Machine Learning Approaches for Early Diagnosis of Liver Diseases using Sri Lankan Patient Data

Abstract

The liver is one of the most vital organs in the human body, responsible for essential func tions such as detoxification, metabolism, protein synthesis, and regulation of biochemical processes nec essary for survival. Damage or impaired liver function can therefore lead to serious health complications and significantly reduce life expectancy. In recent decades, the global burden of liver diseases has steadily increased, affecting both developed and developing countries. Liver disease represents a growing health burden in Sri Lanka, where Alcohol-Associated Liver Disease (ALD) and Metabolic Dysfunction-Associated Steatotic Liver Disease (MASLD, formerly NAFLD) are major causes of illness and death. Importance of early liver disease detection in Sri Lanka is emphasized, addressing a critical real-world healthcare challenge through data-driven approaches that enable timely treatment, reduce complications, and improve patient survival.Clinical data were collected from Jaffna Teaching Hospital and Kilinochchi General Hospital, com prising 538 records, including ALD, MASLD, and healthy subjects. Data preprocessing involved advanced techniques such as Multiple Imputation by Chained Equations (MICE) for missing data, outlier removal using Z-score, Interquartile Range (IQR), and K-Nearest Neighbour (KNN) methods, and feature selection with LASSO and Variance Inflation Factor (VIF). To handle class imbalance, Synthetic Minority Oversam pling Technique (SMOTE) was applied. The dataset was divided into 80% training and 20% testing for model validation.Eight machine learning algorithms were evaluated for both binary (healthy vs. liver dis ease) and multi-class (healthy, ALD, MASLD) classification. Random Forest achieved the highest binary accuracy of 99.03%, with precision 100%, recall of 97.96%, F1-score of 98.97%, and AUC-ROC of 99.92% confirming robust performance. In multi-class classification, Decision Tree performed best, achieving 86.4% accuracy with precision of 88.46%, recall of 86.4%, F1-score of 87.23%, and AUC-ROC of 84.94%. These f indings demonstrate that robust preprocessing and appropriate model selection substantially improve early liver disease prediction. Future work will integrate deep learning techniques, particularly Convolutional Neu ral Networks (CNNs) with ultrasound imaging, to enhance diagnostic precision and support non-invasive, region-specific liver disease detection in Sri Lanka.