[Paper Review] L-SeqSleepNet: Whole-cycle Long Sequence Modelling for Automatic Sleep Staging

Huy P. Phan, Kristian P. Lorenzen+8 more|arXiv (Cornell University)|Jan 9, 2023

EEG and Brain-Computer Interfaces10 citations

TL;DR

L-SeqSleepNet introduces whole-cycle long sequence modelling for sleep staging, achieving state-of-the-art results across multiple EEG setups with efficient long-sequence processing and single-channel inputs.

ABSTRACT

Human sleep is cyclical with a period of approximately 90 minutes, implying long temporal dependency in the sleep data. Yet, exploring this long-term dependency when developing sleep staging models has remained untouched. In this work, we show that while encoding the logic of a whole sleep cycle is crucial to improve sleep staging performance, the sequential modelling approach in existing state-of-the-art deep learning models are inefficient for that purpose. We thus introduce a method for efficient long sequence modelling and propose a new deep learning model, L-SeqSleepNet, which takes into account whole-cycle sleep information for sleep staging. Evaluating L-SeqSleepNet on four distinct databases of various sizes, we demonstrate state-of-the-art performance obtained by the model over three different EEG setups, including scalp EEG in conventional Polysomnography (PSG), in-ear EEG, and around-the-ear EEG (cEEGrid), even with a single EEG channel input. Our analyses also show that L-SeqSleepNet is able to alleviate the predominance of N2 sleep (the major class in terms of classification) to bring down errors in other sleep stages. Moreover the network becomes much more robust, meaning that for all subjects where the baseline method had exceptionally poor performance, their performance are improved significantly. Finally, the computation time only grows at a sub-linear rate when the sequence length increases.

Motivation & Objective

• Motivate the use of long-term, whole-cycle temporal structure in sleep staging to improve accuracy.
• Develop an efficient long-sequence modelling architecture suitable for whole sleep cycles (~90 minutes).
• demonstrate robustness and transferability of the model across diverse EEG modalities and databases.
• Show that long-cycle modelling reduces misclassifications across sleep stages while maintaining computational efficiency.

Proposed method

• Convert each 30-second EEG epoch into a log-magnitude time-frequency image via STFT.
• Encode each epoch using a shared subnetwork with a learnable filterbank, BLSTM, and gated attention to produce epoch embeddings.
• Fold long sequences into non-overlapping subsequences, apply intra-subsequence BLSTM processing with residual connections, then perform inter-subsequence BLSTM modeling, and unfold to a long sequence.
• Train with a cross-entropy loss over the entire sequence with L2 regularization.
• Use a two-layer fully connected classifier with ReLU activations and a softmax output to predict stage probabilities for all epochs in the long sequence.

Experimental results

Research questions

• RQ1 Does modelling information from an entire sleep cycle (long sequences) improve automatic sleep staging over conventional short-context approaches?
• RQ2 Can long-sequence modelling maintain or improve accuracy across different EEG setups (scalp, in-ear, cEEGrid) using single-channel input?
• RQ3 How does whole-cycle modelling affect class-wise performance, particularly for underrepresented stages like N3 and REM?
• RQ4 Is the long-sequence approach computationally efficient and scalable with sequence length?

Key findings

• L-SeqSleepNet consistently outperforms the SeqSleepNet baseline across four databases with random initialization and with pretraining-based initialization.
• On SHHS, L-SeqSleepNet achieves 88.4% accuracy, outperforming SleepTransformer and XSleepNet by 0.7–0.8 percentage points in the reported setup.
• With pretraining, gains on SleepEDF, ear-EEG, and cEEGrid reach 1.0%, 1.9%, and 3.9% respectively over the baseline.
• Long-cycle modelling reduces errors in N3 and REM across datasets, with some trade-off increases in N2, suggesting improved handling of cycle-based structure.
• The approach remains robust, improving performance for subjects where the baseline performed poorly, and computational growth is sub-linear with sequence length.