DeepSleepNet: a Model for Automatic Sleep Stage Scoring based on Raw Single-Channel EEG. Objective: The present study proposes a deep learn- ing model, named DeepSleepNet, for automatic sleep stage scoring based on raw single-channel EEG, and a two-step training algorithm used to effectively train such model. Methods: Most of the existing methods rely on hand-engineered features which require prior knowledge about sleep stage scoring. Only a few of them encode the temporal information such as stage transition rules, which is important to correctly identify the next possible sleep stages, into the extracted features. In the proposed model, we utilize Convolutional Neural Networks (CNNs) to extract time-invariant features, and bidirectional Long Short- Term Memory (bidirectional-LSTM) to learn transition rules among sleep stages from EEG epochs. We implement a two-step training algorithm that pre-trains the model with oversampled dataset to alleviate class-imbalance problems and fine-tunes the model with sequences of EEG epochs to encode the temporal information into the model. Results: We applied our model to the F4-EOG (Left) channel from a set of 62 subjects in an open- access database, containing 58600 EEG epochs ( 488 hours). The results demonstrated that our model scored the EEG epochs with the accuracy of 86.2% and the macro F1-score of 81.7. Conclusion: Without utilizing any hand-engineered features, our model can achieve a similar sleep stage scoring performance with the highest macro F1-score compared to the state-of-the- art methods. Significance: This study proposes a deep learning model that can automatically learn features from raw single- channel EEG, and accurately score EEG epochs as good as the state-of-the-art hand-engineering methods.