A Prediction Model of Stable Warfarin Doses in Patients After Mechanical Heart Valve Replacement Based on a Machine Learning Algorithm.

BACKGROUND

The narrow therapeutic range of warfarin, alongside the response of numerous influencing factors and significant inter-individual variability, presents major challenges for personalized medication. This study aimed to combine clinical and genetic characteristics with machine learning (ML) algorithms to develop and validate a model for predicting stable warfarin doses in patients from Northern China after mechanical heart valve replacement surgery.

METHODS

This study included patients who underwent mechanical heart valve replacement surgery at the Beijing Anzhen Hospital between January 2021 and January 2024 and achieved a stable warfarin maintenance dose. Comprehensive clinical and genetic data were collected, and patients were divided into training and validation cohorts at an 8:2 ratio through random division. The variables were selected using analysis of covariance (ANCOVA). Algorithms for predicting the stable warfarin dose were constructed using a traditional linear model, general linear model (GLM), and 10 ML algorithms. The performance of these algorithms was evaluated and compared using R-squared (R), mean absolute error (MAE), and ideal prediction percentage to identify the optimal algorithm for predicting the stable warfarin dose and verify its clinical significance.

RESULTS

A total of 413 patients were included in this study for model training and validation, and 13 important features were selected for model development. The support vector machine radial basis function (SVM Radial) algorithm showed the best performance of all models, with the highest R value of 0.98 and the lowest MAE of 0.14 mg/day (95% confidence interval (CI): 0.11-0.17). This model successfully predicted the ideal warfarin dose in 93.83% of patients, with the highest ideal prediction percentage found in the medium-dose group (95.92%). In addition, the model demonstrated high predictive accuracy in both the low-dose and high-dose groups, with ideal prediction percentages of 85.71% and 92.00%, respectively.

CONCLUSIONS

Compared to previous methods, SVM Radial demonstrates significantly higher accuracy for predicting the warfarin maintenance dose following heart valve replacement surgery, suggesting it has potential for widespread application. However, this study was based on a relatively small sample size and conducted at a single center. Future research should involve larger sample sizes and multicenter data to validate the predictive accuracy of the SVM Radial model further.