Transportability of Pharmacogenomics-Guided Warfarin Dose Prediction to a Held-Out Asian Cohort and Assessment of Indian Data Availability in the IWPC Dataset: An Exploratory Comparison of Multiple Linear Regression and Random Forest

Background: Pharmacogenomic warfarin dose prediction commonly combines clinical factors with CYP2C9 and VKORC1 genotype information, but model performance may vary across populations. Objective: To compare multiple linear regression (MLR) with random forest regression (RF) for stable weekly warfarin dose prediction, evaluate transportability to a fully held-out Asian subgroup, and assess whether Indian records in the International Warfarin Pharmacogenetics Consortium (IWPC) dataset were sufficiently complete for Indian-specific modelling. Methods: A retrospective secondary analysis of the IWPC dataset accessed via PharmGKB was performed. After complete-case cleaning, 3,449 records remained. All 1,120 records classified as Asian were withheld before model development. The remaining 2,329 non-Asian records were split 70:30 into training (n=1,630) and held-out test (n=699) sets. MLR and RF (500 trees) used the same eight clinical and pharmacogenomic predictors and a square-root-transformed weekly dose outcome. Performance was assessed by mean absolute error (MAE) and percentage of predictions within 20% of observed dose. Five-fold cross-validation, 1,000-resample bootstrap confidence intervals, and a race-removal sensitivity analysis were used as secondary robustness checks. Results: On the main test set, MLR had lower MAE than RF (9.43 vs 10.24 mg/week) and a higher within-20% proportion (47.2% vs 42.9%). In the held-out Asian subgroup, MLR again had lower MAE (6.70 vs 7.24 mg/week) and higher within-20% performance (40.7% vs 38.1%). Five-fold cross-validation preserved the same ordering. Sequential Indian data assessment reduced 43 self-reported Indian records to 3 after valid-dose, CYP2C9 and VKORC1 requirements and to 0 after the full predictor requirements. Conclusion: Under the evaluated model specifications, RF did not improve stable-dose prediction over MLR. The held-out Asian analysis demonstrates population-transportability challenges, while the Indian analysis shows that incomplete covariate and genotype data can prevent population-specific modelling in an international pharmacogenomic resource.