Personal Data Crucial for PPG Emotion Detection AI.

Researchers investigated the effectiveness of Self-Supervised Learning (SSL) in extracting general representations from photoplethysmography (PPG) signals for real-life emotion detection. They first pre-trained a Real-Life PPG encoder (RL-PPG) on unconstrained data and confirmed its strong transferability to an objective task, such as physical activity recognition, where it significantly outperformed baselines. This demonstrated the SSL model's ability to learn robust general features from noisy physiological data. However, when these same general representations were applied to the highly subjective task of real-life emotion detection, they failed to surpass naive baselines in a leave-one-subject-out evaluation. The study found that incorporating an individual's personal data during the fine-tuning phase was the critical factor for achieving predictive performance, outweighing the benefits of population-level pre-training. This suggests that for subjective affective inference, general SSL representations alone are insufficient, and personalization is a key component for successful real-world emotion recognition systems. The code and pre-trained weights are shared to support future research.