<p>The increasing occurrence of stress-related health issues has highlighted the need for accurate and reliable stress detection systems. This study proposes a machine learning-based framework for stress detection using two widely recognized multimodal physiological datasets: NEURO and WESAD. We apply both classical machine learning models and deep learning architectures to classify different types of stress. To effectively capture the temporal and structural patterns in physiological time-series data, the raw signals are converted into image representations using Gramian Angular Summation Fields (GAFs), Gramian Angular Difference Fields (GAFd), and Markov Transition Fields (MTF). While conventional methods transform the data along the time axis, our approach performs transformations along the feature axis. This enables the model to capture feature-centric patterns that are particularly relevant for distinguishing stress states. The classification performance is enhanced by the Convolutional Neural Network (CNN) through the acquisition of spatial features from these visual representations. The Gradient-weighted Class Activation Mapping (Grad-CAM) visualizations highlight the most influential input regions that drive the model’s decision. This, in turn, offers insights into the physiological indicators of stress. Consistent strong generalization is observed across both datasets, with feature-axis transformations based deep learning models achieving around 98.3% accuracy, and classical models like XGBoost and LightGBM achieving approximately 99% accuracy. The integration of Grad-CAM validates the relevance of learned features strengthening the model’s interpretability. Altogether, our approach clearly demonstrates the strong potential of integrating advanced signal transformation techniques with explainable deep learning to create robust and interpretable stress detection systems.</p>