The prompt detection of stress can prevent long-term consequences on people's health, economy and society. In this work, we present EMoCy, a reproducible methodology and analysis pipeline for automatic and continuous stress detection based on physiological signals. By providing reproducible experiments and exploring more constrained settings with state-of-the-art accuracy, we set new benchmarks for future works on stress detection. Our study includes signal selection and preprocessing, feature engineering, and classification using machine learning algorithms. We have tested our approach on WESAD dataset using blood volume pulse, electrodermal activity, and respiration signals. We identify a set of features that allow us to reach effective detection using low sampling frequency and short time windows. This reduces the computational power and the detection delay, going towards real-time applications on wearable devices. Using 60s windows, we reach an accuracy of 0.972 and an F1 score of 0.979 on the stress/baseline discrimination task. Moreover, our machine learning models achieve an accuracy greater than 0.93 on windows of 25s sampled at 64Hz. We have also observed that it is possible to obtain a good estimate of the models' performance by training them with a small window overlap.
