Applicability of LSTM and Double Decoders on U-Net models for Sleep and Arousal Scoring Using In-Home EEG Signals

Machine learning has been widely applied to sleep and arousal analysis using PSG (polysomnography), yet limited work has focused on in-home EEG (electroencephalography). Compared to sleep stage classification, arousal detection remains a more challenging task, often requiring specialized methods. Multitask learning offers a promising alternative by enabling models to learn from related tasks simultaneously, potentially improving arousal scoring performance. In this study, we explore multitask learning using different U-Net architectures for joint sleep and arousal scoring in in-home EEG data. We propose and evaluate two key architectural innovations: (1) the integration of bidirectional LSTM layers at different positions within the U-Net architecture, and (2) the use of a branching double decoder for task-specific outputs. The best-performing configurations for arousal scoring were SAAS U-Net DD-L and D-L (Double/Single Decoder followed by LSTM), while SAAS U-Net L-DD and L-D (LSTM between encoder and Double/Single Decoder) achieved superior results for sleep scoring. Notably, sleep scoring performance was further improved through fine-tuning strategies. Experiments across four datasets-including subjects with sleep apnea and low-noise EEG recordings-demonstrate the robustness and practical potential of the proposed models. These findings highlight the value of multitask learning and architectural enhancements to U-Net for advancing sleep diagnostics using portable EEG data.Clinical Relevance- This study highlights the potential of multitask learning models, enhanced with LSTM layers and double decoders, to improve the accuracy of sleep and arousal scoring using portable EEG data. These advancements could facilitate more reliable in-home sleep diagnostics, offering clinicians a practical and accessible tool for evaluating sleep disorders.