Enhancing Model Stability for Obesity Risk Classification Through Targeted Hyperparameter Optimization and Robust Scaling Against Outliers

İngilizce (OpenAlex)

Obesity is a critical public health concern that demands advanced decision-support systems for early diagnosis and risk classification due to its global prevalence and multidimensional health impacts. This study aims to develop a machine learning model capable of classifying obesity levels with high accuracy. In this research, SMOTE and ADASYN techniques were employed to address data imbalance within an obesity dataset comprising 1,610 individuals. To evaluate the impact of scaling, MinMaxScaler, StandardScaler, and RobustScaler normalization techniques were utilized. A significant scientific contribution of this study is an original GridSearchCV-based strategy developed to optimize the performance of seven different classifiers. Instead of conventional broad search spaces, narrow and targeted hyperparameter ranges specific to the architecture of each algorithm were defined. Consequently, the generalization capability and clinical classification success of the proposed machine learning models were maximized. Experimental results indicated that the Extra Trees classifier, when integrated with the proposed targeted strategy and RobustScaler, achieved the peak performance with an accuracy of 96.09%, an F1-score of 96.08%, a Cohen’s Kappa of 0.941, and an MCC of 0.942. Notably, the proposed optimization strategy provided a 2.4% improvement in classification stability compared to baseline models with default parameters. Compared to the baseline models (Model 1 and Model 2), the proposed Model 3 architecture provided a net performance increase of approximately 8.2% in accuracy. Furthermore, the application of SMOTE data augmentation led to a 10% to 15% average improvement across all tested classifiers compared to raw data results. In conclusion, this comprehensive approach—combining data augmentation, extensive normalization, and algorithm-specific targeted hyperparameter optimization—has achieved significant success in classifying obesity phenotypes. The proposed framework holds the potential to serve as a robust and effective clinical tool to support the decision-making processes of healthcare professionals.