Enhancing Gradient Boosting Performance for Obesity Prediction Through Body Mass Index Feature Engineering

Tree-based algorithms like Gradient Boosting face architectural limitations in implicitly mapping non-linear feature interactions (such as height-to-weight ratios), reducing accuracy across ambiguous obesity classification boundaries. Adopting a data-centric AI paradigm, this study aims to resolve this by executing a Domain-Knowledge Injection of the Body Mass Index (BMI) feature explicitly during preprocessing. Utilizing a self-reported survey dataset of 2,111 records and 17 attributes, the experiment compares a baseline control framework (without BMI) against the proposed model (with BMI) using a default Gradient Boosting Classifier and Stratified 5-Fold Cross-Validation. The dataset contains seven categories of body weight/obesity. Stratification techniques ensure that each fold maintains the same class ratios or percentages as the original dataset. This prevents bias arising from dominant or underrepresented classes during testing. The results show that global model accuracy surged from 95.04% to 98.34% (a +3.3% absolute increase), with a stable mean cross-validation accuracy of 97.06%. At a granular level, the engineered BMI feature acts as a rigid physical anchor that eliminates the ambiguity of overlapping behavioral data. This study concludes that structural interventions within the data pipeline via domain-informed feature extraction are significantly more effective and computationally efficient for optimizing model accuracy than expanding algorithmic complexity.