Enhanced Phenotype Identification of Common Ocular Diseases in Real-World Datasets

Area of Science:

• Ophthalmology
• Health Informatics
• Machine Learning

Background:

• Accurately identifying patient phenotypes in real-world data is crucial but challenging.
• International Classification of Diseases (ICD) billing codes are often used but have limitations for precise patient cohort identification.
• Electronic Health Records (EHR) contain rich data for phenotype identification.

Purpose of the Study:

• To develop and validate enhanced phenotype identification (EPI) algorithms for detecting diabetic retinopathy (DR), age-related macular degeneration (AMD), and glaucoma.
• To compare the performance of EPI models against ICD-only models in EHR data.
• To assess the accuracy of EPI models in identifying patients with these common ocular diseases.

Main Methods:

• Developed EPI algorithms to analyze various EHR fields (e.g., examination findings, orders, medications, surgeries).
• Trained and validated EPI models using gold-standard ophthalmologist assessments of EHR data.
• Compared EPI model performance against ICD-only models and validated on a separate data repository site.

Main Results:

• EPI models demonstrated superior performance over ICD-only models for glaucoma, DR, and AMD, evidenced by higher Area Under the Receiver Operating Characteristic Curve (AUC) and Area Under the Precision-Recall Curve (AUPRC).
• For glaucoma, EPI AUC was 0.97 vs. 0.90 (ICD-only); for DR, 0.997 vs. 0.98; for AMD, 0.99 vs. 0.95.
• Validation on a second site showed high performance, though model calibration was slightly reduced.

Conclusions:

• Machine learning-driven EPI models accurately identify patients with glaucoma, DR, and AMD in real-world datasets.
• EPI models significantly outperform ICD-only approaches, highlighting the value of comprehensive EHR data analysis.
• These findings support improved patient management and clinical outcomes through enhanced phenotype identification.