoint requires fresh labels, and labeling thousands of person‑days of sensor data is expensive and often impossible after the fact.
Google Research’s SensorFM changes the game. It is a foundation model built from more than one trillion minutes of raw sensor data collected from five million people wearing PPG, accelerometer, EDA, skin temperature and altimeter streams. The model uses a ViT‑1D encoder trained with a masked auto‑encoder objective, so it learns a general representation of wearable time‑series.
Because real‑world streams have gaps from charging, off‑wrist moments or power‑saving modes, SensorFM does not rely on naïve imputation or window dropping. Its Adaptive and Inherited Masking (AIM) treats missingness as a signal, letting the decoder reconstruct ablated patches and turning imputation and forecasting into free by‑products.
In practice you keep the encoder frozen, aggregate daily embeddings per person (mean and std to 50 components), and attach a lightweight linear head. Person‑independent five‑fold cross‑validation then yields strong results across 35 downstream tasks without retraining the backbone. The approach beats feature‑engineered baselines on most metrics and works even when labels are scarce; you can first compare against a demographics‑only baseline.
Use cases include: risk‑based screening that flags participants for confirmatory labs, repairing daily summaries when 60 minutes of data are missing, enabling label‑scarce studies by probing frozen embeddings, and powering grounded coaching apps that translate predictions into qualitative guidance rather than raw numbers.
The model scales predictably: larger variants trained on proportionally more data keep improving reconstruction loss, AUC and correlation, showing the trend has not yet saturated.
#AI #WearableTech #HealthAI #FoundationModel #SensorFM #Product