Obesity is a growing and persistent challenge to public health, with repercussions affecting various aspects of human health, particularly in the metabolic, physical, and psychosocial domains. This study compares two modeling strategies applied to structured tabular data. Classic supervised models followed a six-step process: Data acquisition, Preprocessing, Feature selection, Model training, Hyperparameter tuning, and Evaluation. For the tabular Deep Learning models, the sequence was simplified to four stages: Data acquisition, Preprocessing, Model training, and Evaluation, omitting class balancing and hyperparameter tuning. The multi-class evaluation showed that TabICL achieved the best overall performance with a macro-F1 score of 0.993, while CatBoost was the best classical model with a macro-F1 of 0.982. In conclusion, the results demonstrated that deep learning tabular architecture achieved higher levels of accuracy with a more compact workflow compared to traditional methods.