Machine Learning-Directed Explainable Artificial Intelligence and Bayesian Predictions of Warfarin Therapeutic Dose for Orthopedic Indications
Kannan Sridharan, Gowri SivaramakrishnanIntroduction:
Warfarin, a cornerstone anticoagulant for preventing venous thromboembolism following orthopedic procedures such as total hip and knee arthroplasty, is challenged due to a narrow therapeutic window and inter-patient variability. We aimed to apply machine learning (ML) algorithms to predict stable therapeutic warfarin doses in patients with orthopedic indications and to understand the model decisionmaking using explainable artificial intelligence.
Materials and Methods:
This was an observational study carried out on patients who received warfarin for orthopedic indications. The dataset was partitioned into training (80%) and testing (20%) sets. Five ML algorithms were implemented and compared: elastic net, random forest (RF), support vector machine (SVM), gradient boosting machine, and XGBoost. The following metrics were used for assessing the model performance: mean absolute error (MAE), root mean squared error (RMSE), mean absolute percentage error (MAPE), R², and the percentage of predictions within 20% of the actual dose (P20). The predictions from the best ML model were analyzed using SHAP. Uncertainty intervals were estimated using Bayesian linear regression. Exploratory subgroup analyses were carried out across age groups, gender, genetic, racial, and dose strata.
Results:
The mean therapeutic warfarin dose was 34.2 ± 18.7 mg/week. Random forest (RF) was observed to have the lowest MAE (5.84 mg/week) and RMSE (7.63 mg/week) with the highest R² (0.86) following bootstrap validation. These findings should be considered exploratory and hypothesis-generating. SHAP analysis identified VKORC1 G/G (positive contribution) and VKORC1 A/A (negative contribution) as the most influential predictors, followed by CYP2C9 variants and anthropometric factors. SHAP analysis revealed age group 30–39 years and CYP2C9 *1/*1 genotype were the main sources of model miscalibration, as observed by their significantly different values between concordant and discordant predictions. Bayesian analysis confirmed positive linear relationships for height and weight with dose requirements but demonstrated inferior predictive performance (R²=0.38, P20=30.77%). Exploratory subgroup analyses revealed better performance in patients aged 50–59 years and African Americans, while the high-dose category (>75th percentile) and patients aged 60–69 years presented the greatest predictive difficulty.
discussion:
The machine learning models demonstrated only moderate predictive accuracy for warfarin dosing, with genetic factors far outperforming clinical metrics like height and weight—underscoring the limitations of current algorithms in capturing complex pharmacokinetic interactions. Notably, model miscalibration in younger adults and specific CYP2C9 genotypes, alongside variable performance across age and racial strata, emphasizes the need for subgroup-specific refinement rather than a one-size-fits-all approach. While Bayesian methods offered useful uncertainty estimates, their inferior predictive power suggests that integrating genetic drivers with stratified machine learning strategies may be necessary to improve clinical utility.
Discussion:
The ML models demonstrated moderate predictive accuracy for warfarin dosing, with genetic factors outperforming anthropometrics. Varied performance across age and racial groups possibly indicates the need for subgroup-specific refinement rather than a one-size-fits-all approach.
Conclusion:
Machine learning models incorporating genetic and clinical factors can predict warfarin doses in orthopedic patients with moderate accuracy, with genotype variants exerting predominant influence, as in other indications. Prospective validation is warranted before clinical implementation.