An effective multi-step feature selection framework for clinical outcome prediction using electronic medical records

Area of Science:

• Medical Informatics
• Machine Learning
• Clinical Prediction Modeling

Background:

• High-dimensional electronic medical records (EMR) present challenges in selecting key variables for clinical outcome prediction models.
• Existing feature selection (FS) methods face difficulties in method selection, determining optimal variable numbers, and ensuring medical relevance.

Purpose of the Study:

• To develop and validate a practical multi-step feature selection (FS) framework integrating data-driven inference and knowledge verification.
• To assess the framework's effectiveness in reducing EMR dimensionality for predicting clinical outcomes like acute kidney injury (AKI) and in-hospital mortality (IHM).

Main Methods:

• A multi-step FS framework combining statistical inference and expert knowledge validation was developed.
• The framework was tested on two EMR datasets (MIMIC-III for AKI, MIMIC-IV-ED for IHM) using various machine learning (ML) methods.
• Comparative analysis focused on accuracy, stability, similarity, and interpretability, with SHAP used for model explanation.

Main Results:

• The tree-based ensemble method demonstrated the highest accuracy among evaluated ML models.
• Increasing top-ranking features stabilized model accuracy and optimized feature subset stability and inter-method similarity.
• The FS framework significantly reduced feature numbers (e.g., 380 to 35 for AKI prediction) without compromising prediction performance.

Conclusions:

• The developed multi-step FS method effectively reduces EMR feature dimensionality while preserving clinical outcome prediction accuracy.
• Incorporating expert knowledge validation enhances the interpretability of identified risk factors.