Stacking Ensemble Learning Approach for Non-Alcoholic Fatty Liver Disease Identification: Leveraging Explainable Machine Learning for Enhanced Prediction Models

Dolley Srivastava; Himanshu  Pandey; Ambuj Kumar  Agarwal; Richa  Sharma

doi:10.18502/fbt.v13i2.21937

Dolley Srivastava Amity Institute of Information Technology, Amity University, Lucknow, Uttar Pradesh, India
Himanshu Pandey Department of Computer Science and Engineering, Faculty of Engineering and Technology, University of Lucknow, Lucknow, India
Ambuj Kumar Agarwal Department of Computer Science and Engineering, School of Engineering and Technology, Sharda University, Greater Noida, India
Richa Sharma Department of Computer Science & Engineering, Shri Ramswaroop Memorial College of Engineering & Management, Lucknow, India

DOI: https://doi.org/10.18502/fbt.v13i2.21937

Keywords: Machine Learning; Ensemble Learning; Stacking Models; Explainable Artificial Intelligence; Interpretable Machine Learning; Feature Importance.

Abstract

Purpose: The prevalence of Non-Alcoholic Fatty Liver Disease (NAFLD) has significantly increased over the past two decades, becoming a leading cause of liver disease in industrialized nations, particularly among individuals who do not consume alcohol. This study aims to develop an efficient detection method for NAFLD using a stacked ensemble learning approach, which integrates multiple machine learning models to enhance predictive accuracy.

Materials and Methods: The dataset utilized in this research was sourced from an open platform and includes critical attributes such as age, gender, Body Mass Index (BMI), height, time to death or last follow-up, and survival status. We implemented a variety of machine learning algorithms, including XGBoost, CatBoost, Decision Trees, and AdaBoost, within a stacking framework to optimize performance. The proposed methodology involved several steps: data preparation, feature engineering, model training, and evaluation. Additionally, we employed explainable AI techniques to identify the most influential features contributing to NAFLD prediction, thereby enhancing the model's interpretability.

Results: The stacked ensemble model achieved an impressive classification accuracy of 95.9%, outperforming individual models and demonstrating the robustness of the ensemble approach. Confusion Metrics, ROC curves, and Calibration Curves are used to evaluate the proposed approach with state-of-the-art approaches. The suggested stacking methodology demonstrates superior performance in all contexts. When Explainable Machine Learning is applied to the proposed approach, it reveals that NAFLD is more common in middle-aged and elderly individuals, but is also present in younger age groups to some extent. Also, the prevalence of NAFLD is higher in males.

Conclusion: The results underscore the potential of a stacked ensemble approach for clinical applications in NAFLD screening and diagnosis, highlighting its importance in healthcare decision-making. By combining various machine learning techniques, we have developed a reliable and resilient model that improves detection accuracy and offers transparency in its predictions. Combining XGBoost, CatBoost, Decision Tree, and AdaBoost in a stacked ensemble model yielded the best results. The findings also indicate that age and gender are significant predictors of NAFLD, with the model providing valuable insights into the underlying patterns associated with the disease. This research contributes to the growing body of knowledge on machine learning applications in gastroenterology and emphasizes the need for explainable models in clinical settings.