Abstract
Sleep staging traditionally requires massive time and expertise from clinicians. Various automated sleep staging methods have been developed to streamline this task, however, they commonly need extensive labeled clinical sleep data. ActiveSleepLearner is proposed to address this challenge, a transfer learning framework that leverages active learning techniques to achieve weakly supervised sleep staging across various settings, devices, and populations. On one hand, a selection algorithm that identifies the most informative samples for manual labeling is introduced, considering factors like sample representativeness, diversity, and complexity. On the other hand, during the fine-tuning phase with limited samples, model training efficiency is improved by introducing a novel joint loss function, implementing well-designed data augmentation techniques, and adopting layer-wise and multi-step learning rate strategies. Our approach reduces the manual annotation budget required from clinicians, enabling them to label only a small portion of the sizeable unlabeled dataset while the system handles the rest. Across the four datasets evaluated, our method consistently outperforms baselines and other partial-label methods. Especially, even with 5% labels, ActiveSleepLearner reaches 97.5% of the accuracy achieved by the model explicitly trained on this dataset. More importantly, ActiveSleepLearner fosters collaboration between humans and machines, reducing the clinician's burden by requiring them to label only a part of epochs instead of all recordings.