Dosage Individualization of Warfarin Using Artificial Neural Networks

Our objective was to explore artificial neural networks (ANNs) as a possible tool for dosage individualization of warfarin. Demographic, clinical, and genetic data were gathered from a previously collected cohort of patients with a stable warfarin dosage who were able to achieve an observed international normalized ratio of 2-3. Data from a cohort of 3,415 patients were used to develop an ANN dosing algorithm. Data from another cohort of 856 were used to validate the algorithm. The clinical significance of the ANN dosing algorithm was evaluated by calculating the percentage of patients whose predicted dosage of warfarin was within 20 % of the actual stable therapeutic dose. The clinical significance was also compared with a previously published dosing algorithm. A feed-forward neural network with three layers was able to successfully predict the ideal warfarin dosage in 48 % of the patients. The neural network model explained 48 % and 43 % of the dosage variability observed among patients in the derivation and validation cohorts, respectively. ANN analysis identified several predictors of warfarin dosage including race; age; height; weight; cytochrome P450 (CYP)2C9 genotype; VKORC1 genotype; sulfonamide, azole antifungals, or macrolide administration; carbamazepine, phenytoin, or rifampicin administration; and amiodarone administration. An ANN was applied to develop a warfarin dosing algorithm. The proposed dosing algorithm has the potential to recommend warfarin dosages that are close to the appropriate dosages.