WEQA Agent Improves Wearable Health Data Question Answering

This research introduces WEQA (Wearable hEalth Question Answering), a novel query-adaptive agent framework designed to improve the accuracy of answering questions based on wearable health data. Traditional language models struggle with the continuous, high-dimensional, and longitudinal nature of wearable sensor data, which differs significantly from their text-centric training distributions. WEQA addresses these challenges by employing an LLM controller that synthesizes execution plans and dynamically routes queries to appropriate combinations of specialized sensor analysis tools and pretrained models. This allows the framework to handle the diversity of sensor modalities and user intents effectively, performing grounded response auditing with external knowledge. Evaluations on a new benchmark, comprising analytic and predictive tasks across three health domains, show that WEQA is 24% more accurate than existing LLM and agentic baselines. A blinded study involving medical experts and users further confirmed substantial gains in both usefulness and clinical soundness, highlighting its potential for practical application in healthcare.