Explainable AI for NAFLD Prediction Using Clinical and Lifestyle Data

Non-Alcoholic Fatty Liver Disease is an emerging yet often it neglected health problem closely associated with lifestyle and metabolic conditions. In this paper, we present a machine learning framework for the early detection of NAFLD using simple, non-invasive clinical and lifestyle data from the NHANES 2017-2020 dataset. For feature selection, we employed a hybrid feature selection methodology, Principal Component Analysis combined with Information Gain (PCA-IG), for identifying key factors-smoking status, gender, triglyceride level, and liver stiffness-as relevant. We utilized KMeans-SMOTE to address class imbalance. This greatly improved the accuracy of the models. Various machine learning models were utilized, such as Random Forest, SVM, KNN, and Artificial Neural Networks. The performance was improved for our machine learning models by the use of hyperparameter tuning. To understand model predictions and identify the key risk and protective factors, Explainable AI techniques were employed, namely LIME. The best performance from our models was exerted by the Random Forest model with KMeans-SMOTE, with an accuracy of 98.35%, an F1 score of 98.64%, and an AUC of 1.0. This work not only identifies critical NAFLD risk factors but also offers a dependable and interpretable tool that will help in early diagnosis and guide public health interventions.