Agentic-TabDeCap: An Agentic AI Enhanced Multi-Stage Framework for Depression Detection and Support

As the global burden of depression continues to rise, there is an urgent need for efficient and accurate methods for early detection and severity assessment. Agentic-TabDeCap is a novel multi-stage framework that detects depression and estimates its severity. In the first stage, all instances are processed by a TabNet-based classifier, which determines whether an individual is depressed or not and assigns corresponding confidence scores for each case. The DeBERTa-v3 and Capsule Network-based model assesses everyone in the second phase, predicting four categories: no depression, mild, moderate, and severe, thereby allowing the system to correct any misclassifications made in the first stage. To avoid error propagation between stages, a Confidence-Guided Attention-Based Meta-Classifier is introduced as the third stage, merging probabilistic outputs from the two former models to produce the final, robust prediction. The model developed recorded an astonishing 98.21% accuracy, 98.83% precision, 97.47% recall, 97.15% F1-score, and 98.34% AUC. For real-time and empathetic mental health support, an Agentic AI system using Retrieval-Augmented Generation (RAG) is integrated, with a knowledge base embedded via all-MiniLM-L6-v2 and indexed with FAISS, while TinyLlama generates context-aware responses. A prototype web interface has been integrated, demonstrating the feasibility and practical applicability of the complete system, including both depression assessment and human-like supportive responses.