Early detection of mental health disorders through social media analysis presents significant potential but faces challenges in model interpretability for clinical adoption. This study proposes a hybrid explainable AI framework combining bidirectional transformer architectures with clinician validation to detect depression and anxiety from social media text. Our approach integrates a RoBERTa-based language model with temporal BiLSTM layers to analyze both semantic content and behavioral patterns across 112,354 Reddit posts. The system employs SHAP values for feature importance visualization and achieves an F1-score of 0.89 in cross-validated trials, outperforming existing black-box models by 12%. Clinical validation with psychiatric experts demonstrates 82% agreement in risk classification, particularly for early-stage symptom identification. The framework addresses critical ethical concerns through differential privacy preservation and bias mitigation techniques, maintaining 94% accuracy while reducing demographic bias by 37% compared to baseline models. These results are consistent with recent advances in explainable AI for mental health detection [1, 2], suggesting that explainable hybrid architectures can bridge the gap between AI performance and clinical utility in mental healthcare applications.