Clinically Informed Semi-Supervised Learning Improves Disease Annotation and Equity from Electronic Health Records: A...

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Medical Data Analysis

Background:

• Structured Electronic Health Record (EHR) data, like ICD codes, are often insufficient for detailed disease characterization.
• Limitations in EHR data hinder accurate disease pathogenesis studies and AI system development.
• Clinical notes offer rich information but are largely underutilized.

Purpose of the Study:

• To introduce Ci-SSGAN (Clinically Informed Semi-Supervised Generative Adversarial Network), a novel framework to reannotate patient conditions using clinical text.
• To improve the accuracy and equity of patient data for disease research and AI development.
• To address limitations in current EHR data for complex disease characterization.

Main Methods:

• Leveraging large-scale unlabeled clinical text for semi-supervised learning.
• Utilizing a Generative Adversarial Network (GAN) architecture informed by clinical data.
• Integrating demographic conditioning for equitable AI model performance.
• Case study application to glaucoma using 2.1 million ophthalmology notes.

Main Results:

• Ci-SSGAN achieved 0.85 accuracy and 0.95 AUROC for glaucoma annotation.
• Demonstrated a 10.19% AUROC improvement over ICD-based labels (0.85 vs. 0.74).
• Narrowed performance gaps in subgroups, showing F1 gains for Black patients (+0.05), women (+0.06), and younger patients (+0.033).

Conclusions:

• Ci-SSGAN effectively reannotates patient conditions from clinical text with high accuracy and equity.
• The framework significantly outperforms traditional ICD-based labeling for complex diseases like glaucoma.
• Ci-SSGAN's semi-supervised approach and demographic conditioning reduce reliance on expert annotations, enhancing AI accessibility in healthcare.