Abstract: OBJECTIVE To explore the correlation between CYP2C9*2, CYP2C9*3, CYP4F2, and VKORC1 1173C>T polymorphisms and warfarin maintenance dose, and establish an artificial neural network prediction model for international normalized ratio(INR) values after warfarin administration to improve the accuracy of stable dose prediction. METHODS A retrospective study was conducted by collecting clinical data and warfarin pharmacogenetic data from 214 warfarin-treated patients who achieved a stable anticoagulant state from 2019 to 2021. The impact of clinical factors and various gene phenotypes on the patient's warfarin steady-state dose was analyzed. A machine learning prediction model was established by simulating the input of the patient's warfarin dose to calculate the INR target and predict the steady-state dose. The accuracy of the model was compared with the direct dose prediction method and the multiple regression model. RESULTS The multiple regression model had the highest accuracy rate of 56.4% for predicting the patient's steady state dose in the dataset. The machine learning prediction model had a mean absolute error(MAE) of 0.40 and R² of 0.81 when inputting the steady state dose to predict the INR value. Directly predicting the dose resulted in a MAE of 0.52 and R² of 0.68. After group training, the error rate decreased by 20.4% and the accuracy increased by 7.3%. CONCLUSION The artificial neural network model for predicting INR using simulated input of warfarin dose can more accurately predict patient's steady-state dose, which facilitates individualized dosing and promotes the development of precision medicine.