For on-line classification of user states such as emotions or stress levels, we present a new, generic, and efficient physiological fea-ture set. In contrast to common approaches using features specifi-cally tailored to each physiological signal, we break up feature ex-traction into a simple, signal-specific pre-processing step, and the calculation of a comprehensive set of signal-independent features. This systematizes feature design for each physiological signal and facilitates the transfer to other signals. The time complexity of the approach is independent of the size of the analysis window and of the frequency with which feature vectors are computed for classi-fication. We also provide a variant of the feature set that has low memory requirements. Thus, our approach is well suited for im-plementing real-time applications. We evaluate the proposed fea-tures with an emotion and a stress classification task, showing that they are competitive w.r.t. the performance of classifications using signal-tuned state-of-the-art features.