Apple, in collaboration with researchers from the University of Southern California, has introduced a new AI model that leverages behavioral data from wearables.
Highlights
- New Model: Wearable Behaviour Model (WBM) uses daily behavioral data from Apple Watch wearers—such as sleep, mobility, and activity patterns—instead of just raw sensor inputs like heart rate or oxygen levels.
- Massive Dataset: Trained on 2.5+ billion hours of time-series data from over 162,000 Apple Watch users, enabling deep insights across diverse health prediction tasks.
- Superior Accuracy: WBM outperformed traditional sensor-only models in 39 out of 47 health prediction tasks, including chronic disease detection and acute condition monitoring.
- Custom AI Architecture: Apple developed a specialized foundation model with new tokenization methods to handle noisy, unstructured behavioral data over long time periods.
- Four Core Health Domains: The model focuses on sleep quality, daily mobility, cardiovascular fitness, and general activity levels using 27 behavior metrics.
- Hybrid Advantage: The best results came from combining behavioral signals with traditional sensor data (e.g., heart rate), proving both sources are valuable in health prediction.
- Limitations: The model was trained exclusively on U.S.-based Apple Watch users, which may limit global generalizability due to lifestyle and cultural differences.
- Equity Concerns: Reliance on premium wearable data may increase the risk of widening healthcare access disparities, especially in underserved regions.
The model, called the Wearable Behaviour Model (WBM), marks a shift in how health signals are predicted, focusing less on raw sensor readings and more on daily activity patterns.
Trained on more than 2.5 billion hours of data from over 162,000 Apple Watch users, WBM aims to provide more nuanced insights into personal health.
Behavioral Data Over Raw Sensor Signals
Traditionally, wearables have relied on physiological measurements like heart rate or blood oxygen levels to assess health. This study proposes an alternative approach—analyzing data such as step counts, sleep cycles, and mobility patterns.
These behavioral signals, while less direct, may offer a more holistic view of overall health, especially when tracked over long periods.
The challenge has always been that behavioral data is inherently noisy and harder to structure. To solve this, the researchers developed a custom foundation model architecture capable of processing large-scale time-series data.
The result is a system that outperformed sensor-only models in 39 out of 47 health-related prediction tasks.
Custom AI Architecture for Behavioral Health Analysis
According to Apple’s machine learning research team, the WBM was designed from the ground up for behavioral data. A key component of its architecture is a new method of tokenizing and structuring inputs, allowing the model to interpret activity trends over time.
WBM integrates 27 behavioral metrics grouped into four health domains:
- Sleep quality
- Daily mobility
- Cardiovascular fitness
- General activity levels
These metrics allowed the model to perform a wide range of predictions, including detection of chronic illnesses like diabetes and heart disease, as well as short-term conditions such as injury recovery or acute infections.
Sensor Data vs. Behavior
The researchers also tested a model trained only on photoplethysmogram (PPG) data, a sensor commonly used to track heart rate. While the PPG-based model performed well on some tasks, WBM consistently delivered more robust predictions across diverse scenarios.
A hybrid model combining both sensor data and behavioral signals produced the highest accuracy, indicating that each data type contributes unique strengths to health prediction.
Limitations and Considerations
The study notes several limitations that may affect the generalizability of the findings:
- The dataset was limited to Apple Watch users in the United States, raising concerns about its global relevance.
- Behavioral trends can vary significantly across cultures, environments, and lifestyles.
- Access to premium wearables remains limited in many parts of the world, potentially widening existing health disparities.
