Interpretable Multiclass Obesity Classification Using Optimized Logistic Regression Based on Anthropometric and Lifestyle Data

Abstract

Obesity is a global public health challenge associated with increased risks of chronic diseases and significant socioeconomic burdens. Conventional obesity classification relies predominantly on body mass index (BMI), which is static and insufficient to capture the multidimensional nature of lifestyle and behavioral factors. This study aims to develop an adaptive and interpretable machine learning–based framework for multiclass obesity classification that addresses the limitations of BMI-centered approaches. An optimized Logistic Regression model is proposed and evaluated using anthropometric and lifestyle-related features, including dietary habits and physical activity patterns. The methodology involves comprehensive data preprocessing, feature encoding, stratified data splitting, hyperparameter optimization, and performance evaluation using confusion matrix analysis, learning curves, and SHAP-based interpretability. Experimental results demonstrate that the optimized Logistic Regression model achieves a high classification accuracy of 94.26% on the test dataset, accompanied by stable generalization performance, as indicated by a relatively small generalization gap between training and validation data. Learning curve analysis confirms robust learning behavior without significant overfitting, while SHAP analysis reveals that both anthropometric and lifestyle features contribute meaningfully to classification decisions. The findings indicate that Logistic Regression offers a balanced trade-off between predictive performance, generalization ability, and interpretability. This study demonstrates that an interpretable, data-driven machine learning approach can serve as a reliable alternative to conventional obesity classification frameworks and support decision-making in health-related applications.