Systematic review and meta-analysis of AI accuracy in warfarin dose prediction across ethnic groups.

PURPOSE: The primary purpose of this study is to systematically evaluate how accurately artificial intelligence (AI) models can predict optimal warfarin dosing by incorporating both genetic variations-particularly in VKORC1 and CYP2C9-and clinical parameters.

METHODS

We searched PubMed, Scopus, and the Cochrane Library from inception until November 2024 for studies using machine learning to estimate warfarin dosing. Mean absolute error (MAE) was the primary outcome. Subgroup analyses were conducted by ethnicity. A random-effects model was used in R Software.

RESULTS

Seventeen studies involving 50,859 patients were included. The pooled mean absolute error (MAE) using a random-effects model was 7.10 (95% CI: 5.52, 8.67; p < 0.001). Subgroup analysis by ethnicity revealed varying performance of AI-based warfarin dosing models. For the Asian population (4 studies), the pooled MAE was 4.45 (95% CI: 2.92, 5.97; p = 0.01). In contrast, the White population (3 studies) showed a higher pooled MAE of 10.25 (95% CI: 7.82, 12.68; p < 0.001), and the Black population (4 studies) had the highest pooled MAE at 12.27 (95% CI: 10.95, 13.59; p < 0.001). Funnel plot analysis revealed a symmetrical distribution of studies. For initial dosing algorithms, two studies reported MAEs of 0.24 and 5.43, respectively, reflecting variation based on population size and model type.

CONCLUSION

AI-based models enhance warfarin-dosing accuracy, but performance varies by ethnicity. Broader validation, especially in underrepresented groups such as Black populations, and methodological standardization are essential for equitable implementation.