Evaluating Representation Embeddings from LLMs and Time-Series Foundation Models for Wearable Accelerometer-Based Health Prediction.

Abstract

Wearable accelerometer data capture rich behavioral signals relevant for personalized health, yet the comparative evidence on modern representation-learning approaches remains limited. Using accelerometer data from the National Health and Nutrition Examination Survey (NHANES), we evaluated three representation families for predicting multiple clinical outcomes: simple entropy-based features, pretrained large-language-model (LLM) embeddings, and time-series foundation model embeddings. Outcomes included overweight status, lipid biomarkers, glucose, arthritis, and cancers. Across all endpoints, entropy-based features consistently performed comparably to, and often slightly better than, embedding approaches. LLM-derived embeddings offered only marginal improvements (ΔAUC≈0.01-0.05), and time-series foundation model embeddings provided minimal value across varying sequence lengths. Prompt-based LLM reasoning performed worst (AUC≈0.56-0.65), demonstrating limited ability to infer quantitative physiological states from structured text. These results highlight the strength of simple variability features and underscore the need for domain-aligned pretraining in future time-series foundation models for health sensing.