Named Entity Recognition (NER) is a core task in Natural Language Processing (NLP) with applications in conversational AI, healthcare analytics, and search systems. It enables structured extraction of key entities from text, supporting clinical decision-making and automated diagnosis. While NER resources for high-resource languages like English are widely available, low-resource languages such as Marathi lack domain-specific datasets, particularly in the mental health sector. This work introduces ArogyaMIND-NER, a Marathi mental health-specific NER dataset, designed to enhance entity recognition for diseases, symptoms, and treatments. The dataset is manually annotated using BIO tagging, ensuring linguistic accuracy and contextual consistency across formal and conversational text. Annotation policies account for psychiatric terminology nuances, accommodating Marathi’s morphological richness and linguistic diversity. To assess performance, ArogyaMIND-NER is benchmarked using classical machine learning models (SVM, Naïve Bayes, XGBoost, LightGBM, Random Forest), deep learning architectures (CNN, BiLSTM, BiLSTM-CRF), and transformer-based models (mBERT, XLM-RoBERTa, IndicBERT, MahaBERT, MahaMarathi-7B). Comparative evaluation reveals MahaMarathi-7B and Marathi-Social-NER as top-performing models, achieving higher precision and recall in disease recognition, while classical models struggle with contextual understanding. This dataset and trained models establish benchmarks for Marathi NLP research, contributing to mental health analytics, clinical applications, and low-resource language processing. The findings advocate for fine-tuning transformer architectures for domain-specific NER, ensuring better entity identification in Marathi mental health texts. The dataset and models will be made available for further research and development.