Apple Watch becomes a tool for disease prediction thanks to AI Apfelpatient

The Apple Watch has been providing a wealth of health data for years, but only now is its full potential becoming apparent. A new study by MIT and Empirical Health uses approximately 3 million person-days of Apple Watch data to develop a basic AI model that reliably predicts various diseases. The results clearly demonstrate that everyday wearable data is more valuable for medical research than previously thought.

Wearable data is often considered incomplete, irregular, or difficult to interpret. The Apple Watch isn't always worn, many measurements are incomplete, and different sensors deliver data at widely varying intervals. This is precisely where the study comes in. Instead of excluding this data, it demonstrates that even significant gaps can be utilized if AI learns to deduce the meaning of the missing segments from the existing context. This opens a new avenue for identifying patterns in real-world health data without requiring complete records.

Background: From JEPA to a time series model

The foundation is the Joint-Embedding Predictive Architecture (JEPA), a concept developed by Yann LeCun during his time as Chief AI Scientist at Meta. JEPA pursues the idea of not reconstructing missing data, but rather estimating its meaning. In the case of images, this means that a model should not guess exactly what lies behind an obscured area, but rather how that area fits into the visible context.

This principle was applied to time series data, such as that provided by the Apple Watch. Heart rate, activity, sleep stages, and respiration data are recorded irregularly. The JEPA concept was further developed precisely to address such gaps.

LeCun's approach is now considered the starting point for so-called world models, which aim not only to make predictions but also to develop an understanding of how systems function. Meta published the I-JEPA model in 2023, and LeCun has since founded his own company that focuses entirely on such world models.

The study: 3 million days of Apple Watch data

The current work, titled „JETS: A Self-Supervised Joint Embedding Time Series Foundation Model for Behavioral Data in Healthcare,“ has been accepted for a workshop at NeurIPS. The JETS model utilizes the JEPA concept for irregular, multivariate time series, such as those commonly found in wearables.

The study analyzed data from 16,522 individuals, totaling approximately 3 million person-days. The Apple Watch provided 63 different metrics across five areas:

Cardiovascular health
Respiratory health
Sleep
Physical activity
General Statistics

It is noteworthy that only 15 of the participants had a documented medical history. Typically, 85 of the data would therefore have been unusable for a classic, supervised AI model. However, JETS uses self-supervised pre-training. The model initially learns from all data, even unlabeled data, and is only then refined with the available diagnoses.

How JETS learns from Apple Watch data

For each measurement, a data triplet was created, consisting of day, value, and metric type. These triplets were processed as tokens, some of which were masked. The model then learned to estimate the significance of the missing segments. This creates a common embedding space in which all time series are comparable—regardless of the number of gaps.

The model's quality was assessed using AUROC and AUPRC. These metrics measure how well an AI can distinguish between positive and negative cases, not how often it is correct. The results:

High blood pressure: 86.8 % AUROC
Atrial flutter: 70.5 %
Chronic Fatigue Syndrome: 81 %
Sick sinus syndrome: 86.8 %

JETS outperformed several base models, including an older, transformer-based version. Particularly striking was the fact that some values were available for only 0.4 days, while others were available for 99 days. Despite this, the model was able to extract usable patterns from both situations.

Why the results are important for the Apple Watch

The study shows that the Apple Watch can do much more than simply document individual measurements in everyday life. Even with incomplete data, modern models can identify correlations that were previously hidden. Wearables thus provide not just snapshots, but valuable long-term information that can help identify risks and potential illnesses earlier.

Potential of everyday Apple Watch data

Research surrounding JEPA, world models, and the JETS model reveals the immense potential of the data collected daily by the Apple Watch. Even irregular time series contain sufficient structure to enable reliable predictions. The study demonstrates that modern AI architectures can enhance wearable data, transforming it into a tool that contributes to improved disease detection in the long term. This brings us closer to a medicine that recognizes warning signs earlier and understands health trends more precisely—simply through the data generated in everyday life. (Image: Wongphoto / DepositPhotos.com)