Apple’s AI model detects health conditions with 92% accuracy using behavior data

Apple researchers have developed a new AI model called WBM (Wearable Behavior Model) that can detect health conditions with up to 92% accuracy by analyzing behavioral data from wearables rather than raw sensor readings. The breakthrough suggests that movement patterns, sleep habits, and exercise data may be more reliable health indicators than traditional biometric measurements like heart rate or blood oxygen levels.

What you should know: The WBM model was trained on over 2.5 billion hours of data from Apple Watch and iPhone users, focusing on 27 behavioral metrics rather than raw sensor streams.

• The model analyzes higher-level behavioral patterns like step count, gait stability, mobility, and VO₂ max—all metrics that Apple Watch already produces.
• Unlike previous health AI models that relied on raw sensor data like photoplethysmograph (PPG, the heart rate sensor) or electrocardiograph (ECG) readings, WBM learns from processed behavioral information that reflects real-world health trends.
• The research comes from Apple’s Heart and Movement Study (AHMS), involving 161,855 participants.

The big picture: Behavioral data from wearables may be more meaningful for health detection than the raw sensor information most AI models currently use.

• “Consumer wearables, such as smartwatches and fitness trackers, provide rich information across diverse health domains,” the researchers explained.
• The study argues that behavioral metrics are “intentionally chosen by experts to align with physiologically relevant quantities and health states” and are “sensitive to an individual’s behaviors, rather than being driven purely by physiology.”
• This approach proves particularly effective for detecting both static health conditions (like hypertension history) and transient states (like pregnancy or sleep quality).

Key performance metrics: WBM outperformed traditional PPG-based models across most health detection tasks when evaluated on 57 different health-related challenges.

• The model excelled in 18 of 47 static health prediction tasks, including detecting whether someone takes beta blockers.
• It succeeded in all but one dynamic health detection task, with diabetes being the sole exception where PPG sensors proved superior.
• When WBM and PPG data were combined, the hybrid model achieved 92% accuracy for pregnancy detection and showed consistent improvements in sleep quality, infection, injury, and cardiovascular tasks like atrial fibrillation detection.

How it works: The model uses a new architecture built on Mamba-2, which the researchers found performs better than traditional Transformers for this specific use case.

• Data is processed in weekly blocks rather than the second-by-second intervals typical of raw sensor analysis.
• The model focuses on patterns in processed behavioral data that highlight meaningful health events rather than getting overwhelmed by noisy raw sensor signals.
• “Mobility metrics that characterize walking gait and overall activity levels may be important behavioral factors to help detect a changing health state such as pregnancy,” the researchers noted.

Why this matters: The research suggests a complementary approach to health monitoring that could significantly improve early detection of health changes.

• Rather than replacing sensor data, WBM captures long-range behavioral signals while PPG catches short-term physiological changes.
• The combination approach could enable more accurate and earlier detection of meaningful health shifts, potentially transforming how wearable devices monitor and predict health conditions.
• This research positions Apple’s health monitoring capabilities as increasingly sophisticated, moving beyond simple biometric tracking toward comprehensive behavioral health analysis.