Nutritional interventions play a crucial role in preventing and managing numerous health conditions, with proper nutrient analysis being essential for effective healthcare delivery, especially in vulnerable populations. Despite advances in computational methods, current approaches to nutrient analysis struggle to model complex interactions between nutrients and health outcomes while remaining accessible in diverse clinical settings. Existing systems either priorities predictive performance at the expense of interpretability and efficiency or sacrifice accuracy for simpler models, creating a significant barrier to clinical adoption in resource-constrained environments. To address these challenges, we present NUTRINET, a novel graph neural network that combines hierarchical nutrient graph representation, edge-conditioned message passing, sparse attention mechanisms, and transparent prediction modules. Through evaluations on three public datasets (USDA FoodData Central, NHANES, and Framingham Heart Study) with an 80:20 train-test split, NUTRINET demonstrates superior predictive performance with lower MAE (0.09) for nutrient prediction, higher AUC (0.91) for deficiency risk assessment, and improved F1-score (0.85) for personalized recommendations compared to state-of-the-art methods. Notably, our model reduces energy consumption by up to 73% compared to Graph Attention Networks while providing high-fidelity explanations that influenced clinical intervention decisions in 78% of test cases. These results establish NUTRINET as an effective solution for computational nutrient analysis that balances performance, interpretability, and accessibility for diverse healthcare settings, particularly addressing the specific needs of vulnerable populations through contextually aware nutrient interaction pathways.