Cluster-Then-Label Strategy for Sleep Detection U sing Electroencephalogram (EEG)

Data deficiency has long been a major obstacle for many areas of machine learning. Manually labeling data could be utterly time-consuming and expensive, especially in medical fields including the sleep research. To address this challenge, we propose a cluster-then-label strategy, which inserts clustering algorithms into traditional supervised learning pipelines. In this paper, we demonstrate the cluster-then-label strategy for sleep detection using electroencephalogram (EEG) data and show that our proposed strategy can boost the classifier's performance on previously unseen subjects. We also developed a method based on nonlinear transformations that can reshape the feature distributions to resemble normal distributions with which the cluster-then-label algorithm becomes more efficient and robust.