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

Area of Science:

• Artificial Intelligence in Medicine
• Clinical Informatics
• Ophthalmology Research

Background:

• Structured Electronic Health Record (EHR) data, like ICD codes, often lack the granularity needed for accurate disease characterization.
• Limitations in EHR data hinder disease pathogenesis research and the development of robust Artificial Intelligence (AI) systems.
• Clinical notes offer rich information but are underutilized due to challenges in processing unstructured text.

Purpose of the Study:

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

Main Methods:

• Developed Ci-SSGAN, a Clinically Informed Semi-Supervised Generative Adversarial Network framework.
• Leveraged large-scale unlabeled clinical text data for reannotation.
• Applied the framework to glaucoma, a leading cause of blindness with known racial and ethnic disparities.

Main Results:

• Ci-SSGAN achieved 0.85 accuracy and 0.95 AUROC on ophthalmology notes, surpassing ICD-based labels (0.74 accuracy, 0.85 AUROC) by 10.19% AUROC.
• The framework demonstrated improved equity, narrowing performance gaps for Black patients (+0.05 F1), women (+0.06 F1), and younger patients (+0.033 F1).
• The model was trained on 2.1 million ophthalmology notes.

Conclusions:

• Ci-SSGAN effectively reannotates patient conditions from clinical notes, enhancing accuracy and equity.
• The framework's integration of semi-supervised learning and demographic conditioning minimizes reliance on expert annotations.
• Ci-SSGAN offers a more accessible approach to AI development for resource-constrained healthcare systems, particularly for complex diseases with disparities.