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

Area of Science:

• Artificial Intelligence
• Machine Learning
• Pharmacogenomics

Background:

• Warfarin dose prediction is complex due to linear and nonlinear factors.
• Traditional machine learning algorithms struggle with simultaneous linear and nonlinear dose prediction.
• Artificial intelligence (AI) is increasingly applied to warfarin dose prediction.

Purpose of the Study:

• To develop an improved stacking ensemble learning model for accurate warfarin dose prediction in Chinese patients.
• To enhance prediction accuracy by leveraging the specific characteristics of clinical warfarin data.
• To identify additional factors influencing warfarin dosage through feature selection.

Main Methods:

• Collected data from 641 Chinese warfarin patients, including demographics, medical history, genotype, and co-medications.
• Utilized a heuristic-stacking ensemble learning approach for prediction.
• Evaluated model performance using metrics like ideal dose accuracy, mean absolute error, root mean square error, and R-squared.
• Employed feature selection methods to discover associated factors.

Main Results:

• The heuristic-stacking ensemble model achieved higher accuracy in ideal dose prediction (73.44%) compared to traditional stacking (71.88%).
• The new model demonstrated lower mean absolute errors (0.11 mg/day vs. 0.13 mg/day) and root mean square errors (0.18 mg/day vs. 0.20 mg/day).
• The heuristic-stacking model yielded a higher R-squared value (0.87 vs. 0.82), indicating better model fit.

Conclusions:

• The developed heuristic-stacking ensemble learning model accurately predicts warfarin dose.
• Hypertension and a history of severe preoperative embolism were identified as significant factors influencing warfarin dose.
• This AI-driven approach offers a valuable reference for optimizing warfarin dose administration.