Deep Learning Approaches for Predicting Glaucoma Progression Using Electronic Health Records and Natural Language...

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Informatics

Background:

• Predictive models for glaucoma progression exist, but integrating free-text clinical notes remains a challenge.
• Artificial intelligence (AI) offers potential for enhanced clinical prediction.

Purpose of the Study:

• To predict glaucoma progression requiring surgery using deep learning (DL) on electronic health records (EHRs).
• To evaluate the utility of natural language processing (NLP) of clinical free-text notes alongside structured EHR data.

Main Methods:

• A deep learning (DL) predictive model was developed using an observational cohort of adult glaucoma patients.
• Data included structured EHR information (demographics, diagnoses, intraocular pressure, visual acuity) and NLP-processed clinical free-text notes.
• Ophthalmology-specific word embeddings were trained on PubMed abstracts to represent text data for DL models.

Main Results:

• The DL model incorporating both structured data and clinical free-text notes achieved an AUC of 73% and F1 score of 40%.
• Models using only structured data (AUC 66%, F1 34%) or only free-text data (AUC 70%, F1 42%) showed lower performance.
• All developed models outperformed a glaucoma specialist's manual review (F1 29.5%).

Conclusions:

• Deep learning models effectively predict the need for glaucoma surgery using EHR data, particularly when incorporating unstructured clinical notes.
• Free-text clinical notes contain valuable information that significantly enhances predictive model performance compared to structured data alone.
• Future EHR-based predictive models should leverage NLP of clinical notes for improved accuracy; further research into incorporating imaging data is warranted.