The Prediction Model of Warfarin Individual Maintenance Dose for Patients Undergoing Heart Valve Replacement, Based on the Back Propagation Neural Network.

Background And Objective: Because of the narrow therapeutic window and huge inter-individual variation, the individual precision on anticoagulant therapy of warfarin is challenging. In our study, we aimed to construct a Back Propagation Neural Network (BPNN) model to predict the individual warfarin maintenance dose among Chinese patients who have undergone heart valve replacement, and validate its prediction accuracy.

Methods: In this study, we analyzed 13,639 eligible patients extracted from the Chinese Low Intensity Anticoagulant Therapy after Heart Valve Replacement database, which collected data on patients using warfarin after heart valve replacement from 15 centers all over China. Ten percent of patients who were finally enrolled in the database were used as the external validation, while the remaining were randomly divided into the training and internal validation groups at a ratio of 3:1. Input variables were selected by univariate analysis of the general linear model; 2.0, the mean value of the international normalized ratio (INR) range 1.5-2.5, was used as the mandatory variable. The BPNN model and the multiple linear regression (MLR) model were constructed by the training group and validated through comparisons of the mean absolute error (MAE), mean squared error (MSE), root mean squared error (RMSE), and ideal predicted percentage.

Results: Finally, 10 input variables were selected and a three-layer BPNN model was constructed. In the BPNN model, the value of MAE (0.688 mg/day and 0.740 mg/day in internal and external validation, respectively), MSE (0.580 mg/day and 0.599 mg/day in internal and external validation, respectively), and RMSE (0.761 mg/day and 0.774 mg/day in internal and external validation, respectively) were achieved. Ideal predicted percentages were high in both internal (63.0%) and external validation (59.7%), respectively. Compared with the MLR model, the BPNN model showed a higher ideal prediction percentage in the external validation group (59.7% vs. 56.6%), and showed the best prediction accuracy in the intermediate-dose subgroup (internal validation group: 85.2%; external validation group: 84.7%) and a high predicted percentage in the high-dose subgroup (internal validation group: 36.2%; external validation group: 39.8%), but poor performance in the low-dose subgroup (internal validation group: 0%; external validation group: 0.3%). Meanwhile, the BPNN model showed better ideal prediction percentage in the high-dose group than the MLR model (internal validation: 36.2% vs. 31.6%; external validation: 42.8% vs. 37.8%).

Conclusion: The BPNN model shows promise for predicting the warfarin maintenance dose after heart valve replacement.