Stress is a public health issue with a growing global impact. This study proposes a comparative approach between various Machine Learning (ML) and Deep Learning (DL) models to detect stress levels based on biomedical signals from the WESAD dataset, which includes physiological recordings from the chest and wrist. The methodology involves preprocessing, feature extraction using autoencoders, and Leave-One-Subject-Out validation. Convolutional Neural Networks (CNN), Bayesian Neural Networks (BNN), Support Vector Machines (SVM), XGBoost, among others, were evaluated considering both individual and combined sensor modalities. The results show that DL-based models, especially CNNs, outperform traditional ones in terms of accuracy and robustness, reaching up to 94.1% accuracy in the combined modality. This research demonstrates the effectiveness of using multimodal biomedical signals and machine learning techniques for proactive stress monitoring, with promising implications for clinical and occupational settings.