Answers are grounded in this paper and cite their sources.
Background Identifying objective EEG biomarkers for Major Depressive Disorder (MDD) remains a challenge due to poor model robustness and a lack of clarity regarding which neural parameters are sufficiently stable for clinical translation. Objective This study deconstructs the MDD neuro-signature to evaluate the critical tradeoff between predictive power and biomarker stability. Methods Using EEG data from 30 MDD patients and 88 controls, we constructed a high-dimensional feature space comprising connectivity, spectral, and complexity measures. The investigation was multi-stage and a leakage-free cross-validation framework. The feature space was decomposed by modality and frequency band using a non-linear XGBoost pipeline to determine the best sources of predictive information. We compared a model utilizing fold-specific top features against one based on stable biomarkers discovered via a strictly nested, leakage-free Stability-Optimized Consensus Learning (SOCL) framework. Findings The strongest predictive signals were localized to Theta (4-8 Hz) and Alpha (8-13 Hz) bands (max AUC = 0.683). However, these spectral features were highly inconsistent across data subsets (mean Jaccard Index = 0.187), resulting in substantial overfitting. In contrast, the model trained on stable biomarkers, selected by a leakage-free consensus of linear and non-linear feature selection, achieved superior generalization performance (AUC = 0.792 vs 0.662) and showed significantly less overfitting ( P < .05 via permutation test). Conclusion The MDD neuro-signature exists in two forms: a potent but unstable signal associated with Theta–Alpha dysregulation and a weaker but highly stable signal embedded within the functional connectome. These findings suggest that biomarker stability is more important than signal strength for developing clinically reliable diagnostic tools.
No takes yet. Share an insight, caveat, or question.
Khemchandani et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: