Machine learning based prediction of stable warfarin dose in patients undergoing cardiac surgery

Y. Kang¹, J. Kim², S. Sohn¹, J. Choi², H. Hwang², K. Kim^2
1Department of Thoracic and Cardiovascular Surgery, Seoul National University Hospital, SEOUL, Seoul-t'ukpyolsi ²Seoul National University Hospital, Seoul, Seoul-t'ukpyolsi

Purpose: Warfarin dose adjustment requires individualized dosing due to its narrow therapeutic range and individual variation. Previous studies have utilized multiple linear regression models to predict warfarin dosing. However, considering the complexity of warfarin metabolism, machine learning algorithms in pharmacogenetic warfarin dosing can be a powerful tool for predicting warfarin doses.

Methods: We analyzed a cohort of 380 patients who received stable warfarin doses at our institution after cardiac surgery. The study included all demographic data, including genotypes related to warfarin metabolism (VKORC1 and CYP2C9). We utilized both a Multiple Linear Regression model (MLR) and five machine learning algorithms (Extreme Gradient Boost (XGBoost), Neural network (NN), Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbor (KNN)) to build predictive models. To validate the models, we divided the dataset into a training set (n=266) and a test set (n=114) using stratified random sampling with a 7:3 allocation ratio. Stable warfarin doses were predicted using these models, and their performance was evaluated based on mean absolute error (MAE), root mean squared error (RMSE), and R-square value.

Results: The conventional MLR model showed a high MAE of 0.892 mg/day and RMSE of 1.058 mg/day, with a low R-square value of 0.612. In contrast, the XGBoost model demonstrated an MAE value of 0.402 mg/day and RMSE value of 0.579 mg/day, with an improved R-square value of 0.813. Similarly, the RF model also exhibited a low MAE value of 0.332 mg/day and RMSE of 0.530 mg/day, with a high R-square value of 0.813. The important predictor variables in both the XGBoost and RF models included VKORC1 genotype (A/A), age, height, weight, and body mass index. Notably, patients with the VKORC1 A/A genotype had a mean warfarin dose of 2.9 mg/day, whereas those with non-A/A genotypes required a higher mean dose of 6.3 mg/day.

Conclusion: We have successfully developed an effective machine learning-based prediction model for determining stable warfarin doses in patients undergoing cardiac surgery. While external validation is necessary, the results of this study open the possibility of individualized warfarin dose adjustment in cardiac surgery patients through a machine learning approach. This advancement can lead to optimized warfarin dosing by clinicians.

Identify the source of the funding for this research project: None