Evaluating warfarin dosing models on multiple datasets with a novel software framework and evolutionary optimisation.

Area of Science:

• Pharmacogenomics and Computational Biology
• Clinical Pharmacology

Background:

• Warfarin is crucial for preventing thromboembolism but requires precise, individualized dosing due to its narrow therapeutic index and significant inter-individual variability.
• Machine learning (ML) has shown promise in optimizing warfarin dosage, yet performance varies across algorithms and datasets.

Purpose of the Study:

• To evaluate and compare the accuracy of various ML algorithms for predicting warfarin dosage.
• To introduce and assess genetic programming for automated model optimization in warfarin dosing.
• To present a novel software framework, Warfit-learn, to advance warfarin dosing research.

Main Methods:

• Evaluated ML algorithms, including support vectors, linear regression, neural networks, and stacked ensembles, on the International Warfarin Pharmacogenetics Consortium dataset and a South African clinical dataset.
• Employed genetic programming for automated optimization of model architectures and hyperparameters.
• Developed the Warfit-learn software framework incorporating advanced preprocessing, imputation, and parallel evaluation techniques.

Main Results:

• Support vector machines and linear regression demonstrated top performance, comparable to stacked ensembles, across both datasets.
• Neural networks performed poorly in both datasets.
• Genetic programming-generated models achieved performance levels matching those of expert-designed models.

Conclusions:

• Specific ML algorithms like support vectors and linear regression are highly accurate for warfarin dosing prediction.
• Automated optimization via genetic programming can yield expert-level predictive models.
• The Warfit-learn framework offers a robust platform to accelerate reproducible research in warfarin pharmacogenetics.