Stable warfarin dose prediction in sub-Saharan African patients: A machine-learning approach and external validation of a clinical dose–initiation algorithm

Abstract

Warfarin remains the most widely prescribed oral anticoagulant in sub-Saharan Africa. However, because of its narrow therapeutic index, dosing can be challenging. We have therefore (a) evaluated and compared the performance of 21 machine-learning techniques in predicting stable warfarin dose in sub-Saharan Black-African patients and (b) externally validated a previously developed Warfarin Anticoagulation in Patients in Sub-Saharan Africa (War-PATH) clinical dose–initiation algorithm. The development cohort included 364 patients recruited from eight outpatient clinics and hospital departments in Uganda and South Africa (June 2018–July 2019). Validation was conducted using an external validation cohort (270 patients recruited from August 2019 to March 2020 in 12 outpatient clinics and hospital departments). Based on the mean absolute error (MAE; mean of absolute differences between the actual and predicted doses), random forest regression (12.07 mg/week; 95% confidence interval [CI], 10.39–13.76) was the best performing machine-learning technique in the external validation cohort, whereas the worst performing technique was model trees (17.59 mg/week; 95% CI, 15.75–19.43). By comparison, the simple, commonly used regression technique (ordinary least squares) performed similarly to more complex supervised machine-learning techniques and achieved an MAE of 13.01 mg/week (95% CI, 11.45–14.58). In summary, we have demonstrated that simpler regression techniques perform similarly to more complex supervised machine-learning techniques. We have also externally validated our previously developed clinical dose–initiation algorithm, which is being prospectively tested for clinical utility.