Assessing Disorders of Consciousness Using Temporal Sleep Dynamics Extracted From Whole-Night PSG.

Area of Science:

• Neuroscience
• Medical Technology
• Machine Learning

Background:

• Accurate assessment of patients with disorders of consciousness (DoC) is difficult due to limitations in behavior-based evaluations and task-dependent neurophysiological methods.
• Whole-night polysomnography (PSG), a passive and noninvasive monitoring technique, shows potential for assessing residual brain function during sleep in DoC patients.

Purpose of the Study:

• To develop and validate a machine learning framework using temporal-dynamic feature extraction and aggregation from whole-night PSG signals for improved diagnosis and prognosis in DoC patients.
• To enable automated, task-free assessment tools for disorders of consciousness.

Main Methods:

• Whole-night EEG/EOG signals were analyzed using a temporal-dynamic feature extraction and aggregation framework.
• Features were extracted from 30-second epochs, including time-domain, spectral, and nonlinear complexity metrics.
• Sleep staging and unsupervised clustering were used for feature aggregation, followed by a two-stage feature selection process.

Main Results:

• Machine learning classifiers trained on aggregated PSG features achieved high accuracy in distinguishing minimally conscious state (MCS) from vegetative state (VS) (AUC > 0.84).
• The framework demonstrated robust predictive ability for long-term recovery outcomes in DoC patients (AUC = 0.79).

Conclusions:

• Whole-night polysomnography (PSG) analysis using a temporal-dynamic machine learning framework offers significant diagnostic and prognostic value for patients with disorders of consciousness.
• This approach supports the development of automated, non-invasive, and task-free assessment tools for DoC, improving clinical decision-making.