SleepConFormer: Single-Channel EEG Framework Achieves 91.7% Accuracy for Consciousness Assessment in DOC Patients
A novel machine learning framework called SleepConFormer can distinguish patients in a minimally conscious state (MCS) from those with unresponsive wakefulness syndrome (UWS) with 91.7% accuracy using only a single electroencephalography (EEG) channel, according to a study published in IEEE Transactions on Biomedical Engineering. The framework integrates sleep staging and consciousness assessment into a unified pipeline, offering a potential path toward objective, automated bedside monitoring in neurocritical care.
What SleepConFormer Found
The researchers developed SleepConFormer, a three-component deep learning architecture for processing single-channel EEG signals. The first component, MTERL (multi-task EEG representation learning), is a pretraining module trained on large public sleep datasets to learn generalizable EEG features. The second is a SCE-Transformer module that performs confusion-aware temporal modeling, designed to handle the ambiguous boundary transitions between sleep stages that frequently arise in pathological populations. The third component uses logit space aggregation to produce robust coarse-grained classifications for Wake-Sleep and Wake-NREM-REM staging.
On three public datasets of healthy participants, the model achieved five-class sleep staging accuracy ranging from 84.5% to 87.7%, with a cross-dataset macro F1 score of 78.73%. When tested on a clinical cohort of 24 patients with disorders of consciousness (DOC), SleepConFormer maintained 80.78% accuracy for coarse Wake-Sleep differentiation. The critical clinical finding was the model’s ability to discriminate between MCS and UWS: 91.7% accuracy with an area under the curve (AUC) of 0.846. This performance surpassed single-modal feature-based approaches by 12.5 percentage points.
Why It Matters
Patients with disorders of consciousness present a fundamental challenge for clinicians. By definition, these patients cannot participate in standard behavioral assessments, which remain the primary tool for determining consciousness level. The conventional EEG-based sleep staging in DOC patients relies on manual scoring by trained experts, a time-consuming process that is impractical for continuous bedside monitoring.
Sleep architecture is known to correlate with consciousness level. Healthy sleep includes clear cycling through NREM and REM stages, whereas this architecture degrades along the spectrum from MCS to UWS. Capturing this relationship automatically from a single EEG channel could provide a continuous, objective metric for tracking consciousness over time without requiring patient cooperation or expert manual scoring. A single-channel setup is particularly attractive for clinical deployment because it is simpler to apply, less burdensome for patients, and easier to integrate into existing neurocritical care workflows than high-density EEG arrays.
Limitations
The study has several important limitations. The clinical cohort was small, with only 24 DOC patients. The evaluation used subject-independent internal validation, meaning the model was tested on held-out patients from the same cohort rather than on data from an entirely separate institution. External validation on larger, multi-site cohorts will be essential before the framework can be considered for clinical deployment. Additionally, the 80.78% Wake-Sleep accuracy in the clinical cohort, while promising, leaves room for improvement, and the framework’s performance across different EEG hardware setups and recording environments has not been tested.
Bottom Line
SleepConFormer provides a proof of concept that single-channel EEG, combined with representation learning and transformer-based temporal modeling, can simultaneously perform sleep staging and consciousness assessment in patients with disorders of consciousness. The 12.5 percentage point improvement over single-modal features suggests that integrating sleep staging information into the assessment pipeline adds meaningful discriminative power. If validated in larger, diverse clinical populations, this approach could enable continuous, objective monitoring of consciousness at the bedside using equipment already available in many intensive care units.
Source
Man Li, Xiaoyu Bao, Di Chen, Wei Gao, Pengmin Qin, Xinyi Jin, Xiaochun Yang, Yanbin He, Jiahui Pan, Yuanqing Li. “SleepConFormer: A Single-Channel EEG Framework for Sleep Staging and Consciousness Assessment in Patients with Disorders of Consciousness.” IEEE Transactions on Biomedical Engineering, July 1, 2026. DOI: 10.1109/TBME.2026.3708665. PMID: 42384537.