Analyzing Anthropometric Feature Dependency in Obesity Classification Using Feature Ablation

Obesity classification models often rely heavily on anthropometric features, particularly weight and height, because both variables directly contribute to body mass index (BMI) calculation. This study investigates the extent of such dependency using feature ablation and error analysis. Experiments were conducted on the NObeyesdad dataset using Logistic Regression and Random Forest under three feature configurations: (1) baseline using all features, (2) ablation without weight, and (3) ablation without height. Model performance was evaluated using Balanced Accuracy, Macro F1-score, confusion matrices, and class-level analysis. Results show that Random Forest consistently outperformed Logistic Regression across all scenarios and demonstrated greater robustness to feature removal. Under the baseline setting, Random Forest achieved a Balanced Accuracy of 0.90 and a Macro F1-score of 0.90, while Logistic Regression obtained 0.89 for both metrics. Performance degradation became more pronounced when weight was removed compared with height, indicating stronger model dependence on weight-related information. Error analysis further revealed increased confusion among adjacent obesity categories, particularly overweight classes. These findings highlight the importance of feature dependency evaluation to improve robustness and interpretability in obesity classification models.