An adaptive machine learning framework integrating large language models to assess and enhance emotional intelligence in adolescents.

BACKGROUND: Emotional intelligence (EI) is a critical determinant of children's socio-emotional and neurocognitive development, and deficits in EI dimensions are associated with long-term psychological and behavioral risks. Existing interventions are often static, resource-intensive, and lack personalization, interpretability and reasoning beneath the suggestions. Scalable, explainable AI approaches may help to deliver individualized and effective EI-enhancing activities. METHODS: A total of 120 children (aged 3-16 years) were recruited from the Asian region with informed consent and PGRC ethical approval (CS-00411). EI was systematically assessed using the Strengths and Difficulties Questionnaire (SDQ) to identify weaker dimensions. A stacked ensemble machine learning framework then predicted suitable accurate EI-enhancing activities by incorporating child age, personal interests, and identified weak dimensions. Before generating the final recommendation, the system cross-checked highly rated activities from children with similar profiles to ensure relevance. The activity repository combined validated sources with expert-verified, LLM-generated activities, each mapped to underlying neurocognitive functions. SHAP explained predictive contributions, while LLMs translated outputs into family-friendly narratives. A reassessment module recalculated post-intervention EI scores and integrated parental feedback to dynamically refine subsequent recommendations. RESULTS: Children in the experimental group demonstrated significant improvement in EI scores (p < 0.001, d = 2.58), with empathy and self-regulation showing the largest gains. Parental satisfaction was also significantly higher (p < 0.001, d = 1.88). Weak EI dimension, Interest, and Age Group emerged as the most influential predictors of activity effectiveness. CONCLUSION: This framework directly enhances children's emotional intelligence while ensuring personalization, transparency, and neurocognitive relevance, offering a sustainable and scalable pathway for precision socio-emotional interventions with potential applications in neuropsychiatric prevention and care.