<p>Liver disease remains a significant global health concern, requiring early and accurate diagnosis to improve patient outcomes. Traditional diagnostic methods often involve invasive procedures, delays, and subjective interpretation, limiting their effectiveness. Machine Learning (ML) and Deep Learning (DL) offer promising alternatives by improving diagnostic accuracy and accessibility. This study proposes a comprehensive comparative framework integrating classical ML classifiers such as Support Vector Machine, K-Nearest Neighbors, Decision Tree, Logistic Regression, and Random Forest, with DL models including Multilayer Perceptron, Long Short-Term Memory, Recurrent Neural Network, Convolutional Neural Network, and Deep Neural Network for liver disease prediction. Unlike prior studies that rely on benchmark datasets such as ILPD or BUPA, we utilize a real-world, multi-source dataset collected from several hospitals across Bangladesh. This region-specific dataset captures unique demographic, genetic, dietary, and healthcare characteristics, significantly enhancing clinical relevance and local applicability. To ensure robustness, feature selection was performed using Pearson correlation, and class imbalance was addressed using the Synthetic Minority Over-Sampling Technique (SMOTE). Ablation analyses on SMOTE and outlier detection were conducted to explicitly evaluate their impact on classification performance. A tenfold cross-validation framework was employed, with expanded evaluation metrics (ROC-AUC, PR-AUC, F1, class-wise scores, and error bars) ensuring rigorous model assessment. Among the models, Random Forest achieved the best performance, with an accuracy of 80.6% and recall of 80.6% on the nine-feature dataset, and an accuracy of 76.6% with precision and recall of 76.7% and 76.6%, respectively, on the seven-feature dataset. Sensitivity analysis further identified Alkaline Phosphatase as the most influential biomarker, while gender showed minimal impact. Statistical significance testing via the Wilcoxon signed-rank test validated that Random Forest outperformed most models (p &lt; 0.05). The novelty of this research lies in (i) developing and validating predictive models on a region-specific, real-world Bangladeshi dataset, (ii) designing a comprehensive comparative ML–DL framework for liver disease prediction, (iii) introducing robust experimental analyses including ablation, significance testing, and interpretability, and (iv) highlighting deployment potential in low-resource healthcare systems. Together, these contributions deliver a reliable, interpretable, and clinically actionable solution for liver disease prediction tailored to the Bangladeshi healthcare context.</p>