Active feature elicitation: An unified framework.

Area of Science:

• Machine Learning
• Clinical Informatics
• Data Science

Background:

• Collecting complete patient data, such as full Electronic Health Records (EHRs), can be resource-intensive due to expensive or sensitive features like lab tests.
• Existing methods may not efficiently guide the process of acquiring additional data for specific patient examples.

Purpose of the Study:

• To develop a general active learning approach for efficient feature elicitation in clinical settings.
• To identify patient examples where collecting additional information is most beneficial, optimizing resource allocation.

Main Methods:

• Proposed a classifier-independent and similarity metric-independent active learning strategy.
• Identified examples dissimilar to those with complete feature sets for targeted data acquisition.
• Evaluated the approach across four diverse clinical tasks using various divergence metrics and classifiers.

Main Results:

• Demonstrated the effectiveness of the proposed active learning approach in identifying key examples for data collection.
• Achieved consistent results across different domains, showcasing the generalization capabilities of the method.
• The approach successfully guided the acquisition of complete feature sets for selected dissimilar examples.

Conclusions:

• The developed active learning method offers an efficient strategy for feature elicitation in healthcare.
• The approach is robust and generalizable, applicable across various clinical scenarios and data types.
• This work provides a valuable tool for optimizing data collection in resource-constrained clinical research and practice.