Variable importance evaluation with personalized odds ratio for machine learning model interpretability with...

Area of Science:

• Machine Learning
• Medical Informatics
• Statistical Modeling

Background:

• Machine learning models offer excellent prediction performance but lack interpretability.
• Understanding variable importance is crucial for practical applications of supervised learning.
• Existing interpretation methods may not fully capture subject heterogeneity.

Purpose of the Study:

• To develop a novel, computationally efficient, and model-agnostic framework for evaluating variable importance.
• To introduce the Personalized Odds Ratio (POR) for quantifying local variable importance.
• To establish a global interpretation strategy using a hierarchical tree for predictor grouping and ranking.

Main Methods:

• Proposed the Variable Importance based on Personalized Odds Ratio (VIPOR) framework.
• Utilized personalized odds ratio (POR) for local interpretation, accounting for subject heterogeneity.
• Employed a hierarchical tree for global interpretation, grouping predictors into five categories (positive, negative, dominated, neutral).

Main Results:

• VIPOR effectively interprets multilayer perceptron (MLP) models for predicting subarachnoid hemorrhage (SAH) patient mortality using electronic health records (EHR).
• Consistent identification of top importance variables across MLP, tree-based models, and L1-regularized logistic regression.
• VIPOR demonstrated comparable or superior performance to existing interpretation methods on public datasets.

Conclusions:

• VIPOR provides a robust and versatile approach to enhance machine learning model interpretability.
• The method successfully identifies key predictors in clinical prediction tasks, aiding in understanding patient outcomes.
• VIPOR's model-agnostic nature and ability to handle heterogeneity make it valuable for diverse machine learning applications.