Detecting depression through healthcare technology holds significant potential for enhancing clinical decision-making. Traditional diagnostic approaches such as self-reported surveys often rely on subjective assessments and may fail to capture the nuanced nature of mental health conditions. In this research, we propose a novel interpretable approach for depression detection that integrates both global and local explanation mechanisms. Global explanations are derived using a prototype-based learning method, where representative samples are identified via clustering techniques. Local explanations are generated by highlighting key sentences from user input, guided by attention scores and sentence-level sentiment analysis. We introduce the use of SetFit, a few-shot learning framework, to fine-tune pre-trained sentence transformers for effective classification on the DAIC-WOZ dataset. Our model achieves high classification performance, with an accuracy of 83.33%, precision of 86.67%, recall of 83.33%, and an F1-score of 81.48%. Beyond accuracy, our approach significantly enhances interpretability through clear, human-understandable justifications for predictions. This work aims to bridge the gap between model performance and explainability in mental healthcare applications, offering clinicians not only reliable detection tools but also intuitive insights that support informed and effective clinical decisions.