Retrieval-Augmented Personalization Improves Wearable Stress Detection

Personalizing stress detection systems based on wearable data is challenging due to the wide variability in how individuals physiologically and behaviorally respond to stress. Traditional methods often involve extensive user-specific fine-tuning or costly self-supervised pre-training on large datasets. This research introduces a more efficient alternative: retrieval-augmented personalization. The proposed method utilizes frozen, pre-trained foundation models that are out-of-domain. These models are used to retrieve patterns from a target user's historical data that are similar to current inputs. These retrieved patterns are then encoded into a compact, personalized embedding. This embedding subsequently modulates the representations extracted by a lightweight transformer network, effectively tailoring the model's interpretation to the individual user. Evaluated on the WESAD stress detection dataset, which includes physiological and activity signals from wrist-worn devices, the approach demonstrated notable improvements. It achieved gains of nearly 4% in accuracy and almost 5% in macro F1-score compared to a non-personalized transformer baseline. Crucially, this performance approaches that of supervised fine-tuning but without the need for any labeled user-specific data. The study also showed robustness to limited user history and successful cross-dataset personalization, highlighting its practical applicability.