Data driven modeling of nutritional profiles for caloric prediction using advanced machine learning techniques

Caloric estimation remains a cornerstone of nutritional modeling, yet formula‑based methods often fail to capture nonlinear interactions among nutrients that influence real energy yield. This study aimed to develop and evaluate machine learning models capable of accurately predicting food caloric content from standardized nutrient profiles, thereby improving estimation precision for diverse food types. A dataset of 410 food items from the USDA FoodData Central was standardized and examined for outliers using leverage statistics, ensuring uniform predictor distribution and data integrity. Seven nutritional variables, including protein, fat, carbohydrates, sugar, fiber, sodium, and potassium, served as predictors of caloric content. Six algorithms were trained and validated: Decision Tree, AdaBoost, Random Forest, Support Vector Regression, and Multi‑Layer Perceptron, with model optimization via 5‑fold cross‑validation. Among tested algorithms, the MLP achieved the highest coefficient of determination (R²≈0.996) reflecting the strong deterministic relationship between nutrient composition and calories within the standardized USDA dataset. SHAP analysis identified carbohydrates, sugar, fat, and protein as the most influential predictors, consistent with physiological expectations. The findings demonstrate that data‑driven models can replicate theoretical caloric relationships with high fidelity while capturing minor nonlinear effects absent in conventional methods. This work contributes a transparent and reproducible framework for machine learning‑based caloric estimation and underscores the potential of deep learning architectures for enhanced nutritional prediction.