SleepXAI study finds Integrated Gradients most faithful for explaining single-channel EEG sleep staging
An arXiv paper extends the SleepXAI framework to compare five explainable AI methods, reporting that Integrated Gradients gave the most faithful attributions across two datasets.
In plain words
Researchers tested ways to make an automated sleep-stage classifier show which parts of a brain-signal recording drove its decisions. Using a single EEG channel, they report that a method called Integrated Gradients produced explanations that best matched clinically defined EEG frequency bands.
What was announced
The authors, writing on arXiv, say automated sleep staging already reaches strong benchmark accuracy but sees limited clinical adoption because its outputs lack interpretability. To address this, they extend the SleepXAI framework to systematically compare five explainable AI (XAI) methods, focusing on single-channel electroencephalogram (EEG), which they argue is more practical to deploy than full polysomnography setups. Attribution fidelity was measured with Spearman rank correlation against AASM-defined EEG frequency bands, treated as clinically grounded reference signals.
Experiments used the Sleep-EDF benchmark and the DREAMT dataset, where subjects were stratified into four Apnea-Hypopnea Index (AHI) severity groups based on obstructive sleep apnea. The team reports that SleepXAI achieved strong classification performance on both datasets, and that Integrated Gradients (IG) consistently achieved the highest mean attribution fidelity across all five sleep stages on Sleep-EDF and across all AHI groups on DREAMT, outperforming the other evaluated methods under the Spearman rank correlation metric. The authors present this as evidence that choosing the attribution method is a key factor for reliable, clinically interpretable automated sleep staging.
What it does not show
- The source is an abstract only; experimental details, exact scores and peer review status are not available in the supplied passage.
- Findings are limited to the two datasets and five XAI methods evaluated; no independent confirmation is provided.
For security and IT teams
- Scope
- The reported results cover single-channel EEG sleep staging on the Sleep-EDF and DREAMT datasets, not full polysomnography deployments.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.