BackgroundObstructive jaundice is a common problem associated with diverse etiologies which has not been thoroughly investigated in large-scale cohorts. Our study involved the largest retrospective cohort of obstructive jaundice to date, exploring the spectrum of diseases while establishing a diagnostic system with machine learning (ML) methods based on routine laboratory tests.MethodsThis study involves two retrospective observational cohorts from China. The biliary surgery cohort (BS cohort, n=349) served for initial data exploration and external validation of ML models, while the large general cohort (LG cohort, n=5726) enabled comprehensive data analysis and ML model construction. Interpretable ML techniques were employed to derive insights from the models.ResultsThe LG cohort exhibited a more diverse disease spectrum compared to the BS cohort, with pancreatic adenocarcinoma, common bile duct stones, distal cholangiocarcinoma, perihilar cholangiocarcinoma, and acute pancreatitis (non-calculous) identified as the top five causes of obstructive jaundice. Traditional serum markers such as CA 19-9 and CEA did not emerge as standalone diagnostic markers for obstructive jaundice. Leveraging ML techniques, we developed two models collectively named as the MOLT model: one effectively distinguishes between benign and malignant causes (AUROC=0.862), while the other provides nuanced insights by further categorizing malignancies into three tiers and benign diseases into two (ACC=0.777). Interpretable ML tools revealed key features contributing to the decision-making process of each model.ConclusionsThrough our study, we uncovered the diagnostic potential of routine laboratory tests in obstructive jaundice, enabling the development of a practical diagnostic tool based on interpretable ML models. These findings may pave the way for personalized and user-friendly diagnosis of obstructive jaundice, thereby aiding clinical decision-making.