Apple Watch Data Powers AI That predicts Medical Conditions with High Accuracy
Table of Contents
A new study demonstrates the potential of artificial intelligence to unlock life-saving insights from everyday wearable data, achieving notable accuracy in predicting medical conditions using data from 3 million Apple Watch “person-days.” The research, conducted by MIT and Empirical Health, offers a promising path toward proactive healthcare and early disease detection.
The Rise of “World Models” in AI
the breakthrough hinges on a novel AI architecture called the Joint-Embedding Predictive Architecture (JEPA), initially proposed by Yann LeCun while he was Chief AI Scientist at Meta.Unlike customary AI systems that focus on reconstructing input data, JEPA learns to predict future states based on an internal “world model.”
This approach is a cornerstone of a growing field exploring “world models,” a departure from the token-prediction methods used in large language models (LLMs) like GPT. LeCun, who recently left Meta to found a company dedicated to world models, believes this technology represents the true path to Artificial General Intelligence (AGI).
JETS: Adapting JEPA for Wearable Health Data
The study, titled JETS: A Self-Supervised Joint Embedding Time Series Foundation Model for Behavioral Data in Healthcare and recently accepted to a workshop at NeurIPS, adapts JEPA to handle the unique challenges of wearable data. This data is frequently enough “irregular multivariate time-series,” meaning measurements like heart rate,sleep patterns,and activity levels are recorded inconsistently and with gaps over time.
Researchers analyzed data from 16,522 individuals, totaling approximately 3 million person-days, with 63 distinct metrics tracked across five key areas: cardiovascular health, respiratory health, sleep, physical activity, and general statistics. A notable hurdle was the limited availability of labeled medical histories – only 15% of participants had this facts.
“Interestingly, only 15% of participants had labeled medical histories for evaluation, which means that 85% of the data would have been unusable in traditional supervised learning approaches,” the study noted. To overcome this, the team employed a self-supervised pre-training approach, allowing JETS to learn from the complete dataset before being fine-tuned on the smaller labeled subset.
Impressive Prediction Accuracy
The researchers converted each data point into a “token” – representing the day, value, and metric type – and then used a masking process to train the model to predict missing information. when tested against baseline models, JETS demonstrated significant advantages in predicting several conditions.
Specifically, JETS achieved an Area Under the Receiver Operating Characteristic curve (AUROC) of 86.8% for high blood pressure, 70.5% for atrial flutter, 81% for chronic fatigue syndrome, and 86.8% for sick sinus syndrome. While not always the top performer across all conditions, the results clearly indicate the model’s potential.
It’s significant to note that AUROC and Area Under the Precision-Recall Curve (AUPRC) measure a model’s ability to rank likely cases, not simply its overall accuracy. This means JETS excels at identifying individuals who are most at risk, even if it doesn’t always correctly predict every case.
The Future of Proactive Healthcare
This study underscores the immense potential of leveraging data already collected by wearable devices like the Apple Watch, even when usage isn’t consistent. Health metrics were recorded as infrequently as 0.4% of the time for some individuals, yet JETS was still able to extract meaningful insights.
“All in all, this study presents an interesting approach to maximizing the insight and life-saving potential of data that could be written off as incomplete or irregular,” a leading AI analyst commented. The research reinforces the idea that innovative models and training techniques can unlock valuable information from the vast amounts of data generated by everyday wearables, paving the way for a future of more proactive and personalized healthcare.
You can read the full study here.
