Nonlinear Machine Learning in Warfarin Dose Prediction: Insights from Contemporary Modelling Studies

Area of Science:

• Pharmacogenomics
• Machine Learning in Medicine
• Clinical Prediction Models

Background:

• Warfarin dosing is complex, influenced by genetic and clinical factors.
• Nonlinear machine learning (ML) algorithms show promise for improving warfarin dose prediction.
• Systematic evaluation of existing ML models for warfarin dosing is needed.

Purpose of the Study:

• To systematically assess the characteristics of studies using nonlinear ML for warfarin dose prediction.
• To evaluate the risk of bias and methodological quality of these studies.

Main Methods:

• Systematic literature search across multiple databases (PubMed, Embase, CNKI, etc.) up to March 2022.
• Assessment of study characteristics: participants, predictors, model development, and evaluation.
• Risk of bias and methodological quality evaluation using the Prediction model Risk of Bias Assessment Tool (PROBAST).

Main Results:

• 23 studies were included, all retrospective, with 11 focusing on Asian populations.
• Common predictors included age, weight, height, amiodarone, CYP2C9, and VKORC1.
• All studies exhibited a high risk of bias due to factors like small sample size, poor data handling, and inappropriate participant exclusion.

Conclusions:

• Most nonlinear ML-based warfarin prediction models suffer from poor methodological quality and high risk of bias.
• Lack of external validity and model reproducibility are significant limitations.
• Future studies should prioritize enhancing external validity, reducing bias, and improving clinical relevance.