Retrieval-Augmented Personalization with Foundation Models for Wearable Stress Detection

Researchers propose a lightweight retrieval-augmented personalization method for wearable-based stress detection that uses frozen foundation models to retrieve similar patterns from a user's history, achieving 3.92% accuracy gains over non-personalized baselines without requiring labeled data. The approach demonstrates that personalized AI models for health monitoring can be built efficiently by leveraging historical user data rather than expensive fine-tuning, with performance remaining robust even with limited user history.

This research addresses a fundamental challenge in wearable health technology: adapting AI models to individual physiological differences without requiring extensive labeled data or computational resources. Stress detection through wearables has garnered significant attention due to growing mental health awareness and the proliferation of consumer devices, yet traditional personalization methods remain costly and impractical for deployment at scale. The proposed retrieval-augmented approach represents a meaningful advancement by using frozen foundation models—pre-trained neural networks that don't require updating—to identify similar patterns in a user's historical data, then encoding these into compact personalized embeddings that guide a lightweight transformer network. This method achieves near supervised fine-tuning performance while avoiding the data labeling burden entirely.

The technical contribution gains practical importance in the broader context of edge AI and federated learning, where computational constraints and privacy concerns limit traditional deep learning approaches. The 4.76% macro F1-score improvement demonstrates measurable performance gains that could translate to meaningfully better stress detection accuracy in real-world applications. Cross-dataset retrieval experiments using the K-Emocon dataset for WESAD personalization suggest the approach's generalizability across different wearable data sources, addressing a common fragmentation problem in health tech.

For the wearable and digital health sectors, this work reduces barriers to deploying personalized health monitoring at scale. Developers can implement sophisticated personalization without accumulating large labeled datasets, accelerating product development cycles. The temporal retrieval findings—showing performance with only past user samples—align well with real deployment scenarios where historical data naturally accumulates. Future work examining larger user populations and extended temporal windows will determine whether this approach achieves production-grade reliability for clinical applications.

• →Retrieval-augmented personalization enables effective stress detection from wearables without labeled user data, reducing deployment barriers for health tech applications
• →The method achieves 3.92% accuracy and 4.76% F1-score improvements over non-personalized baselines by leveraging frozen foundation models and user historical patterns
• →Temporal retrieval using only prior user samples maintains near-optimal performance, making the approach practical for real-world scenarios with limited history
• →Cross-dataset personalization suggests the technique generalizes across different wearable sources, addressing fragmentation in health monitoring ecosystems
• →Lightweight transformer networks with personalized embeddings provide computational efficiency suitable for edge deployment on resource-constrained wearable devices