WEQA: Bridging LLMs and Wearable Health Data

While Large Language Models (LLMs) excel at medical question answering, their efficacy diminishes when confronted with the complexities of wearable health data. The continuous, high-dimensional, and longitudinal nature of sensor outputs poses a significant challenge for LLMs trained primarily on text. Addressing this gap, researchers have introduced WEQA, a query-adaptive agent framework designed to unify LLM reasoning with specialized wearable analytical and modeling tools.

Synthesizing LLM Reasoning with Sensor Analytics

WEQA employs an LLM controller to dynamically synthesize execution plans. This controller routes each query to an appropriate combination of sensor analysis techniques and pretrained models, enabling a more nuanced approach than fixed workflows or single foundation models. Crucially, it performs grounded response auditing using external knowledge, ensuring accuracy and relevance in its outputs. This integration is key to effectively handling the diversity of sensor modalities and user intents inherent in wearable health data LLM applications.

A Benchmark for Wearable Health Data Analysis

To rigorously evaluate such systems, the researchers curated a comprehensive benchmark. This benchmark spans four open wearable datasets and includes both analytic and predictive tasks across three distinct health domains. Experiments conducted on this benchmark demonstrate that WEQA outperforms existing LLM and agentic baselines by a significant 24%. Furthermore, a blinded study involving medical experts and users confirmed substantial improvements in usefulness and clinical soundness, validating WEQA's potential for real-world healthcare applications.