Depression remains a key global mental health concern, with social media emerging as a promising source for early detection through user-generated content. This study presents a comprehensive comparison between message-level and user-level natural language processing (NLP) approaches for identifying early signs of depression on social media. Using the eRisk 2025 dataset, encompassing over 1.4 million Reddit posts from 3,061 users, we evaluated multiple machine learning and deep learning models, including Logistic Regression, Random Forest, XGBoost, and BERT. Results showed that while message-level models achieved high macro-level accuracy and AUC scores (up to 0.90), they struggled to reliably detect depressive messages, as evidenced by low F1-scores and precision in the depressive class. In contrast, user-level models, which aggregate information across multiple posts per user, demonstrated superior performance in identifying depressive users, with higher recall and F1-scores (up to 0.88 recall). These findings highlight the importance of analytical granularity in mental health detection tasks: user-level approaches offer a more robust and context-aware strategy for early identification of individuals at risk. The study demonstrates the potential of integrating artificial intelligence and NLP for proactive mental health monitoring, while also acknowledging challenges related to class imbalance, generalizability, and clinical validation. Future research should explore hybrid models, multi-platform data integration, real-time systems, and ethical frameworks to enhance practical applicability and societal impact.