Non-alcoholic fatty liver disease (NAFLD) is a hepatic manifestation of metabolic syndrome, characterized by fat accumulation in the liver among individuals who do not consume excessive alcohol. Over the past three decades, its prevalence has risen globally, posing a significant public health challenge. NAFLD can progress to cirrhosis, liver failure, and an increased risk of cardiovascular disease, ultimately contributing to higher overall mortality. Despite its widespread occurrence, early detection remains a challenge due to limitations in current screening methods. Here, we aimed to develop an AI-driven model for diagnosing NAFLD based on blood parameters and anthropometric indices. This study utilized data from the Persian Fasa cohort, originally comprising ۱۰,۱۳۸ records and ۲۲۶ features, categorized into discrete and continuous features. After preprocessing, normalization, and dimensionality reduction, statistical analyses were conducted using Python. Patients were categorized into three groups based on the Fatty Liver Index (FLI), including healthy (<۳۰), borderline (۳۰–۶۰), and NAFLD (>۶۰). The dataset was divided into training (۷۰%) and testing (۳۰%) subsets. Seven feature selection methods, including ANOVA, Mutual Information (MI), Independent Component Analysis (ICA), Non-negative Matrix Factorization (NMF), Principal Component Analysis (PCA), Penalized Support Vector Machine (SVM_L۱), and Elastic Net Logistic Regression, were applied to extract common features. The Random Forest algorithm identified the most important extracted features, which were validated through Receiver Operating Characteristic (ROC) curve analysis. A variety of machine learning models, including Random Forest, Support Vector Machine (SVM), Logistic Regression, K-Nearest Neighbors (KNN), Decision Tree, CatBoost, AdaBoost, and XGBoost, were trained to evaluate classification performance using a ۵-fold cross-validation approach. Model diversity was assessed using Kappa statistics and error analysis to ensure robustness. To further improve performance, Optimized Weighted Averaging (OWA) and Sugeno Fuzzy Integral methods were applied for model combination. Finally, a Convolutional Neural Network (CNN) was trained with ۵-fold cross-validation to integrate robust models and enhance classification results. The final dataset comprised ۷۰ clinical and lifestyle variables, including hypertension, smoking status, and others, collected from ۱۰,۰۰۷ patients (۴۵.۲% male and ۵۴.۸% female). The number of patients in each category was as follows: healthy (۴,۴۴۴), borderline (۲,۸۹۲), and NAFLD (۲,۶۷۱). Five key features, including BMI, waist-to-hip ratio, triglycerides, and GGT, were identified as the most significant predictors using the Random Forest method. The diagnostic value of these features was confirmed through ROC curve analysis, achieving an Area Under the Curve (AUC) greater than ۰.۷. SVM and CatBoost models demonstrated exceptional performance, with a Kappa score of ۰.۹۶ and an error rate of ۰.۰۱, indicating high model diversity and minimal error. Combining these two models using Sugeno Fuzzy Integral, OWA, and CNN-based meta-learning produced outstanding results: Accuracy ۰.۹۹, Precision ۰.۹۹, Recall ۰.۹۹, F۱ Score ۰.۹۹, and an AUC of ۱.۰۰. By highlighting factors that could improve the diagnosis of NAFLD, we
