Toward Optimal Heparin Dosing by Comparing Multiple Machine Learning Methods: Retrospective Study.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine
• Pharmacology

Background:

• Heparin is a critical medication in intensive care units (ICUs).
• Current standard practice relies on weight-based heparin dosing nomograms for thrombosis treatment.
• Machine learning (ML) offers potential for advanced clinical decision support in medication management.

Purpose of the Study:

• To predict heparin treatment effects using ML methods.
• To optimize heparin dosing in ICUs based on ML predictions.
• To classify patient states into subtherapeutic, normal therapeutic, and supratherapeutic activated partial thromboplastin time ranges.

Main Methods:

• Retrospective analysis of data from MIMIC-III and eICU databases.
• Comparison of ML models including random forest, SVM, adaptive boosting, XGBoost, and shallow neural networks.
• Evaluation of model performance using precision, recall, F1 score, and accuracy for 3-class patient state prediction.

Main Results:

• Shallow neural network demonstrated superior performance across datasets (F1 scores up to 87.55%).
• Highest F1 scores for the shallow neural network were observed in predicting subtherapeutic, normal therapeutic, and supratherapeutic states.
• Model validation indicated performance exceeding standard heparin dosing practices.

Conclusions:

• A data-driven ML model can effectively predict heparin treatment outcomes.
• ML facilitates optimized and personalized heparin dosing for improved patient safety in ICUs.
• This approach represents a feasible advancement over traditional heparin management strategies.