In a collaborative effort with the University of Southern California, Apple researchers have unveiled a groundbreaking artificial intelligence (AI) model designed to monitor behavioral data derived from sensor signals. This new study builds on insights from the Apple Heart and Movement Study (AHMS) and explores whether behavioral metrics, including sleep patterns and step counts, can be more effective indicators of individual health than traditional measures such as heart rate and blood oxygen saturation. The findings indicate that the AI model performed well overall, despite some limitations.
New Apple Study Shows Benefits of Moving Beyond Traditional Health Data
The research paper, titled “Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions,” has been published in the pre-print journal arXiv and awaits peer review. The authors aimed to create an AI model referred to as the Wearable Behaviour Model (WBM), which processes behavioral data collected by wearables, including sleep duration, REM cycles, daily step counts, and variations in activity patterns throughout the week.
Traditionally, wearable health research has focused on raw sensor data—like continuous heart rate and blood oxygen monitoring—when assessing health. Although these metrics can sometimes provide insight, they often fail to capture the complete picture of an individual’s health and can present inconsistencies.
Notably, behavioral data, which many wearables track, has not been leveraged effectively as a dependable indicator of health, due to two primary challenges identified in the study. First, the volume of behavioral data is significantly larger than that of sensor data, which can introduce considerable noise. Second, developing algorithms capable of collecting and analyzing this data to reliably predict health outcomes presents a substantial challenge.
This is where a large language model (LLM) plays a crucial role in addressing data analysis issues. To minimize the noise in the data, researchers trained the model using structured and processed inputs. These inputs were derived from over 162,000 Apple Watch users who participated in the AHMS, generating more than 2.5 billion hours of wearable data.
Once the model was trained, it utilized 27 distinct behavioral metrics, categorized into areas such as activity levels, cardiovascular health, sleep quality, and mobility. The model was tested across 57 health-related scenarios, which included determining the presence of specific medical conditions, such as diabetes and heart disease, and monitoring acute health changes like recovery from injuries or infections. The WBM outperformed baseline accuracy in 39 out of 47 assessments, according to the researchers.
Comparison between performance of the WBM model the test model and the combination of both
Photo Credit: Apple
Further analysis compared the model’s findings with another test model that utilized only raw heart data, known as photoplethysmogram (PPG) data. When assessed separately, neither model demonstrated a definitive superiority. However, the combination of both models yielded enhanced accuracy in predictions and health assessments.
The researchers propose that integrating traditional sensor data with behavioral metrics may significantly improve the prediction of health conditions. They noted that behavioral data are often easier to interpret, align more closely with real-world health outcomes, and suffer less from technical errors.
The study also acknowledges significant limitations, as the data were solely derived from Apple Watch users in the United States, leaving out representation of the global demographic. Furthermore, the cost of wearable devices that can accurately gather and maintain behavioral data poses a challenge to accessibility in preventive healthcare.
