arXiv News

First-hand information for everyone

Site

AboutPrivacy PolicyTerms of Use

© 2026 arXiv News

New Benchmark Shows Gradient‑Based Explainable‑AI Methods Yield Fairly Accurate Sleep‑Stage Classifications with CNN‑CRF Models | arXiv News

arXiv News

EnglishJapanese

Switch language

EnglishJapanese
Loading account…

First-hand information for everyone

Back to front page
Artificial IntelligenceEnglishPublished

New Benchmark Shows Gradient‑Based Explainable‑AI Methods Yield Fairly Accurate Sleep‑Stage Classifications with CNN‑CRF Models

November 17, 2026arXiv: 2611.18434v1

Researchers expanded the SleepXAI framework to include five popular gradient‑based explainable‑AI (XAI) techniques—modified Grad‑CAM, Vanilla Saliency, DeepLIFT, SHAP, and Integrated Gradients. They used a convolutional neural network (CNN) encoder that transforms single‑channel EEG epochs into 64‑dimensional vectors, followed by a linear‑chain conditional random field (CRF) that decodes the full sequence of sleep stages.

The CNN‑CRF model was trained with leave‑one‑subject‑out cross‑validation. On the Sleep‑EDF reference dataset it achieved 86.81 % overall accuracy, compared with 84.79 % for a baseline CNN‑CNN model. On the DREAMT dataset, the CNN‑CRF improved accuracy from 74.73 % to 76.10 % for normal subjects, and from 66.00 % to a similar level for mild obstructive sleep apnea. The per‑stage results show consistent gains across wake, N2, N3, and REM, but lower accuracy for the minority N1 stage.

Attribution fidelity was measured by Spearman rank correlation between the XAI maps and clinically defined EEG frequency bands assigned by the American Academy of Sleep Medicine (AASM). The study confirms that higher classification accuracy aligns with stronger correlation to these clinical reference signals.

Limitations include that only gradient‑based explanations were evaluated; approaches such as LIME and attention‑based methods were not considered. Additionally, the explanations were generated on the CNN encoder output, not on the final CRF prediction, so they do not capture the sequence‑level smoothing performed by the CRF. Class imbalance—particularly the low number of N1 and N3 epochs, which account for fewer than 15 % of total epochs—may reduce the reliability of attribution maps for these stages. Finally, the findings are specific to the SleepXAI CNN‑CRF architecture and may not generalise to other sleep‑staging models.

These results demonstrate that carefully chosen gradient‑based XAI methods can provide interpretable sleep‑stage predictions while maintaining high accuracy, but further work is needed to broaden the range of explanations and to address class imbalance.

Source paper

Towards Explainable Sleep Stage Classification using Single-Channel Electroencephalogram

Artificial IntelligenceSignal Processing
View on arXiv