Do not sleep on traditional machine learning

Over the last few years, research in automatic sleep scoring has mainly focused on developing increasingly complex deep learning architectures. However, recently these approaches achieved only marginal improvements, often at expense of requiring more data and expensive training procedures. Despite all efforts their satisfactory performance, staging solutions are not widely adopted a clinical context yet. We argue that most for limited real-world applicability as they hard to train, deploy, reproduce. Moreover, lack interpretability transparency, which key increase adoption rates. In this work, we revisit problem stage classification using classical machine learning. Results show competitive performance can be with conventional pipeline consisting preprocessing, feature extraction, simple model. particular, analyze linear model non-linear (gradient boosting) Our approach surpasses state-of-the-art (that uses same data) two public datasets: Sleep-EDF SC-20 (MF1 0.810) ST 0.795), while achieving results SC-78 0.775) MASS SS3 0.817). that, task, expressiveness an engineered vector is par internally learned representations models. This observation opens door adoption, representative allows leverage both successful track record traditional