Improving condition severity classification with an efficient active learning based framework.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Data Analysis

Background:

• Classifying condition severity is crucial for healthcare decisions and patient prioritization.
• Electronic Health Records (EHRs) offer valuable labeled data but expert labeling is costly and time-consuming.
• Active Learning (AL) techniques can potentially reduce the burden of expert labeling in EHR analysis.

Purpose of the Study:

• To demonstrate the effectiveness of Active Learning (AL) techniques in reducing expert labeling efforts for condition severity classification.
• To introduce the Active Learning Enhancement (CAESAR-ALE) framework, integrating AL methods into the existing CAESAR framework.
• To evaluate the performance of CAESAR-ALE in terms of labeling reduction and classification accuracy.

Main Methods:

• Developed the Active Learning Enhancement (CAESAR-ALE) framework by incorporating three AL methods into the original CAESAR framework.
• Applied CAESAR-ALE to the "CAESAR dataset" comprising 516 conditions labeled by seven experts from 1.9 million patient records.
• Compared the performance of AL methods against the original CAESAR framework's training approach.

Main Results:

• All three AL methods integrated into CAESAR-ALE reduced labeling efforts by 48% to 64% compared to the original CAESAR framework.
• CAESAR-ALE achieved over 13% absolute improvement in Positive Predictive Value (PPV) for classifying severe conditions.
• The AL methods proved robust to variations in human labeler expertise.

Conclusions:

• Active Learning (AL) methods significantly decrease the need for expert labeling in EHRs for condition severity classification.
• CAESAR-ALE enhances classification accuracy while reducing the volume of data requiring expert annotation.
• The developed framework offers substantial time and cost savings for healthcare data analysis and supports diverse expert input.