The WBM model was trained using over 2.5 billion hours of wearable data collected from participants in the Apple Heart and Movement Study (AHMS), which included more than 160,000 volunteers. Researchers also created a dedicated pregnancy dataset using data from 430 pregnancies and included data from more than 24,000 women under the age of 50 who were not pregnant, to increase accuracy.
Rather than looking for short-term biometric spikes, the model tracks subtle, cumulative changes in behaviour over time. It uses an advanced AI architecture called Mamba-2 — well-suited for analysing time-series data like daily routines — to spot week-by-week physiological shifts that may indicate pregnancy, infections, or recovery from injury.
In the case of pregnancy detection, the AI identified high-level behavioural indicators such as altered gait, reduced mobility, and disrupted sleep as reliable early signals. Combined with biometric data like photoplethysmography (PPG), the WBM achieved pregnancy detection with 92% accuracy.
The findings suggest the Apple Watch and iPhone could one day play a more active role in reproductive health, potentially enabling non-invasive, early-stage pregnancy detection as a built-in feature. However, researchers stress that this approach is not intended to replace raw sensor data entirely. Instead, they advocate for a hybrid system, where behavioural trends provide a broader health context and traditional sensors capture real-time events.
Beyond pregnancy, the AI model showed promising results across 57 different health prediction tasks, including early detection of respiratory infections and insights into medication adherence, such as beta blocker usage.
As Apple continues to explore the vast potential of wearable data, the study highlights how behavioural AI could transform the Apple Watch into a more intelligent, proactive health companion.