Improving the performance and interpretability on medical datasets using graphical ensemble feature selection.

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Medicine

Background:

• Machine learning (ML) in medicine faces challenges with high-dimensional data and small sample sizes, leading to overfitting.
• Existing feature selection methods often fail to fully utilize dependencies identified by component algorithms.
• Ensemble techniques offer robustness but often neglect inter-algorithm dependencies.

Purpose of the Study:

• To introduce Graphical Ensembling (GE), a novel graph-theory-based feature selection method for ML on medical datasets.
• To enhance the stability and relevance of selected features in high-dimensional medical data.
• To address the limitations of current ensemble methods in leveraging component algorithm dependencies.

Main Methods:

• Developed Graphical Ensembling (GE), a technique utilizing graph theory for ensemble feature selection.
• Applied GE to four distinct datasets, including patient stratification for rheumatoid arthritis and sub-cellular network data.
• Compared GE's performance against baseline methods in terms of classification accuracy and feature set size.

Main Results:

• GE significantly improves classification performance, achieving 9% higher Balanced Accuracy in rheumatoid arthritis patient stratification.
• The method selects fewer features while enhancing predictive power across tested datasets.
• Analysis of sub-cellular networks revealed selected features (proteins) are biologically closer to known disease genes, uncovering more diverse mechanisms.

Conclusions:

• Graphical Ensembling (GE) offers a robust solution for feature selection in ML applied to complex medical data.
• The approach effectively handles intricate correlations between biological variables, improving model stability and relevance.
• GE is anticipated to significantly advance the application of machine learning in medical research and clinical practice.