[Paper Review] Sequence-based Sleep Stage Classification using Conditional Neural Fields

Intan Nurma Yulita, Mohamad Ivan Fanany+1 more|arXiv (Cornell University)|Oct 6, 2016

EEG and Brain-Computer Interfaces23 references3 citations

TL;DR

This paper proposes a sequence-based sleep stage classification framework using Conditional Neural Fields (CNF) with Deep Belief Networks (DBN) and Fuzzy C-Means (FCM) as feature extractors. It demonstrates that DBN-CNF achieves 99.96% accuracy on a private dataset with moderate training time, outperforming CRF and other variants, while FCM-CNF offers a strong balance of accuracy (82.7%) and speed on the same data, validating sequence modeling and feature extraction choices for sleep staging.

ABSTRACT

Sleep signals from a polysomnographic database are sequences in nature. Commonly employed analysis and classification methods, however, ignored this fact and treated the sleep signals as non-sequence data. Treating the sleep signals as sequences, this paper compared two powerful unsupervised feature extractors and three sequence-based classifiers regarding accuracy and computational (training and testing) time after 10-folds cross-validation. The compared feature extractors are Deep Belief Networks (DBN) and Fuzzy C-Means (FCM) clustering. Whereas the compared sequence-based classifiers are Hidden Markov Models (HMM), Conditional Random Fields (CRF) and its variants, i.e., Hidden-state CRF (HCRF) and Latent-Dynamic CRF (LDCRF); and Conditional Neural Fields (CNF) and its variant (LDCNF). In this study, we use two datasets. The first dataset is an open (public) polysomnographic dataset downloadable from the Internet, while the second dataset is our polysomnographic dataset (also available for download). For the first dataset, the combination of FCM and CNF gives the highest accuracy (96.75\%) with relatively short training time (0.33 hours). For the second dataset, the combination of DBN and CRF gives the accuracy of 99.96\% but with 1.02 hours training time, whereas the combination of DBN and CNF gives slightly less accuracy (99.69\%) but also less computation time (0.89 hours).

Motivation & Objective

• To address the limitation of treating sleep signals as non-sequence data in existing classification methods.
• To evaluate the effectiveness of sequence modeling using Conditional Neural Fields (CNF) and its variants for sleep stage classification.
• To compare DBN and FCM as unsupervised feature extractors in the context of sequential sleep signal classification.
• To assess the trade-off between classification accuracy and computational efficiency (training/testing time) across multiple models and datasets.

Proposed method

• Employed two polysomnographic datasets: one public (St. Vincent’s) and one private (Mitra Keluarga Kemayoran) with five sleep stages: Awake, S1, S2, SWS/S3, and REM.
• Used hand-crafted features extracted from EEG, EOG, and EMG signals to preserve domain-specific patterns.
• Applied Deep Belief Networks (DBN) and Fuzzy C-Means (FCM) clustering as unsupervised feature extractors to generate high-level representations.
• Evaluated sequence-based classifiers: Hidden Markov Models (HMM), Conditional Random Fields (CRF), HCRF, LDCRF, and their neural variants—Conditional Neural Fields (CNF) and LDCNF.
• Conducted 10-fold cross-validation to ensure robust performance estimation across all configurations.
• Optimized hyperparameters such as number of clusters (g), fuzziness (w), and network depth to maximize accuracy and minimize computation time.

Experimental results

Research questions

• RQ1 Does modeling sleep signals as sequences using CNF improve classification accuracy compared to non-sequential methods?
• RQ2 How do DBN and FCM compare as unsupervised feature extractors in terms of accuracy and computational efficiency for sleep stage classification?
• RQ3 Which combination of feature extractor and sequence classifier yields the best trade-off between accuracy and training/testing time?
• RQ4 Is the performance of FCM-CNF robust across different datasets, particularly when DBN-based models are less effective?

Key findings

• On the St. Vincent’s dataset, the FCM-CNF combination achieved the highest accuracy (96.75%) with the shortest training time (0.33 hours), outperforming all other configurations.
• On the Mitra Keluarga Kemayoran dataset, DBN-CRF achieved the highest accuracy (99.96%) but required 1.02 hours of training, while DBN-CNF achieved 99.69% accuracy in 0.89 hours, offering a strong balance of performance and efficiency.
• FCM-CNF achieved an average accuracy of 82.7% on the private dataset with an average computation time of 0.34 hours, demonstrating its viability as a fast and effective alternative to deep learning-based extractors.
• The CNF model consistently outperformed CRF, HCRF, LDCRF, and LDCNF in both accuracy and computational efficiency across both datasets.
• DBN was found to be a superior feature extractor compared to FCM on the private dataset, though FCM showed strong performance on the public dataset, suggesting data-dependent effectiveness.
• The study confirms that treating sleep signals as sequences using CNF significantly enhances classification performance, especially when combined with effective feature extractors like DBN.