Knockoff-ML: a knockoff machine learning framework for controlled variable selection and risk stratification in...

Area of Science:

• Computational Biology and Bioinformatics
• Clinical Informatics
• Machine Learning in Healthcare

Background:

• Effective risk stratification is crucial for optimizing resource allocation and patient outcomes in clinical practice.
• Machine learning models in electronic health records (EHR) aid risk prediction but often lack interpretable decision rules for clinicians.
• Existing interpretability metrics struggle to identify specific patient features influencing outcomes.

Purpose of the Study:

• To introduce Knockoff-ML, a model-free machine learning framework for simultaneous outcome prediction and risk feature identification.
• To integrate a knockoff framework with predictive machine learning algorithms for enhanced variable selection.
• To enable false discovery rate (FDR) control in identifying complex, nonlinear associations in EHR data.

Main Methods:

• Developed Knockoff-ML by augmenting traditional machine learning models with a knockoff framework.
• Employed variable selection with FDR control for identifying significant risk features.
• Evaluated performance through simulations and real-world application on the MIMIC-IV database.

Main Results:

• Knockoff-ML demonstrated high statistical power in identifying risk features while controlling FDR in simulations, outperforming conventional methods.
• Identified significant risk features associated with short- and long-term mortality in 50,591 intensive care unit (ICU) patients.
• Achieved comparable prediction accuracy to full models and superior predictive power and clinical utility compared to SOFA and SAPS II scoring systems.

Conclusions:

• Knockoff-ML provides a robust and interpretable tool for clinical decision-making by identifying key patient risk factors.
• The framework enhances prediction accuracy and clinical utility, offering potential improvements in patient outcomes and healthcare delivery.
• Knockoff-ML represents a significant advancement in leveraging EHR data for personalized risk stratification.