Stable feature selection for clinical prediction: exploiting ICD tree structure using Tree-Lasso.

Area of Science:

• Medical Informatics
• Machine Learning
• Biostatistics

Background:

• Electronic Medical Records (EMR) are increasingly vital for clinical prediction models.
• EMR data uses hierarchical codes (e.g., ICD-10) where correlated features pose challenges for traditional feature selection.
• Existing methods like Lasso struggle with correlated features, leading to unstable risk factor identification.

Purpose of the Study:

• To evaluate the stability and performance of the Tree-Lasso model for feature selection in EMR data.
• To compare Tree-Lasso against traditional and other advanced feature selection techniques.
• To demonstrate the utility of Tree-Lasso in identifying stable clinical risk factors.

Main Methods:

• Applied the Tree-Lasso model to synthetic and real-world datasets (Cancer, Acute Myocardial Infarction).
• Assessed feature selection stability by comparing Tree-Lasso with Information Gain, T-test, ReliefF, and Lasso.
• Evaluated classification performance using logistic regression, naive Bayes, SVM, decision trees, and Random Forest.

Main Results:

• Tree-Lasso demonstrated significantly higher stability compared to Lasso.
• Tree-Lasso's stability was comparable to Information Gain, T-test, and ReliefF.
• Classification performance with Tree-Lasso was on par with Lasso and superior to other methods.

Conclusions:

• Tree-Lasso provides a more stable feature selection approach for EMR data, addressing limitations of Lasso.
• This method enhances the identification of reliable risk factors, crucial for clinical decision-making.
• Tree-Lasso holds significant implications for improving medical prognosis and healthcare analytics.