mDOT Center Srivastava Lab Unveils LLMSense for High-Level Reasoning in mHealth from Sensor Data

The mDOT Center’s TR&D3 Srivastava Lab at UCLA has unveiled LLMSense, a cutting-edge framework that leverages the reasoning capabilities of Large Language Models (LLMs) to perform high-level inferences over long-term spatiotemporal sensor traces in mHealth applications. Moving beyond basic sensor perception, LLMSense enables the analysis of complex human activity routines and behaviors, providing insights critical for health monitoring and intervention.

One of the key applications of LLMSense is in dementia diagnosis, where the framework interprets behavioral traces collected over time, achieving over 80% accuracy. To manage the challenges of extensive sensor datasets, LLMSense employs strategies such as summarization before reasoning and the selective inclusion of historical traces, ensuring efficient and effective performance without sacrificing detail or fidelity.

Designed with a hybrid edge-cloud architecture, LLMSense allows smaller LLMs to run on edge devices for preliminary data summarization, while higher-level reasoning occurs on the cloud to preserve user privacy. This setup ensures that sensitive sensor data remains secure while enabling comprehensive analysis of long-term behavioral patterns.

Published at the 2024 IEEE 3rd Workshop on Machine Learning on Edge in Sensor Systems (SenSys-ML), LLMSense provides both a technological breakthrough and a set of practical guidelines for applying LLMs to complex mHealth reasoning tasks. Its development represents a significant step toward smarter, privacy-preserving health monitoring systems that can extract actionable insights from rich sensor datasets.

Study citation
X. Ouyang and M. Srivastava, “LLMSense: Harnessing LLMs for High-level Reasoning Over Spatiotemporal Sensor Traces,” 2024 IEEE 3rd Workshop on Machine Learning on Edge in Sensor Systems (SenSys-ML), Hong Kong, Hong Kong, 2024, pp. 9-14, doi: 10.1109/SenSys-ML62579.2024.00007.