Abstract
This paper investigates the application of artificial intelligence and natural language processing techniques for the early detection of critical behavioral risks based on textual data analysis. A transformer-based BERT architecture is employed as the core model, enabling context-aware identification of latent linguistic indicators of psycho-emotional distress. Experimental validation demonstrates high sensitivity of the proposed system and confirms its effectiveness for early warning tasks related to critical behavioral risks. From an engineering perspective, particular attention is given to computational efficiency and the energy-related feasibility of deploying pre-trained models when scaling digital preventive platforms. It is shown that reducing repeated training costs and optimizing inference processes are key factors for integrating intelligent systems into distributed information environments of educational institutions and urban digital infrastructures. The obtained results confirm the potential of the proposed approach for developing energy-efficient and sustainable decision-support systems in the field of psycho-pedagogical support.