Improved stacking ensemble learning based on feature selection to accurately predict warfarin dose

Mingyuan Wang, Yiyi Qian, Yaodong Yang, Haobin Chen, Wei-Feng RaoFrontiers in Cardiovascular Medicine 2024 · DOI

Abstract

The developed heuristic-stacking ensemble learning can satisfactorily predict warfarin dose with high accuracy. A relationship between hypertension, a history of severe preoperative embolism, and warfarin dose is found, which provides a useful reference for the warfarin dose administration in the future.

Introduction

Warfarin remains a cornerstone of anticoagulant therapy, yet its clinical application is fraught with risk due to its narrow therapeutic index. The margin between an effective dose and a life-threatening hemorrhage is exceptionally slim, necessitating precise dosing strategies. While genetic markers such as CYP2C9 and VKORC1 provide a foundation for dosing, they fail to account for the full spectrum of clinical variability. This study introduces a "heuristic-stacking" ensemble learning framework designed to capture the complex, non-linear relationships between clinical factors and stable warfarin requirements.

Methodology

Feature Selection via Bidirectional Search (BDS)

To optimize model performance, the researchers employed Bidirectional Search (BDS) to identify the most impactful features. Unlike standard filtering methods, BDS seeks a globally optimal subset of variables, effectively reducing noise and preventing the model from overfitting on irrelevant data. This process identified nine core variables as the primary drivers of dose variance.

Heuristic-Stacking Architecture

The proposed framework utilizes a "stacking" approach where multiple base learners contribute to a final prediction. The architecture consists of five base models: Random Forest (RF), Gaussian Process Regression (GPR), Gradient Boosting (GB), Multiple Linear Regression (MLR), and Stochastic Gradient Descent (SGD).

To refine the ensemble, the authors implemented a heuristic weighting system. Rather than using a simple average, the contribution of each base learner is weighted by the reciprocal of its Root Mean Square Error (RMSE). This ensures that higher-performing models have a greater influence on the final output. A Support Vector Regressor (SVR) serves as the meta-learner to synthesize these weighted predictions.

Results and Feature Analysis

The heuristic-stacking model demonstrated superior performance over individual algorithms and traditional stacking methods. Key metrics included an accuracy of 73.44% (within ±20% of the actual dose), a Coefficient of Determination ($R^2$) of 0.87, and a Mean Absolute Error (MAE) of 0.11 mg/day.

Radar charts illustrating the correlation coefficients (Spearman, Pearson, Kendall tau-b, and Average) between clinical features and stable warfarin dose.

Analysis of the feature correlations (as shown in the radar charts above) highlights the importance of clinical history. Notably, hypertension and a history of severe preoperative embolism (HSEBO) emerged as significant predictors. This suggests that a patient's physiological and inflammatory state prior to surgery significantly alters their sensitivity to warfarin.

Discussion and Clinical Implications

The study utilized data from 641 patients in southern China. The findings emphasize that clinical factors are just as vital as genetic ones in determining the "Goldilocks" dose. The identification of hypertension and HSEBO as critical variables provides clinicians with new markers to consider when initiating therapy.

However, the study's geographic focus on southern China and its relatively modest sample size suggest that further validation is needed. Future research should aim to incorporate larger, more diverse datasets to ensure the model's generalizability across different ethnicities and dietary habits (e.g., Vitamin K intake).

Conclusion

The heuristic-stacking ensemble learning model offers a high-precision tool for warfarin dose prediction. By combining advanced feature selection with weighted ensemble techniques, the model achieves a level of accuracy that could significantly reduce the risks associated with anticoagulant therapy. Integrating such models into clinical decision-support systems represents a promising step toward personalized medicine in cardiovascular care.

Summary updated

Abstract

The developed heuristic-stacking ensemble learning can satisfactorily predict warfarin dose with high accuracy. A relationship between hypertension, a history of severe preoperative embolism, and warfarin dose is found, which provides a useful reference for the warfarin dose administration in the future.
BibTeX
@inproceedings{wang2024improved,
  author = {Mingyuan Wang and Yiyi Qian and Yaodong Yang and Haobin Chen and Wei-Feng Rao},
  title = {{Improved stacking ensemble learning based on feature selection to accurately predict warfarin dose}},
  booktitle = {Frontiers in Cardiovascular Medicine},
  year = {2024},
  doi = {10.3389/fcvm.2023.1320938},
}
Figure