A new study by researchers at MIT and Empirical Health used 3 million days of Apple Watch data to develop a baseline model that predicts medical conditions with impressive accuracy.
See also: Apple Watch – AirPods: Expanding health features to more countries
While Yann LeCun was still AI Chief Scientist , he proposed the Joint Integration Prediction Architecture, or JEPA, which teaches an AI to infer the meaning of missing data instead of the data itself. When faced with gaps in the data, the model learns to predict what the missing parts represent, rather than trying to guess and reconstruct their exact values.
For an image where some parts are covered and others are visible, JEPA will integrate both the visible and covered areas into a common space and have the model infer the representation of the covered area from the visible frame, rather than the exact content that was hidden.
In 2023, Meta released a model called I-JEPA, which was based on LeCun’s original JEPA study. This architecture has become the basis for a field exploring “world models,” a departure from the token prediction focus of LLMs and GPT-based systems. LeCun recently left Meta to start a company focused solely on world models, which he argues are the real path to AGI.
The paper, titled “JETS: A Self-Supervised Joint Integration Time Series Base Model for Behavioral Data in Healthcare,” was accepted into a workshop at NeurIPS. It adapts the JEPA joint integration approach to irregular multivariate time series, such as long-term data from wearable devices where heart rate, sleep, activity, and other metrics appear inconsistently or with large gaps over time.
See also: AI can reveal deeper heart data than Apple Watch's optical sensor
The study uses a longitudinal dataset that includes wearable device data collected from a group of 16,522 individuals, totaling approximately 3 million days. For each individual, 63 discrete time-series metrics were recorded at daily or lower resolution, categorized into five physiological and behavioral domains: cardiovascular health, respiratory health, sleep, physical activity, and general statistics.
Interestingly, only 15% of participants had labeled medical history for assessment, meaning that 85% of the data would be unusable in traditional supervised learning approaches. Instead, JETS first learned from the full dataset via self-supervised pre-training and then adapted to the labeled subset.
To run the model, the researchers created data triplets of observations corresponding to day, price, and metric type. This allowed them to convert each observation into a token, which was masked, encoded, and then fed through a predictive model to predict the inclusion of missing parts.
Once completed, the researchers compared JETS to other baseline models, including an earlier version of JETS based on the Transformer architecture, and evaluated them using AUROC and AUPRC, two standard measures of how well an AI distinguishes between positive and negative cases.
See also: Apple Watch users complain about the Workout app
JETS achieved AUROC of 86.8% for high blood pressure, 70.5% for atrial fibrillation, 81% for chronic fatigue syndrome, and 86.8% for sinus arrhythmia, among others. Although it did not always win, the advantages of JETS are evident in its performance metrics.
