Prediction of mortality in Intensive Care Units: a multivariate feature selection

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Critical Care Medicine

Background:

• Intensive Care Units (ICUs) require intensive monitoring, leading to high costs and administrative challenges.
• Existing clinical scores like APACHE and SOFA for predicting patient mortality are imprecise and outdated.
• Machine learning offers potential for improved mortality prediction in ICUs, as explored in the 2012 PhysioNet Challenge.

Purpose of the Study:

• To improve the prediction of mortality in ICUs using data from the 2012 PhysioNet Challenge.
• To simplify the prediction problem by reducing data dimensionality.
• To decrease classifier dependence on training data and reduce uncontrolled variance.

Main Methods:

• A methodology involving sample filtering and data normalization was applied.
• Multivariate data analysis techniques, including Principal Component Analysis and Factor Analysis, were used for feature selection and variance reduction.
• Machine learning classifiers were developed and evaluated using metrics from the 2012 PhysioNet Challenge.

Main Results:

• The best performing classifier, a linear Support Vector Machine (SVM), achieved approximately 0.73 accuracy.
• This represents a significant improvement over previous work, which achieved approximately 0.54 accuracy.
• The proposed approach required only twelve features, substantially fewer than prior methods.

Conclusions:

• The developed approach demonstrated a 20% increase in predicting death risks with reduced dependence on training data.
• The method offers lower costs for ICU monitoring due to the reduced number of features required.
• Future research should explore incorporating additional physiological features not available in the 2012 dataset.