Most knowledge tracing systems prioritise performance prediction over the delivery of personalised, actionable feedback. This paper proposes SPAR-GNN (Selective Pattern-Aware Reasoning with Graph Neural Networks), a novel framework that combines heterogeneous GNN-based modelling with selectively triggered feedback from a Large Language Model (LLM). SPAR-GNN represents students, problems, skills and virtual pattern nodes that encode latent behavioural traits such as frustration and hint overuse within a unified graph structure. A customised Heterogeneous Graph Layer learns student representations, while the LLM is activated only for at-risk learners identified through behavioural and performance indicators. This reduces computational overhead while enhancing pedagogical relevance. SPAR-GNN consistently outperforms strong baselines across multiple metrics, demonstrating both predictive accuracy and robustness. Observational analysis also reveals interpretable links between behaviours and learning progression, supporting fairness-aware interventions.