Machine learning in bleeding risk assessment for low-molecular-weight heparin or fondaparinux: a predictive model...

Area of Science:

• Medical Informatics
• Clinical Pharmacology
• Machine Learning in Healthcare

Background:

• No established model exists for predicting bleeding risk in patients using low-molecular-weight heparin (LMWH) or fondaparinux.
• Accurate bleeding risk prediction is crucial for patient safety and treatment optimization.

Purpose of the Study:

• To develop and evaluate machine learning models for predicting short-term bleeding risk in hospitalized patients receiving LMWH or fondaparinux.
• To implement the best-performing model as a clinical decision support tool.

Main Methods:

• Retrospective analysis of 1,691 hospitalized patients treated with LMWH or fondaparinux.
• Development and comparison of seven machine learning algorithms (Logistic Regression, SVM, GBM, NN, XGBoost, AdaBoost, CatBoost).
• LASSO regression was used to identify key predictors for model development.

Main Results:

• The CatBoost model demonstrated the highest discrimination (AUC=0.659) and accuracy (86.0%) in the validation cohort.
• Key predictors included surgical site, INR, hemoglobin, platelet count, renal function, BMI, indication, and comorbidities.
• While models showed strong negative predictive value, positive predictive performance for identifying bleeding events was modest.

Conclusions:

• CatBoost is the optimal model for predicting bleeding risk in this patient population, offering superior discrimination and clinical utility for ruling out bleeding.
• A web-based risk calculator utilizing the CatBoost model has been developed for internal use.
• Prospective multicenter validation is necessary before widespread clinical implementation.