Active Learning for Medical Text Classification: Insights from Ophthalmology and Broader Medical Domains

In the realm of healthcare informatics, curated datasets play a crucial role in advancing text classification tasks, which are essential for clinical decision support and knowledge discovery. However, the scarcity of labeled medical data and the high cost of annotation, particularly in specialized domains such as ophthalmology, pose significant challenges. To address these issues, this study explores the use of active learning strategies to optimize the annotation process and improve performance in medical text classification. Six active learning strategies are evaluated on five novel datasets derived from ClinicalTrials.gov, DrugCentral, and PubMed, focusing on ophthalmology and broader medical domains. Leveraging five BERT-based models, the results demonstrate that active learning strategies not only outperform random sampling but also achieve comparable or superior results to full-dataset training while significantly reducing annotation requirements. By highlighting the efficiency and effectiveness of active learning, particularly in data-scarce domains, this work provides a scalable framework for reducing annotation costs and advancing medical text classification, with a focus on clinical relevance in ophthalmology.