Currently, most studies on mental stress evaluation mainly focus on classification tasks, while research on accurately estimating continuous stress levels using deep learning for early identification remains limited. This study proposes an end-to-end continuous stress assessment framework based on a deep hybrid learning architecture. The framework employs efficient channel attention convolution to extract local pattern features from the signals, utilizes a bidirectional long short-term memory (BiLSTM) network to model contextual dependencies, and incorporates emotional cross-attention to assign importance weights to different emotional states. In addition, an adaptive ridge stacking ensemble learning method is proposed. To enhance feature representation, pulse rate variability (PRV) and discrete pulse signals (dPS) extracted from Photoplethysmography (PPG) signals are encoded into markov transition field (MTF) and recurrence plot (RP) images, respectively. The results show that, for PRV, MTF- and RP-based representations reduce the detection error by 6.93% and 2.57% compared with the time-domain baseline. For dPS, the error reductions reach 6.97% and 15.05%. Furthermore, the proposed fusion strategy of PRV-MTF and dPS-RP achieves the best performance (MAE = 3.29, RMSE = 4.05). Compared with the previous state-of-the-art method based on time-domain fusion of PRV and dPS signals (MAE = 4.38, RMSE = 5.19), the proposed approach yields substantial reductions of 24.88% in MAE and 21.96% in RMSE, reaching the current state-of-the-art performance. These results demonstrate that transforming time-domain signals into structured encoding images enables more effective capture of deep patterns associated with psychological states, thereby significantly improving the accuracy of mental health detection.
Li et al. (2026) studied this question.