Development and Evaluation of a Two-Stage Machine Learning Multivariable Model to Support Clinical Decision-Making in...

Thrombophilia is a disease characterized by an increased risk of venous thromboembolism. Diagnosing and treating thrombophilic patients requires the nuanced interpretation of a great number of medical data. In this single-center, retrospective study, we evaluated the use of a two-stage machine learning (ML) model to predict the thrombophilia risk category and subsequent anticoagulation recommendation. The model was trained and tested using clinical and laboratory data obtained from 496 patients who underwent a complete thrombophilia work-up at the University Hospital Bonn between 2019 and 2024. The hybrid nature of the diagnostic categories, which combine ordinal classes including thrombophilia diagnosis, severity, and treatment recommendations and categorical classes (e.g. antiphospholipid syndrome), necessitated a two-stage approach. First, XGBClassifier was used to distinguish between the two classes, and then an ordinal-specific XGBOrdinalV2 model was used. Performance was assessed across 100 random 80/20 train-test splits. For exact class prediction (primary analysis), mean macro-averaged sensitivity, specificity, and precision were 58%, 91%, and 56.2% for diagnosis and 54.4%, 91.3%, and 58.5% for treatment. Performance was highest for frequent classes and poor for rare high-risk classes. When predictions within one adjacent class were accepted (secondary analysis), macro-averaged sensitivity was 91.9% and 86.2%, respectively. Antiphospholipid antibody status, age at first thrombosis, and unprovoked thrombosis were the most influential features for diagnosis. Model predictions and physician decisions were concordant in 395 of 496 patients (79.6%). This study demonstrates that ML models can support standardized thrombophilia management and improve diagnostic accuracy. However, external and prospective validation is required before clinical use.