[Paper Review] Transparency in Sleep Staging: Deep Learning Method for EEG Sleep Stage Classification with Model Interpretability

Shivam Sharma, Suvadeep Maiti+3 more|arXiv (Cornell University)|Sep 10, 2023

EEG and Brain-Computer InterfacesNeuroscience3 citations

TL;DR

This paper proposes a novel end-to-end deep learning model for single-channel EEG-based sleep stage classification that integrates squeeze-and-excitation residual networks (SE-ResNet) and stacked Bi-LSTM layers to capture spatial and temporal features. It introduces the first explainable deep learning approach in sleep staging using 1D-GradCAM, achieving high alignment with expert annotations and macro-F1 scores of 82.5, 78.9, and 81.9 on SleepEDF-20, SleepEDF-78, and SHHS datasets, respectively, while accelerating training by 8x via increased stride.

ABSTRACT

Automated Sleep stage classification using raw single channel EEG is a critical tool for sleep quality assessment and disorder diagnosis. However, modelling the complexity and variability inherent in this signal is a challenging task, limiting their practicality and effectiveness in clinical settings. To mitigate these challenges, this study presents an end-to-end deep learning (DL) model which integrates squeeze and excitation blocks within the residual network to extract features and stacked Bi-LSTM to understand complex temporal dependencies. A distinctive aspect of this study is the adaptation of GradCam for sleep staging, marking the first instance of an explainable DL model in this domain with alignment of its decision-making with sleep expert's insights. We evaluated our model on the publically available datasets (SleepEDF-20, SleepEDF-78, and SHHS), achieving Macro-F1 scores of 82.5, 78.9, and 81.9, respectively. Additionally, a novel training efficiency enhancement strategy was implemented by increasing stride size, leading to 8x faster training times with minimal impact on performance. Comparative analyses underscore our model outperforms all existing baselines, indicating its potential for clinical usage.

Motivation & Objective

• Address the challenge of high variability and complexity in single-channel EEG signals for accurate sleep stage classification.
• Improve model interpretability in automated sleep staging to support clinical adoption and trust.
• Develop a data-efficient training strategy to reduce computational cost without sacrificing performance.
• Establish a benchmark for automated sleep staging by outperforming existing state-of-the-art methods.

Proposed method

• Employ a hybrid architecture combining SE-ResNet for channel-wise attention-based feature extraction and stacked Bi-LSTM for modeling long-range temporal dependencies in EEG sequences.
• Adapt 1D-GradCAM to generate class activation maps that visualize which EEG signal segments influence the model’s classification decisions.
• Implement a training efficiency enhancement by increasing the stride size in convolutional layers, reducing computation with minimal performance drop.
• Use public datasets (SleepEDF-20, SleepEDF-78, SHHS) for training and evaluation, with standard 30-second epoch-based scoring.
• Conduct a blind validation with sleep experts using identical EEG windows to compare model attention regions with expert-annotated microstructures.
• Leverage AASM and R&K scoring guidelines to ensure alignment with clinical standards in data labeling and evaluation.

Experimental results

Research questions

• RQ1 Can a deep learning model achieve state-of-the-art performance in single-channel EEG sleep staging while maintaining interpretability?
• RQ2 To what extent does the model’s attention, as visualized via 1D-GradCAM, align with clinically relevant EEG microstructures identified by sleep experts?
• RQ3 How effective is the proposed training efficiency strategy (increased stride) in reducing training time without degrading model performance?
• RQ4 Can the integration of squeeze-and-excitation blocks with residual networks improve feature representation in EEG sleep staging?
• RQ5 Does the model generalize well across diverse public datasets such as SleepEDF-20, SleepEDF-78, and SHHS?

Key findings

• The model achieved macro-F1 scores of 82.5 on SleepEDF-20, 78.9 on SleepEDF-78, and 81.9 on SHHS, outperforming existing state-of-the-art methods.
• 1D-GradCAM visualizations demonstrated strong alignment with expert-identified EEG microstructures such as alpha waves, sleep spindles, K-complexes, and delta waves.
• The model successfully detected key markers like K-complexes in N2 and delta waves in N3, confirming its ability to focus on diagnostically relevant features.
• The training efficiency enhancement via increased stride size reduced training time by 8x with negligible impact on performance.
• Blind validation with sleep specialists confirmed that the model’s attention regions closely matched the signal segments experts prioritize during manual scoring.
• The integration of SE-ResNet and stacked Bi-LSTM significantly improved feature extraction and temporal modeling, contributing to superior classification accuracy.