Resting-State EEG Features for Distinguishing VS/UWS from MCS: A Comprehensive Univariate Evaluation with Exploratory Multivariate Analysis

Background/Objectives: Consciousness assessment in Disorders of Consciousness (DoC) lacks reliable biomarkers. The relative utility of different EEG feature families for distinguishing Vegetative State/Unresponsive Wakefulness Syndrome (VS/UWS) from Minimally Conscious State (MCS) remains unclear. This study systematically compares their discriminative capacity. Methods: Resting-state EEG was recorded from 53 VS/UWSs, 55 MCSs, and 48 healthy controls (HCs). Ten feature families were extracted: spectral power/ratios, Mean Microstate Duration (MMD), Microstate frequency of occurrence (MFO), Microstate coverage (MC), microstate transitions, nonlinear complexity, weighted Phase-Lag Index (wPLI) connectivity, Autocorrelation Window (ACW), and aperiodic 1/f parameters. Features were screened for between-group differences with monotonic trends across VS/UWS, MCS, and HC, correlated with Coma Recovery Scale-Revised (CRS-R) scores. To obtain unbiased performance estimates and address potential circularity, an exploratory Sparse Group Lasso (SGL) analysis was embedded within a strict nested cross-validation (CV) framework, where all univariate screening, standardization, and hyperparameter tuning were performed exclusively within each outer training fold. Results: Of 101 initial features, 35 wPLI-based metrics were excluded after quality control, leaving 66; of these, 29 showed significant monotonic between-group differences. Partial correlations suggested EEG–CRS-R associations were primarily driven by between-group differences. Microstate transition probabilities showed the strongest univariate discrimination, with the highest AUC for D to A (0.708; sensitivity 60.0%, specificity 77.4%). In the nested CV, the SGL linear predictor achieved a mean outer-fold AUC of 0.632 ± 0.107 (median 0.636) across 37 valid folds, and a nested stable-feature logistic regression yielded a comparable AUC of 0.638 ± 0.099. Notably, family-level selection was highly consistent: Complexity and Microstate Transition families were selected in 74% of outer training folds, followed by MFO (70%) and MMD (70%). Conclusions: Although individual EEG features showed limited discriminative power, and the nested-CV multivariate model did not outperform the best single-feature AUC (0.708), SGL analysis within a nested CV framework identified microstate dynamics and nonlinear complexity as the most robust feature families for distinguishing VS/UWS from MCS. Furthermore, partial correlation analyses indicated that EEG–CRS-R associations were driven by between-group differences, positioning this framework as state-discriminative rather than recovery-monitoring. These findings are exploratory and require further validation in independent external cohorts.