<p>Stress is an increasingly pervasive aspect of modern life that can adversely affect health by disrupting multiple physiological systems. Early detection of stress is critical for mitigating its potential impact on cardiovascular and reproductive functions. Among various physiological signals, photoplethysmogram (PPG) has attracted significant attention due to its non-invasive nature, ease of acquisition, and cost-effectiveness. In this paper, we introduce CEM-Net, a lightweight network designed for psychological stress detection using wrist PPG signals. CEM-Net efficiently extracts and models both local and global features from short time-series windows while maintaining a low parameter count and minimal computational overhead, making it highly suitable for deployment in resource-constrained environments. The model synergistically integrates a convolutional neural network for local feature extraction, an Efficient Channel Attention module for dynamic channel weighting, and a Mamba module for robust long-term temporal modeling. Comprehensive experiments on the WESAD and MAUS datasets demonstrate that CEM-Net outperforms both traditional machine learning and state-of-the-art deep learning methods. Furthermore, experiments on personalized modeling and noise interference confirm the model’s adaptability and robustness under diverse and challenging conditions.</p>