Accurate text extraction from scanned medical documents is essential for data management and clinical decision-making. This study evaluates Large Language Models (LLMs) as an enhancement to traditional Optical Character Recognition (OCR) methods. By leveraging language and context, LLMs offer improved accuracy and relevance in text interpretation. We compared the EasyOCR model and the multimodal \\\"gpt-4o-mini\\\" LLM on a dataset of 110 medical transcript samples. Performance was assessed by comparing extracted texts against clinical data embeddings, using cosine similarity for semantic accuracy. The OCR model achieved an F1-score of 0.59, while the LLM scored 0.70, demonstrating LLMs\\\' potential to advance text extraction in healthcare

Keywords