Using Machine Learning to Predict the Effect of Warfarin on Heart Patient
Lara Rennie
Department of Computer Science
University of Canterbury
Abstract
In this study several machine learning approaches were compared to the accuracy of more traditional ways of predicting the effect of a dose of Warfarin, an anticoagulant, in heart-valve transplant patients. The twin motivations for this project derived from its potential contribution to the feeld of time-series machine learning, as well as the medical applications. A new `two-layer' approach was attempted to account for the fact that the Warfarin problem consists of multiple, potentially related data-sets. Many different attribute combinations were attempted to provide the best representation of the data and any temporal patterns observed that could help with prediction. Its value in a medical sense derived from the desirability of an accurate web-based system with which self-management of patients could be facilitated. Machine learning was considered a viable solution to the diffculty of Warfarin dose prediction as machine learning algorithms endeavour to cope in a heuristic manner with problems in real data sets such as non-linearity and noise. When tested on the data of heart-valve transplant patients, it was found that the effect of a Warfarin dosage could be predicted with the most accuracy by machine learning algorithms learning on the history of multiple patients. However, the best performing algorithm and attributes differed from patient to patient, making a one-fits-all solution unlikely. The potential for machine learning solutions to out-perform physicians was demonstrated, meaning further work to increase their accuracy would be recommended in this area.