Abstract
Depressive disorder (DD), a leading global cause of disability, lacks objective diagnostic biomarkers due to reliance on subjective clinical criteria. This study introduces an algorithm-driven framework integrating multidimensional EEG features, dynamic time-window optimization, feature selection and machine learning to address this gap. Resting-state EEG signals were acquired from 70 DD patients and 30 healthy controls (HC). Three-dimensional neurophysiological features, including power spectral density (PSD), sample entropy (SE), and phase lag index (PLI), were systematically extracted across variable time windows. The SVM-RFE algorithm eliminated redundant features, identifying an optimal subset that maximized classification accuracy through leave-one-subject-out cross-validation. Our model achieved exceptional classification accuracy of 94.48% using 10-second windows, outperforming conventional approaches. Critical biomarkers included beta rhythm alterations and cross-frequency functional connectivity patterns, demonstrating superior discriminative power for DD patients. The optimal feature subset emphasized the combined significance of spectral, nonlinear dynamic, and network-level characteristics in differentiating DD from HC. This framework establishes the first evidence-based integration of time-window and feature selection optimized multidimensional EEG features for DD identification, resolving key limitations in replicability and clinical translatability of existing methods. Beyond enabling high-precision objective diagnosis, the biomarker profile provides mechanistic insights into DD neuropathology, particularly beta rhythm dysregulation and aberrant cross-frequency coupling. These findings advance EEG-based precision psychiatry by offering a validated protocol for therapeutic monitoring and treatment personalization, bridging the critical gap between computational neuroscience and clinical practice in mood disorder management.