<p>Precise sleep stage classification is vital for advancing sleep research and enhancing the clinical assessment of sleep-related disorders. In this study, a comprehensive and scalable machine learning framework is introduced for automated sleep staging, leveraging polysomnographic (PSG) data that incorporates electroencephalogram (EEG), electrooculogram (EOG), and snoring-based acoustic features. The methodological novelty of this work lies in three distinct aspects: (1) the first integration of snoring-derived acoustic features with traditional EEG and EOG signals within an autoencoder-based feature selection framework for sleep staging; (2) a novel application of Autoencoder-Based Feature Selection (AES) that compresses multimodal features into a discriminative latent space while preserving inter-class separability; and (3) a systematic comparative analysis of five state-of-the-art ensemble methods (XGBoost, LightGBM, CatBoost, Gradient Boosting, Random Forest) under identical preprocessing and feature selection conditions. After the initial extraction and preprocessing of signals, a combined feature extraction strategy is utilized, drawing on both statistical metrics and wavelet-based analysis with the Daubechies-4 (db4) wavelet. To counteract class imbalance and bolster model generalization, Gaussian Noise Data Augmentation (GNDA) is applied exclusively to the training data after stratified train-test splitting (80-20 ratio), with augmentation parameters (noise standard deviation <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\sigma = 0.05\)</EquationSource> </InlineEquation>) optimized through validation to ensure no data leakage occurs between training and test sets. Autoencoder-Based Feature Selection (AES) is then used to compress the feature space and retain the most informative attributes from the augmented training data, reducing dimensionality from 124 to 64 latent dimensions while preserving 99.2% of discriminative information. For classification, five machine learning models are employed: XGBoost, Gradient Boosting, CatBoost, Random Forest, and LightGBM. Among these, XGBoost achieves the highest performance, attaining an accuracy of 98.39% and an AUC of 99.94% when applied to the EEG–EOG–snoring dataset enhanced through GNDA and AES. These exceptionally high performance metrics are attributable to three key synergistic factors: (1) the multimodal fusion of complementary electrophysiological and acoustic modalities; (2) the application of AES for eliminating redundant and noisy features; and (3) the use of GNDA to address class imbalance. To rigorously validate these results, we performed 5-fold stratified cross-validation with statistical significance testing (paired t-test p < 0.001, Wilcoxon signed-rank test p < 0.01). Comparative experiments demonstrate that integrating snoring features with EEG and EOG signals significantly improves classification accuracy over unimodal configurations, with the multimodal approach achieving 6.05% higher accuracy than EEG-EOG only. Feature importance analysis reveals that snoring-derived features, particularly MFCC coefficients and spectral centroid, contribute significantly to N3 and REM stage classification. Overall, the proposed framework offers a reliable and high-performing solution for automated sleep stage classification.</p>

Wavelet-Based Feature Extraction and Autoencoder Compression for EEG–EOG–Snoring Sleep Stage Classification

期刊论文
发表日期：

2026年6月15日

查看全文

Annpurna Singh,
Ratnesh Prasad Srivastava,
Pooja Thawait,
Reetu Rani Pradhan

摘要

Precise sleep stage classification is vital for advancing sleep research and enhancing the clinical assessment of sleep-related disorders. In this study, a comprehensive and scalable machine learning framework is introduced for automated sleep staging, leveraging polysomnographic (PSG) data that incorporates electroencephalogram (EEG), electrooculogram (EOG), and snoring-based acoustic features. The methodological novelty of this work lies in three distinct aspects: (1) the first integration of snoring-derived acoustic features with traditional EEG and EOG signals within an autoencoder-based feature selection framework for sleep staging; (2) a novel application of Autoencoder-Based Feature Selection (AES) that compresses multimodal features into a discriminative latent space while preserving inter-class separability; and (3) a systematic comparative analysis of five state-of-the-art ensemble methods (XGBoost, LightGBM, CatBoost, Gradient Boosting, Random Forest) under identical preprocessing and feature selection conditions. After the initial extraction and preprocessing of signals, a combined feature extraction strategy is utilized, drawing on both statistical metrics and wavelet-based analysis with the Daubechies-4 (db4) wavelet. To counteract class imbalance and bolster model generalization, Gaussian Noise Data Augmentation (GNDA) is applied exclusively to the training data after stratified train-test splitting (80-20 ratio), with augmentation parameters (noise standard deviation \(\sigma = 0.05\) ) optimized through validation to ensure no data leakage occurs between training and test sets. Autoencoder-Based Feature Selection (AES) is then used to compress the feature space and retain the most informative attributes from the augmented training data, reducing dimensionality from 124 to 64 latent dimensions while preserving 99.2% of discriminative information. For classification, five machine learning models are employed: XGBoost, Gradient Boosting, CatBoost, Random Forest, and LightGBM. Among these, XGBoost achieves the highest performance, attaining an accuracy of 98.39% and an AUC of 99.94% when applied to the EEG–EOG–snoring dataset enhanced through GNDA and AES. These exceptionally high performance metrics are attributable to three key synergistic factors: (1) the multimodal fusion of complementary electrophysiological and acoustic modalities; (2) the application of AES for eliminating redundant and noisy features; and (3) the use of GNDA to address class imbalance. To rigorously validate these results, we performed 5-fold stratified cross-validation with statistical significance testing (paired t-test p < 0.001, Wilcoxon signed-rank test p < 0.01). Comparative experiments demonstrate that integrating snoring features with EEG and EOG signals significantly improves classification accuracy over unimodal configurations, with the multimodal approach achieving 6.05% higher accuracy than EEG-EOG only. Feature importance analysis reveals that snoring-derived features, particularly MFCC coefficients and spectral centroid, contribute significantly to N3 and REM stage classification. Overall, the proposed framework offers a reliable and high-performing solution for automated sleep stage classification.