An ensemble learning based framework to estimate warfarin maintenance dose with cross-over variables exploration on...

Area of Science:

• Pharmacogenomics and Computational Biology
• Clinical Pharmacology
• Artificial Intelligence in Medicine

Background:

• Warfarin dosing is complex due to narrow therapeutic windows and individual variability.
• Existing machine learning models for warfarin dose prediction are limited by insufficient exploration of variable interactions and inability to handle missing data.

Purpose of the Study:

• To develop a novel method for predicting optimal warfarin maintenance doses.
• To address limitations in current models by incorporating variable interactions and handling missing data.

Main Methods:

• Utilized an observational cohort of 377 patients with 1173 warfarin order events.
• Employed univariate analysis to select significant variables, followed by automated cross-over variable generation.
• Applied ensemble learning (LightGBM) to model incomplete data using selected single and cross-over variables.

Main Results:

• The proposed method achieved an average R² of 75.0% in 5-fold cross-validation.
• Demonstrated superior performance compared to baseline methods, particularly in medium- and high-dose subgroups.
• Outperformed the International Warfarin Pharmacogenomics Consortium (IWPC) dosing prediction model.

Conclusions:

• The novel method effectively explores variable interactions and learns from incomplete data for warfarin dose prediction.
• This approach shows significant promise and warrants further research in clinical applications.