This study proposes the approach for early detection of Narcissistic Personality Disorder in online communication texts by leveraging the transformer-based Distilled Bidirectional Encoder Representations from Transformers (DistilBERT) architecture combined with explainable artificial intelligence techniques. A balanced dataset was constructed from social media messages, where the narcissistic disorder target class was extracted from the corpus, while control texts were sourced from other mental disorder categories and messages without clinically relevant signs. Experimental results demonstrate the effectiveness of the proposed method, achieving a classification accuracy of 98%, which outperforms comparable state-of-the-art approaches by at least 5%. High F1, Matthews Correlation Coefficient (MCC), and Area Under the Curve (AUC) values confirm the model’s ability to accurately classify the target category and differentiate it from related psychopathologies. The integration of the Local Interpretable Model-agnostic Explanations (LIME) framework ensured transparency and interpretability of model decisions, allowing expert validation and enhancing clinical relevance. The scientific novelty of this work lies in combining a Bidirectional Encoder Representations from Transformers (BERT)-like architecture with local decision interpretability in a task characterized by limited data availability for the target class. The findings demonstrate that even a small number of relevant examples can be sufficient for effective fine-tuning using transfer learning (TL) when supported by a balanced dataset. Practically, the proposed system can serve as a tool for preliminary screening, risk group identification, and decision support in mental health contexts.