Wearable foundation models: a brief history
A history of wearable foundation models tracing their evolution from early sensor data processing to modern transformer-based architectures that analyze heart rate and activity for personalized health insights.
Background
- Wearables (Apple Watch, Fitbit, Oura Ring) collect continuous health data—heart rate, sleep, activity—but today's analysis relies on simple, device-specific rules rather than general AI.
- "Foundation models" (like GPT) are large AI models trained on broad data, then adapted to many tasks. The article argues one foundation model pre-trained on sensor data could detect flu, monitor stress, predict ovulation, etc.—replacing today's one-algorithm-per-problem approach.
- Key challenge: wearable data is messy—different sampling rates, noise, and no clean labels ("this heart rate pattern = early COVID"). Researchers are tackling this with self-supervised learning that doesn't need labeled data.
- Why it matters: success could turn your watch into a general-purpose health monitor catching early signs of infection, metabolic shifts, or cardiac issues—no separate app or algorithm needed. The field is early; no production model exists yet.