Development and Evaluation of a Computable Phenotype for Normal Tension Glaucoma.

Area of Science:

• Ophthalmology
• Medical Informatics
• Health Data Science

Background:

• Normal tension glaucoma (NTG) is a subtype of glaucoma requiring accurate identification for effective management.
• Electronic Health Records (EHRs) contain vast amounts of patient data that can be leveraged for disease phenotyping.
• Developing computable phenotypes is crucial for enhancing disease detection and research from EHR data.

Purpose of the Study:

• To develop and validate computable phenotypes for normal tension glaucoma (NTG) using EHR data.
• To improve the accuracy and efficiency of identifying NTG patients within large datasets.
• To compare the performance of different phenotyping approaches, including those using ICD codes, structured data, and natural language processing (NLP).

Main Methods:

• A retrospective cohort study was conducted using deidentified EHR data from 1851 patients aged ≥40 years with glaucoma.
• Four rule-based NTG computable phenotypes were developed: 1) ICD codes only, 2) structured IOP and medication data, 3) structured IOP data only, and 4) structured data combined with NLP for notes.
• Phenotypes were validated internally and externally using chart review for gold standard diagnoses.

Main Results:

• Chart review confirmed NTG in 30% of patients, with only 7% having NTG ICD codes.
• The ICD code-based phenotype (Phenotype 1) showed low sensitivity (24.1%) but high specificity (97%).
• Phenotypes incorporating structured data (Phenotypes 2 and 3) and NLP (Phenotype 4) significantly improved performance, with Phenotype 4 achieving the highest F1 score (77.4%), sensitivity (82.8%), and accuracy (85.5%).
• Phenotypes 2-4 increased NTG case detection fourfold compared to Phenotype 1.

Conclusions:

• Computable phenotypes for NTG utilizing NLP achieve the best performance in identifying patients from EHRs.
• Phenotypes incorporating structured data outperform those relying solely on ICD codes, which lack sensitivity.
• These findings may guide the development of computable phenotypes for other disease subtypes.