Three-Test Liver Disease Screening Using Machine Learning: A Cost-Effective Approach Through Feature Prioritization

Abstract

Standard liver disease screening requires testing eight or more laboratory parameters, which increases costs and limits accessibility in resource-limited settings. This study aimed to identify a minimal test combination that maintains diagnostic accuracy while reducing costs. Seven machine learning algorithms—Support Vector Machine, Boosting, Multilayer Perceptron, Bagging, Random Forest, K-Nearest Neighbors, and J48 Decision Tree—were evaluated using the Indian Liver Patient Dataset (583 patients, 10 laboratory parameters). Features were ranked based on their diagnostic value to identify the most informative tests. Algorithm performance was assessed across six scenarios with progressive feature removal, measuring accuracy, Kappa statistic, ROC area, and F-measure. The Kappa statistic revealed classification issues that accuracy metrics alone masked. Feature ranking identified Total Bilirubin, Direct Bilirubin, and SGPT as the three most informative parameters. Random Forest achieved the best performance using only these three tests (accuracy 72.04%, ROC 0.70, Kappa 0.24), matching the accuracy obtained with all eight parameters. This approach reduces the required tests by 62.5% and screening costs by 60%. Support Vector Machine reached 71% accuracy, but its zero Kappa value indicated failure to learn real diagnostic patterns, instead exploiting the dataset's 71.4% disease rate. Bagging achieved the highest ROC (0.73) but showed lower overall balance than Random Forest. The three-parameter Random Forest model enables accurate and cost-effective liver disease screening, correctly identifying 87% of diseased patients. This approach reduces testing from eight parameters to three without compromising diagnostic accuracy. The feature ranking method can be applied to other diseases where comprehensive testing increases healthcare costs.