Study Finds SSL Representations Fail in Subjective PPG Emotion Detection

A recent study published on arXiv (2608.14675) investigates the role of Self-Supervised Learning (SSL) in analyzing photoplethysmography (PPG) signals to identify strong emotions in everyday situations. Researchers trained a Real-Life PPG encoder (RL-PPG) using unconstrained, real-world data. They confirmed the effectiveness of these representations by demonstrating a nearly fivefold improvement in performance on an objective physical activity recognition task through a leave-one-subject-out (LOSO) evaluation. Conversely, when applied to a subjective emotion detection task, these representations did not outperform basic baselines under the same LOSO protocol. The study revealed that integrating personal data during fine-tuning is essential for enhancing emotion detection, emphasizing the need for personalized methods in subjective applications.