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

Area of Science:

• Biomedical Engineering
• Digital Health
• Wearable Technology

Background:

• Wearable accelerometers collect detailed behavioral data for personalized health insights.
• Limited comparative studies exist on modern representation-learning methods for accelerometer data.

Purpose of the Study:

• To compare the effectiveness of different data representation methods for predicting clinical outcomes using accelerometer data.
• To evaluate simple entropy-based features, large-language-model (LLM) embeddings, and time-series foundation model embeddings.

Main Methods:

• Utilized accelerometer data from the National Health and Nutrition Examination Survey (NHANES).
• Assessed prediction accuracy for outcomes including overweight status, lipid biomarkers, glucose, arthritis, and cancers.
• Compared entropy-based features, LLM embeddings, and time-series foundation model embeddings.

Main Results:

• Entropy-based features performed comparably or better than embedding approaches across all outcomes.
• LLM embeddings provided only marginal predictive improvements (AUC difference of 0.01-0.05).
• Time-series foundation model embeddings showed minimal utility, and prompt-based LLM reasoning performed poorly.

Conclusions:

• Simple variability features derived from accelerometer data are highly effective for health outcome prediction.
• Current advanced AI embeddings offer limited advantages over basic features for this data.
• Future health sensing foundation models require domain-aligned pretraining for improved performance.