[Paper Review] MSSC-BiMamba: Multimodal Sleep Stage Classification and Early Diagnosis of Sleep Disorders with Bidirectional Mamba

Chao Zhanga, Weirong Cuia+1 more|arXiv (Cornell University)|May 30, 2024

Restless Legs Syndrome Research4 citations

TL;DR

This paper proposes MSSC-BiMamba, a novel multimodal deep learning model that combines Efficient Channel Attention (ECA) with a Bidirectional State Space Model (BiMamba) for automated sleep stage classification and early diagnosis of sleep disorders using polysomnography (PSG) data. The model achieves 95.2% accuracy in sleep health prediction and demonstrates superior computational efficiency and generalization over Transformer-based models.

ABSTRACT

Monitoring sleep states is essential for evaluating sleep quality and diagnosing sleep disorders. Traditional manual staging is time-consuming and prone to subjective bias, often resulting in inconsistent outcomes. Here, we developed an automated model for sleep staging and disorder classification to enhance diagnostic accuracy and efficiency. Considering the characteristics of polysomnography (PSG) multi-lead sleep monitoring, we designed a multimodal sleep state classification model, MSSC-BiMamba, that combines an Efficient Channel Attention (ECA) mechanism with a Bidirectional State Space Model (BSSM). The ECA module allows for weighting data from different sensor channels, thereby amplifying the influence of diverse sensor inputs. Additionally, the implementation of bidirectional Mamba (BiMamba) enables the model to effectively capture the multidimensional features and long-range dependencies of PSG data. The developed model demonstrated impressive performance on sleep stage classification tasks on both the ISRUC-S3 and ISRUC-S1 datasets, respectively containing data with healthy and unhealthy sleep patterns. Also, the model exhibited a high accuracy for sleep health prediction when evaluated on a combined dataset consisting of ISRUC and Sleep-EDF. Our model, which can effectively handle diverse sleep conditions, is the first to apply BiMamba to sleep staging with multimodal PSG data, showing substantial gains in computational and memory efficiency over traditional Transformer-style models. This method enhances sleep health management by making monitoring more accessible and extending advanced healthcare through innovative technology.

Motivation & Objective

• To address the limitations of manual sleep staging, which is time-consuming and subjective, by developing an automated, efficient, and accurate sleep stage classification system.
• To improve the detection of sleep disorders by enabling early diagnosis using multimodal polysomnographic (PSG) data.
• To enhance computational and memory efficiency compared to existing Transformer-based models in sleep staging tasks.
• To leverage bidirectional long-range temporal modeling for improved capture of complex sleep dynamics in PSG signals.
• To validate the model’s generalization across diverse datasets, including healthy and disordered sleep patterns.

Proposed method

• The model integrates an Efficient Channel Attention (ECA) mechanism to dynamically weight contributions from multiple PSG sensor channels, enhancing feature representation.
• A Bidirectional State Space Model (BiMamba) is employed to capture long-range temporal dependencies in both forward and reverse directions, improving sequence modeling.
• The architecture is specifically designed for multimodal PSG data, combining EEG, EOG, EMG, and other signals to improve classification robustness.
• The model is trained and evaluated on the ISRUC-S1, ISRUC-S3, and Sleep-EDF153 datasets, using K-fold cross-validation to assess generalization.
• The framework replaces self-attention mechanisms with Mamba’s state space formulation, reducing computational cost while maintaining high performance.
• Performance is evaluated using standard metrics including accuracy, F1-score, and Kappa coefficient across multiple sleep stages (N1, N2, N3, REM).

Experimental results

Research questions

• RQ1 Can a bidirectional Mamba-based architecture effectively model long-range temporal dependencies in multimodal polysomnographic data for sleep stage classification?
• RQ2 How does the integration of Efficient Channel Attention improve feature representation across diverse PSG sensor channels?
• RQ3 To what extent does MSSC-BiMamba generalize across different datasets with varying sleep patterns (healthy vs. disordered)?
• RQ4 Can the model achieve high diagnostic accuracy for early sleep disorder detection while maintaining computational efficiency?
• RQ5 How does the model’s performance compare to existing Transformer-based models in terms of accuracy, inference speed, and memory usage?

Key findings

• The MSSC-BiMamba model achieved an overall accuracy of 0.952 in sleep health status prediction on a combined dataset of Sleep-EDF153 and ISRUC, demonstrating strong generalization for early diagnosis.
• On the ISRUC-S3 dataset, the model achieved an overall accuracy of 0.857, F1-score of 0.851, and Kappa of 0.817, indicating strong performance on standard sleep staging tasks.
• The model trained on ISRUC-S1(50) showed robust generalization, achieving 0.852 accuracy on the S1(50) test set and maintaining high performance across S1(100) and S3 subsets.
• The model demonstrated superior computational and memory efficiency compared to Transformer-based models, making it suitable for real-time and resource-constrained applications.
• Cross-validation results confirmed the model’s stability and reliability, with consistent F1-scores across all sleep stages, including REM (0.893 on S3).
• Increasing the training set size improved model performance, with the 9:1 split achieving the highest accuracy of 0.952, indicating strong scalability.