With the accelerating pace of societal development and the escalating competitive pressures, accurate and reliable stress assessment plays a crucial role in the monitoring of mental health and personalized intervention. Electrocardiogram (ECG) signals have emerged as a critical modality for stress assessment due to their non-invasiveness and capacity to convey abundant physiological information. However, existing methods find it difficult to achieve high-precision classification under multi-level stress states. Meanwhile, they are also faced with challenges such as high feature redundancy, weak temporal correlation, and insufficient fusion of multimodal information. In this study, ECG data of 23 subjects under the pressure induced by a mental arithmetic task at three phase were collected. A time-series dataset was constructed through signal preprocessing and segment division. In addition, a multi-scale feature extraction framework was proposed, covering time-domain, frequency-domain, and nonlinear dynamics indices. A CNN-LSTM network was designed to extract deep spatiotemporal features from the original signals. Furthermore, the TabNet network was adopted for model training and prediction to achieve cross-modal feature interaction and stress state classification. The experimental results show that this method achieves an accuracy of 91% in the three phase stress classification task and has good robustness against inter-individual differences. Compared with existing technologies, this study innovatively integrates HRV features with deep learning features, constructs an ECG multi-scale feature fusion framework, and introduces TabNet for efficient feature modeling, which significantly improves the accuracy and reliability of stress assessment. It provides a new idea for real-time stress monitoring by wearable devices.