<p>Depression detection through computational models has gained traction, especially with advances in natural language processing (NLP). However, existing models face challenges in handling noisy data, linguistic diversity, cultural context, and fine-tuning for optimal performance, particularly for low-resource languages like Marathi. To address these issues, this paper proposes a novel DepreLex-BERT-Att-LSTM framework. This model integrates depression-specific lexicons, BERT embeddings, and an auxiliary attention mechanism with an LSTM classifier for depression detection in text data. The model enhances feature extraction using N-grams and a lexicon tailored to depressive language, while BERT generates rich contextual embeddings. Auxiliary attention, inspired by human brain processing, refines the embeddings by focusing on relevant depressive patterns and mitigating irrelevant information. LSTM is then used to capture temporal dependencies and improve classification performance. This novel combination of deep learning and the attention mechanism offers a robust and scalable solution for accurate depression detection across diverse datasets. The proposed method was evaluated using metrics such as balanced accuracy, precision, recall, and the F1-score. The results demonstrated its superiority over traditional models, by achieving an enhanced performance in detecting depressive tendencies from Marathi text data with precision of 97.40%, precision of 97.73%, recall of 97.40%, and F1-score of 97.40%. The findings suggest that DepreLex-BERT-Att-LSTM is an effective approach for early intervention and mental health support. Such real-time Marathi text-based depression detection system can be applied to social media monitoring, mental health apps, and healthcare platforms to identify early signs of depression. Also, analyzing Marathi texts to provide timely intervention for depression detection will help and provide support to Marathi speaking communities.</p>