Individual nutrition is a key issue in preventive care, but there is no reliable method of risk evaluation in the dieting process since the health data are heterogeneous, classes are disproportionate, and the probability estimates are inaccurate. It is suggested in the study that a hybrid deep learning architecture should be used to predict diet-quality and that includes transformer-based representation learning, gradient- boosted decision trees, latent-space oversampling, and calibrated ensemble modeling. The five actual health samples were combined into a single dataset of 8,737 valid cases of demographics, dietary intake, laboratory biomarkers, examination measurements and lifestyle questionnaires. A masked-feature pretraining plan was applied to acquire contextual nutritional representations, after which the embedding-level augmentation was taken to enhance the sensitivity of the minority-class. FT-Transformer, MLP and XGBoost models were combined using an optimized weighted ensemble with isotonic calibration to make them more reliable. The experimental results show the enhanced minority-class recall, the enhanced probability calibration, and the overall performance as opposed to individual models. These results suggest that the combination of calibrated learning methods with hybrid techniques can assist in efficient dietary risk assessment, as well as be used as a scalable base of individualized nutrition decision support systems.