NLP-ROPCare: predicting retinopathy of prematurity with admission notes using natural language processing

Area of Science:

• Ophthalmology
• Neonatology
• Medical Informatics

Background:

• Retinopathy of prematurity (ROP) is a significant cause of childhood blindness globally.
• Efficient prediction models are crucial for timely treatment of ROP.
• Current prediction methods may not fully leverage available clinical data.

Purpose of the Study:

• To develop and evaluate a novel prediction model, NLP-ROPCare, for ROP occurrence and severity.
• To utilize natural language processing (NLP) on unstructured clinical notes for ROP prediction.
• To compare NLP model performance against traditional machine learning methods.

Main Methods:

• A retrospective observational study included 3922 preterm infants (gestational age ≤32 weeks or birth weight ≤2000 g).
• Four pretrained language models (BERT, RoBERTa, MC-BERT, NEZHA) were trained on free-form admission notes.
• Performance was compared to Random Forest and Support Vector Machine models using structured data, evaluated by accuracy, precision, recall, F1 score, and AUC.

Main Results:

• NLP models showed superior performance in predicting ROP occurrence compared to severity.
• The NEZHA model achieved the highest accuracy for ROP occurrence prediction (F1 score: 89.35%, AUC: 0.90), outperforming machine learning models.
• For ROP severity, RoBERTa yielded the highest AUC (0.91) with an F1 score of 78.44%.

Conclusions:

• NLP-ROPCare, integrating NEZHA and RoBERTa, enables early prediction of ROP occurrence and severity from unstructured admission notes.
• This model demonstrates significant potential for the early prevention of ROP in preterm infants.
• External validation is recommended for further model refinement and broader applicability.