A machine learning-based framework for predicting metabolic syndrome using serum liver function tests and high-sensitivity C-reactive protein.

Metabolic Syndrome (MetS) comprises a clustering of conditions that significantly increase the risk of heart disease, stroke, and diabetes. Timely detection and intervention are crucial in preventing severe health outcomes. In this study, we implemented a machine learning (ML)-based predictive framework to identify MetS using serum liver function tests-Alanine Transaminase (ALT), Aspartate Aminotransferase (AST), Direct Bilirubin (BIL.D), Total Bilirubin (BIL.T)-and high-sensitivity C-reactive protein (hs-CRP). The framework integrated diverse ML algorithms, including Linear Regression (LR), Decision Trees (DT), Support Vector Machine (SVM), Random Forest (RF), Balanced Bagging (BG), Gradient Boosting (GB), and Convolutional Neural Networks (CNNs). This framework is designed to develop a robust, scalable, and efficient predictive tool. We evaluated our approach on a large-scale cohort comprising 9,704 participants from the Mashhad Stroke and Heart Atherosclerotic Disorder (MASHAD) study, spanning 2010-2020. After preprocessing, a final dataset of 8,972 individuals (3,442 with MetS and 5,530 without) was used for model development and validation. Among the tested models, GB and CNN demonstrated superior performance, achieving specificity rates of 77% and 83%, respectively. The Gradient Boosting model achieved the lowest error rate of 27%, indicating robust predictive capability. Additionally, SHAP analysis identified hs-CRP, BIL.D, ALT, and sex as the most influential predictors of MetS. These findings suggest that leveraging liver function biomarkers and hs-CRP within an automated ML pipeline can facilitate early, non-invasive detection of MetS, supporting clinical decision-making and risk stratification efforts in healthcare systems.