Efficient and sparse feature selection for biomedical text classification via the elastic net: Application to ICU...

Area of Science:

• Biomedical Informatics
• Clinical Informatics
• Machine Learning in Healthcare

Background:

• Sparsity is crucial for interpretable statistical models, yet its application in biomedical text classification for intensive care unit (ICU) mortality risk stratification is underexplored.
• Developing interpretable models is key for understanding patient outcomes in critical care settings.

Purpose of the Study:

• To create and evaluate sparse classifiers using free-text nursing notes for predicting ICU patient mortality.
• To identify key text features associated with mortality risk from clinical notes.

Main Methods:

• Utilized nursing notes from the first 24 hours of ICU admission for 25,826 adult patients from the MIMIC-II database.
• Developed elastic-net regularized classifiers using stochastic gradient descent.
• Analyzed performance-sparsity trade-offs by varying regularization parameters.

Main Results:

• The best classifier achieved a 10-fold cross-validated Area Under the Curve (AUC) of 0.897 with L2 regularization.
• L1 regularization resulted in an AUC of 0.889 while using only 0.00025% of input features.
• Log loss yielded better performance (AUCs 0.889-0.897) than hinge loss (0.850-0.876), though hinge loss produced sparser models.

Conclusions:

• Elastic-net regularized classifiers demonstrate strong performance in predicting ICU mortality.
• These sparse models significantly reduce feature count (over thousandfold) with minimal performance impact.
• Identified clinically relevant features and novel informative text features associated with mortality risk.