This study evaluates five machine learning models, such as AutoKeras, Bayesian Neural Network (BNN), Multilayer Perceptron (MLP), Random Forest, and Support Vector Machine (SVM), on the publicly available WESAD dataset. The models are tested under three sensor configurations: chest-only, wrist-only, and combined. While MLP achieved the highest accuracy (83.17%), SVM and Random Forest delivered comparable performance with much lower computational requirements, making them suitable for wearable deployment. BNN offered the advantage of uncertainty estimation, though at the cost of significantly increased resource usage. In addition to classification accuracy, we report training and inference times, memory usage, and provide preliminary explainability results via feature importance (Random Forest, SVM) and Bayesian uncertainty estimation. These findings underscore the trade-offs between accuracy, efficiency, and explainability in stress detection systems.
Pataca et al. (Thu,) studied this question.