Clinical interviews remain the gold standard for diagnosing depression, yet the global shortage of trained professionals underscores the need for interpretable and automated assessment systems. Current LLM-based approaches often overlook the personalised nature of patient expressions, producing generic analyses that fail to capture individual linguistic and psychological nuances. Moreover, effective diagnosis requires both empathic understanding and clinical abstraction—abilities rarely combined in existing models. We propose PADE (PersonAlisation–DEpersonalisation Framework), which mirrors the reasoning process of human clinicians by bridging personal understanding and clinical interpretation. In the Personalisation stage, Patient-Guided Query Rewriting (PGQR) employs a role-playing LLM to simulate each patient’s behavioural and emotional profile, generating adaptive queries for precise evidence retrieval. In the Depersonalisation stage, Dual-Channel Evidence Elicitation (DCEE) transforms retrieved content into clinically interpretable cues by integrating explicit symptoms and implicit causes. Experiments on real-world depression detection benchmarks demonstrate that PADE outperforms current state-of-the-art methods in both diagnostic accuracy and interpretability, offering a clear, transparent pathway from personalized dialogue to clinical decision-making.