SeqSleepNet: End-to-End Hierarchical Recurrent Neural Network for Sequence-to-Sequence Automatic Sleep Staging (1809.10932v3)

Abstract: Automatic sleep staging has been often treated as a simple classification problem that aims at determining the label of individual target polysomnography (PSG) epochs one at a time. In this work, we tackle the task as a sequence-to-sequence classification problem that receives a sequence of multiple epochs as input and classifies all of their labels at once. For this purpose, we propose a hierarchical recurrent neural network named SeqSleepNet. At the epoch processing level, the network consists of a filterbank layer tailored to learn frequency-domain filters for preprocessing and an attention-based recurrent layer designed for short-term sequential modelling. At the sequence processing level, a recurrent layer placed on top of the learned epoch-wise features for long-term modelling of sequential epochs. The classification is then carried out on the output vectors at every time step of the top recurrent layer to produce the sequence of output labels. Despite being hierarchical, we present a strategy to train the network in an end-to-end fashion. We show that the proposed network outperforms state-of-the-art approaches, achieving an overall accuracy, macro F1-score, and Cohen's kappa of 87.1%, 83.3%, and 0.815 on a publicly available dataset with 200 subjects.

• The paper introduces SeqSleepNet, a hierarchical RNN that integrates epoch-level and sequence-level processing for improved automatic sleep staging.
• It achieves 87.1% accuracy, a macro F1-score of 83.3%, and superior sensitivity in challenging stages like N1 and REM.
• The end-to-end training framework streamlines feature extraction, offering a robust solution for clinical sleep monitoring and related biomedical applications.

Analysis of SeqSleepNet: End-to-End Hierarchical Recurrent Neural Network for Sequence-to-Sequence Automatic Sleep Staging

The paper delineates an innovative approach to automatic sleep staging by conceptualizing it as a sequence-to-sequence classification task. The authors introduce SeqSleepNet, a hierarchical recurrent neural network (RNN) designed to process sequences of polysomnography (PSG) epochs, rather than treating each epoch independently. This method leverages both epoch-level and sequence-level modeling through a comprehensive network architecture, offering an advancement over traditional one-to-one classification models.

Hierarchical Network Architecture

SeqSleepNet employs a sophisticated network structure with two primary processing levels. At the epoch-level, a filterbank layer is introduced for preprocessing via frequency-domain filtering, followed by an attention-based bidirectional recurrent layer for short-term sequential modeling. The sequence-level processing involves an additional recurrent layer deployed to handle long-term inter-epoch dependencies. This hierarchical approach ensures the model captures not only the intra-epoch dynamics through sequential features but also inter-epoch relationships across sequences.

The model is trained in an end-to-end manner, a significant stride over previous methodologies that often necessitated separate training stages for different network layers. This form of training facilitates a seamless integration of layer operations, enhancing the model's capacity for feature extraction and classification.

Performance Evaluation

The proposed SeqSleepNet was evaluated on the Montreal Archive of Sleep Studies (MASS) dataset, comprising recordings from 200 subjects. The model achieved remarkable performance, with an overall accuracy of 87.1%, a macro F1-score of 83.3%, and Cohen’s kappa of 0.815. These results signify an enhancement in accuracy and reliability over existing models, particularly highlighting substantial improvements in classifying challenging sleep stages like N1 and REM.

Comparison with the E2E-DeepSleepNet baseline, as well as other methods, underscores SeqSleepNet’s efficacy. While the difference in overall accuracy between SeqSleepNet and E2E-DeepSleepNet might seem marginal, the former exhibits a finer class-wise sensitivity, especially in recognizing N1 sleep stages, indicating its robustness in modeling transitional sleep phenomena.

Implications and Future Directions

The application of SeqSleepNet to automatic sleep staging holds substantial implications. Its ability to handle sequences of multiple epochs simultaneously positions it as a viable candidate for clinical settings where reliable automatic scoring is invaluable. Moreover, the end-to-end training framework offers a blueprint for developing similar models in other sequential tasks within biomedical engineering and beyond.

Future research directions could explore integrating SeqSleepNet with other advanced network architectures or enhancing the model with additional data types, such as wearable sensor inputs, to elevate its utility in diverse environments. Furthermore, refining the model to reduce latency during online applications could bridge the gap to real-time deployment, enhancing home-based sleep monitoring systems.

By unifying epoch-level and sequence-level analyses, SeqSleepNet sets a benchmark in sleep staging tasks, showcasing the potential of hierarchical RNNs in extracting nuanced patterns from complex datasets. This advancement, while tested within the context of sleep studies, may well stimulate parallel innovations across related domains in time-series analysis and healthcare informatics.