Apple AI Model Detects Health Conditions with 92% Accuracy: Study

A new Apple-supported study developed an AI foundation model, trained on behavioral data from wearables, that can detect health conditions with up to 92% accuracy — outperforming traditional sensor-based methods in several cases (via 9to5Mac).

The study, titled Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions, is based on data from the Apple Heart and Movement Study (AHMS). In a bid to show that behavioral data — like movement and sleep — can be a stronger health indicator than traditional biometrics such as heart rate or blood oxygen, researchers trained the Wearable Behavior Model (WBM) on over 2.5 billion hours of Apple Watch and iPhone data from 161,855 participants.

This is the latest in a growing list of studies commissioned by Apple as part of its push into the health space. Rather than relying on noisy raw sensor data like heart rate or ECG readings, researchers gave WBM 27 human-interpretable behavioral metrics to analyze. These included active energy, walking pace, heart rate variability, respiratory rate, and sleep duration — metrics that Apple Watch users generate daily.

What makes WBM stand out is its focus on longer-term behavioral trends, such as changes in gait or physical activity, which may offer better insight into certain health conditions than momentary sensor readings. As the study explains:

“Unlike raw sensors, these higher-level behavioral metrics are calculated using carefully validated algorithms derived from the raw sensors. These metrics are intentionally chosen by experts to align with physiologically relevant quantities and health states. Importantly, these data are sensitive to an individual’s behaviors, rather than being driven purely by physiology.”

In tests across 57 health-related tasks, WBM outperformed sensor-based models in many static health predictions (like hypertension or beta blocker use) and nearly all dynamic tasks — such as detecting pregnancy, sleep quality, or respiratory infections.

What’s even more impressive is that combining WBM with traditional PPG sensor data yielded the strongest results. This hybrid approach reached a remarkable 92% accuracy for pregnancy detection and delivered better outcomes for sleep tracking, detecting cardiovascular issues like atrial fibrillation, flagging injuries and infection, and more.

While it will take a long time for WBM to make its way to Apple’s products (provided the company chooses to integrate it in the first place), it’s clear that AI-powered health predictions are rapidly advancing and could soon play a larger role in your Apple Watch or iPhone.

Apple is rumoured to be working on an AI-powered health coach to offer personalized health advice, much like an actual doctor, and WBM could play a key role in that — assuming Apple can manage to get all its AI ducks in a row, that is.