Retrieval-Augmented Personalization with Foundation Models for Wearable Stress Detection

Louis Simon, Mohamed Chetouani

Summary

Frozen, out-of-domain foundation models retrieve similar patterns from a target user's own history; retrieved items are encoded into a compact personalized embedding that modulates a lightweight transformer. WESAD (N=15, wrist BVP/EDA/TEMP/ACC): +3.92% accuracy and +4.76% macro-F1 over a non-personalized transformer, approaching supervised fine-tuning without user labels; temporal (past-only) retrieval ≈ full intra-user retrieval. Cross-dataset cold start with K-EmoCon as retrieval DB gave no significant gain.