We evaluate fifteen techniques from sixteen papers on warfarin dose prediction, highlighting artificial neural networks (ANNs), support vector regression (SVR), boosted regression trees (BRTs), and random forest regression (RFR) as promising avenues for further research. We also investigate the use of pharmacogenetic factors in model development, finding that they may not be necessary for high accuracy, and suggest four avenues for novel research in this field.
This project had the goal of investigating avenues for improved warfarin dosing in South African patients through the use of machine learning. To achieve this goal, two research aims were declared. The first was to determine whether training models on local data could lead to tangible benefits for South African clinicians and patients. The seconds was to compare the performance of “wide” and “deep” algorithms for warfarin dosing.
This study evaluated the accuracy of 17 learning algorithms on both a South African warfarin dataset and the international IWPC dataset. The first 10 algorithms used default or manually-optimised hyperparameters, but the remaining 7 algorithms were developed using genetic programming. These automated algorithms produced the most accurate models and outperformed the best published results in this field. This study also examined the effects of parameter sets and missing data treatments on model accuracy, which informed guidelines on how to implement dosing models in a South African clinical context.
Supplementary content, including raw data, more graphics, and details of some techniques.