A Comparative Analysis of Ensemble Machine Learning Models for Herniated Lumbar Disk Classification

Abstract

Herniated Lumbar Disk (HLD) poses a significant and growing health burden in Nigeria, yet diagnosis is hampered by limited resources. Existing machine learning (ML) solutions are often constrained by class imbalance and the selection of the most effective models. This study attempts to address these gaps by experimenting with four ensemble ML models to determine the most effective for HLD classification. A clinical dataset that contains 310 samples in size was sourced from Kaggle because, to the best of our knowledge, there are no local data currently accessible in Nigeria. A comparative analysis of bagging (Random Forest, Extra Trees) and boosting (AdaBoost, CatBoost) ensemble models was employed in addition to SMOTE, to rectify class imbalance for both binary and multi-class tasks. Results demonstrated strong performance across models. In bagging, Extra Trees excelled, achieving 92.10% accuracy, and a 98.59% ROC-AUC in binary classification. For multi-class, it established 89.50% accuracy and a leading 99.20% ROC-AUC. Among boosting models, CatBoost also performed well with 89.90% binary accuracy and a 98.74% ROC-AUC. The top-performing Extra Trees model was subsequently deployed into a web application to enhance clinical accessibility. The non-Nigerian dataset was limited by generalizability and explainable AI in complex ensemble models, indicating a need for future validation on local datasets and integration of interpretability techniques.