Few-Shot Personalized Stress Detection from ECG via Self-Supervised Learning

Physiological stress detection using wearable sensors is challenging due to inter-individual variability and limited availability of labeled data for personalization. This study investigates the potential of self-supervised learning (SSL) to enable few-shot personalization in electrocardiogram (ECG)- based stress detection.We compare four approaches: a support vector classifier (SVC), a supervised convolutional neural network (SCNN), and two SSL-based CNN models pretrained via contrastive learning. All models are pretrained on the WESAD dataset and evaluated on a target dataset using transfer learning and few-shot personalization. The results indicate that the SSL-based CNN model exhibits faster adaptation when trained with fewer than five samples, achieving superior performance compared with the SVC. These findings highlight the potential of SSL for real-world affective computing applications, where labeled data is limited and rapid adaptation to new users is required.