MCFN: A Multichannel Fusion Network for Sleep Apnea Syndrome Detection

Xingfeng Lv, Jinbao Li, Qianqian RenJournal of Healthcare Engineering 2023 · DOI

TL;DR

This paper introduces MCFN, a novel multichannel fusion network designed to improve the accuracy of sleep apnea syndrome (SAS) detection. By effectively fusing multilevel features from multiple respiratory signals using an attention mechanism, MCFN achieves a state-of-the-art accuracy of 87.3% on a large public dataset, outperforming existing methods.

What Problem Does This Paper Address?

Sleep apnea syndrome (SAS) is a prevalent sleep disorder with significant health implications. Traditional diagnosis relies on polysomnography (PSG), which is expensive and inconvenient. While deep learning methods have shown promise for automated SAS detection using physiological signals, they often overlook the varying importance of features from different physiological channels and the complementary information present in both shallow and deep-level features. This paper addresses the challenge of effectively integrating these diverse features to enhance SAS detection accuracy.

Methods

The proposed Multichannel Fusion Network (MCFN) is a deep learning architecture designed to address the limitations of previous methods. Its core components include:

  • Signal Preprocessing: Respiratory signals (airflow, thoracic, and abdominal) are segmented into 30-second epochs, standardized, and labeled according to AASM guidelines.
  • Multilevel Feature Concatenation Block: This block utilizes skip connections to combine shallow and deep-level features extracted by convolutional neural networks (CNNs). This approach aims to retain valuable high-frequency information often lost in traditional deep CNN architectures.
  • Multichannel Attention Fusion Block: This novel component employs an attention mechanism to adaptively recalibrate channel-wise feature responses. It learns to assign different weights to features from different respiratory signals, emphasizing informative channels and suppressing less useful ones. This allows the network to dynamically learn the importance of each feature channel for SAS detection.
  • Classification: The fused features are then passed through convolutional and pooling layers, followed by fully connected layers with a sigmoid activation function for final classification into normal sleep, hypopnea, or obstructive sleep apnea (OSA) events.

Datasets

The study utilizes the Multi-Ethnic Study of Atherosclerosis (MESA) dataset, a large public dataset comprising PSG recordings of 2056 subjects. The network was trained and tested on this dataset, with a focus on three types of respiratory signals: nasal airflow, thoracic, and abdominal.

Key Insights and Ideas

  • Complementarity of Respiratory Signals: Different respiratory signals (airflow, thoracic, abdominal) provide unique and complementary information for SAS detection. Combining them significantly improves accuracy compared to using any single signal.
  • Value of Multilevel Features: Incorporating both shallow and deep-level features, rather than relying solely on deep features, leads to improved detection performance, particularly for signals like thoracic and abdominal respiration.
  • Attention for Feature Fusion: An attention mechanism is crucial for adaptively weighting and fusing features from different channels, allowing the model to prioritize the most informative signals and features for accurate SAS detection.
  • Variable Importance of Channels: The attention mechanism reveals that not all feature channels are equally important for SAS detection. Some channels contribute significantly, while others have minimal impact, highlighting the benefit of adaptive weighting.

Key Results

  • The proposed MCFN model achieved an overall accuracy of 87.3% on the MESA dataset.
  • This accuracy is higher than other state-of-the-art SAS detection methods evaluated on the same dataset.
  • The model demonstrated particularly high performance in detecting OSA events, with a recall of 93.7% and an F1 score of 93.5%.
  • Combining all three respiratory signals (airflow, thoracic, and abdominal) yielded the highest accuracy (87.3%), an improvement of 9.1% over using a single signal.
  • The multilevel feature concatenation block showed a notable improvement in accuracy for thoracic and abdominal respiratory signals, suggesting the importance of shallow features for these signals.

Limitations and Future Work

  • Misclassification of Normal Sleep and Hypopnea: The paper notes some misclassifications between normal sleep and hypopnea events, attributed to similar waveform characteristics, with differences primarily in amplitude.
  • Limited Impact of Multilevel Features on Airflow: For airflow respiratory signals, the improvement from the multilevel feature concatenation block was minimal, suggesting that deep features might be sufficient for this signal type.
  • Future Work: The authors suggest exploring the effect of sleep apnea on sleep staging in future research.

Summary updated

Abstract

Sleep apnea syndrome (SAS) is the most common sleep disorder which affects human life and health. Many researchers use deep learning methods to automatically learn the features of physiological signals. However, these methods ignore the different effects of multichannel features from various physiological signals. To solve this problem, we propose a multichannel fusion network (MCFN), which learns the multilevel features through a convolution neural network on different respiratory signals and then reconstructs the relationship between feature channels with an attention mechanism. MCFN effectively fuses the multichannel features to improve the SAS detection performance. We conducted experiments on the Multi-Ethnic Study of Atherosclerosis (MESA) dataset, consisting of 2056 subjects. The experiment results show that our proposed network achieves an overall accuracy of 87.3%, which is better than other SAS detection methods and can better assist sleep experts in diagnosing sleep disorders.
BibTeX
@article{lv2023mcfn,
  author = {Xingfeng Lv and Jinbao Li and Qianqian Ren},
  title = {{MCFN: A Multichannel Fusion Network for Sleep Apnea Syndrome Detection}},
  journal = {Journal of Healthcare Engineering},
  year = {2023},
  doi = {10.1155/2023/5287043},
}
Figure