Depression is a complex mental health disorder with significant personal and societal impacts. With the growing interest in bio-signals for mental health diagnostics, electroencephalogram (EEG) signals have emerged as a non-invasive, cost-effective, and reliable source for detecting neural abnormalities associated with Major Depressive Disorder (MDD). This study proposes an innovative framework for MDD detection leveraging advanced signal processing and learning-based methods. The methodology integrates Empirical Mode Decomposition (EMD), Hilbert Transform (HT), and Recurrence Quantification Analysis (RQA) to extract dynamic and non-linear features from EEG data, emphasizing the critical role of Instantaneous Phase (IP). Focusing on Instantaneous Phase (IP) rather than amplitude, the approach uncovers subtle neural dysfunction patterns linked to connectivity and coordination. Key RQA metrics-Recurrence Rate (RR), Determinism (DET), and Laminarity (LAM) were analyzed, with EMD improving feature quality. XGBoost achieved 94.74% accuracy, outperforming non-EMD models. SHAP analysis emphasized the importance of temporal, frontal, and central brain regions. These findings demonstrate the potential of phase-based EEG analysis for advancing MDD diagnosis with robust, interpretable results.