Maximally informative feature selection using Information Imbalance: Application to COVID-19 severity prediction

Area of Science:

• Biostatistics
• Clinical Informatics
• Data Science

Background:

• Clinical databases contain diverse patient data, including medical history, symptoms, and test results.
• Identifying the most informative features is crucial for accurate clinical predictions, especially for minority patient groups.

Purpose of the Study:

• To adapt and apply the Information Imbalance statistical approach for selecting maximally informative patient features in a clinical setting.
• To identify key features predictive of clinical fate and disease severity in COVID-19 patients.

Main Methods:

• Utilized the Information Imbalance statistical approach, adapted for categorical and incomplete clinical data.
• Applied the algorithm to a dataset of approximately 1300 COVID-19 patients treated before October 2021.
• The method automatically determines the optimal number of features and handles missing data without imputation.

Main Results:

• Identified combinations of 10-15 patient features, measurable at admission, that are highly informative of clinical outcomes and disease severity.
• Demonstrated the approach's effectiveness even with features available for only a fraction of patients.
• The selected features exhibited low inter-feature correlation.

Conclusions:

• The adapted Information Imbalance method provides an effective way to select parsimonious, informative feature sets from complex clinical data.
• This approach can enhance predictive modeling for diseases like COVID-19, offering valuable clinical insights for patient management.