Apple Watch data usage of about 3 million person-days for a new successful study by researchers from MIT and Empirical Health. The Apple Watch data usage was to develop a foundation model that predicts medical conditions with splendid accuracy even when that data is incomplete or irregular.
This is a Self-Supervised Joint Embedding Time Series Foundation Model for Behavioral Data in Healthcare. The study was recently accepted for presentation at a NeurIPS workshop, a leading AI research conference.
Apple Watch for Health Research
The team of researchers used a large dataset of around 3 million person-days of Apple Watch, Fitbit/Pixel Watch, and Samsung data from more than 16,000 individuals to build the model. The model was named ‘JETS’ (Joint-Embedding Time Series).
The major reason was to solve the real-world data problem. This research was able to handle the messy reality of consumer wearable data. The researchers adapted a concept known as Joint-Embedding Predictive Architecture (JEPA), that was proposed first by AI pioneer Yann LeCun. Here instead of artificially reconstructing or ‘guessing’ the missing data points the model learns to infer the context of the missing information from surrounding data.
For each individual, 63 distinct time series metrics were recorded at a daily or lower resolution. These metrics ranged from cardiovascular health, respiratory health, sleep, physical activity, and general statistics. Though just 15% of participants had labeled medical histories for evaluation that shows that around 85% of the data would have been unusable in traditional supervised learning approaches. Here JETS first learned from the dataset using self-supervised pre-training and then fine-tuned on the labeled subset.
Purpose of Research
After evaluation JETS model achieved an AUROC of 86.8% for high blood pressure, 70.5% for atrial flutter, 81% for chronic fatigue syndrome, 86.8% for sick sinus syndrome, among others which is impressive.
The model was able to perform much better than the existing baseline models in predicting biomarkers like HbA1c and glucose levels. Though the scores represent risk prediction and not a definitive clinical diagnosis, the study is a significant milestone.
This study suggested that the devices like the Apple Watch can effectively function as long-term health monitoring tools without requiring 24/7 adherence from users.
This study is a great way to reinforces the notion that novel models and training techniques show great promise if one explores the already collected data by wearables such as the Apple Watch, even when they’re not worn all the time.
No Comments