Study compares CNN, Transformer, and Mamba deep learning architectures for PPG-based emotion recognition from wearable devices

Researchers evaluated four deep learning models—CNN, CNN-LSTM hybrid, Transformers, and Mamba—for classifying arousal, valence, and relaxation states from wrist-based PPG (photoplethysmography, a signal used to measure heart activity) signals under a subject-independent 5-fold cross-validation protocol.

CNNs achieved the highest accuracy with the smallest model size overall, while Transformer and Mamba models showed performance comparable to CNNs but did not consistently outperform them across all tasks; Transformers achieved a better balance of F1 scores for Arousal and Relaxation.

This is the first evaluation of Transformer and Mamba models for PPG-based affect recognition, offering guidance on model selection for wearable affective monitoring systems in consumer devices.