Improving Warfarin dosing by applying machine learning to a small African data set

Snibbe, Claire

Improving Warfarin dosing by applying machine learning to a small African data set

Thesis / Dissertation

2026

Publisher

University of Cape Town

Department

Department of Computer Science

Faculty

Faculty of Science

Abstract

Warfarin remains the mainstay of anticoagulants in sub-Saharan Africa (SSA). The dose of Warfarin is critical to providing a therapeutic effect without the potentially devastating side effects of thrombotic complications such as strokes, Deep Vein Thrombosis or Pulmonary Embolism associated with subtherapeutic dosing or haemorrhaging due to supratherapeutic dosing. This study deploys supervised regression machine learning (ML) models to predict optimal patient dosing of Warfarin (a continuous value expressed in mg per week). These models are evaluated using the metrics- Mean Absolute Error (MAE) and Percentage of Patients with Predicted Dose Within 20% of Actual Dose (PW20). Our data are provided by the War-PATH study, a collaboration to investigate Warfarin dose in black African patients with cardiovascular disease in SSA. Unfortunately, this data set is very small, comprising 364 patients (one record per patient) with a low-dimensional (d=10) feature space. Herein lies a core challenge of our study– ML algorithms are data hungry, typically performing optimally on training sets of abundant examples. Not only is the data set very small, but it also suffers from missing data over multiple features. Multiple Imputation by Chained Equations (MICE) is applied before ML algorithms are trained on the data, but extensive changes to the study are necessary to accommodate the effect of imputed data on model training and evaluation. Earlier work on the War-PATH data set (which involved training 21 ML algorithms) suggested that Linear Regression (LR) was sufficient to model the data distribution. The overarching objective of this study is to determine whether any benefit exists in applying more complex ML models (excluding Deep Learning (DL) models because such models are best trained on very high volume data) to such a limited data set and how to implement this. We explored a range of standard supervised regression ML models- Random Forest (RF), Decision Tree (DTR), XGBoost (XGBR), K-Nearest Neighbours (KNNR), Support Vector (SVREG), Multi-Layer Perceptron (MLPR), Ridge(RR), Least Absolute Shrinkage and Selection Operator (LASSO), ElasticNet (ELNET) and LR, our benchmark model; and various strategies to counter the limitations of the data set size and data imputation. Barring RF and DTR, all trained models outperformed LR with lower MAEs and higher PW20s; MLPR's and SVREG's MAEs were 11.94 compared with LR's MAE of 12.44. Our results thus show that ML has the potential to improve on the less complex LR model and that it is possible to develop ML models that are more accurate than LR even on small, low-dimensional data sets.

Keywords

Warfarin

machine learning

Reference:

Collections

Masters

Full item page