Year: 2025
Type: article
Abstract: Named entity recognition of digitized medical records accessed through Optical Character Recognition (OCR) poses significant problems, since the character distortions, uneven formatting, and domain-specific acronyms render it quite difficult. Such artifacts worsen the quality of rule-based or machine learning models that do not perform well under such noisy conditions by retaining consistent entity extraction. The use of hybrid techniques, i.e., the combination of deterministic rule-based modules with neural convolutional models like transformer-based models, is a stable remedy to these problems. Hybrid systems show better tolerance to OCR-induced noise by combining lexicon-based rules with contextual embeddings and error correction mechanisms, and ensemble strategies to maximize precision and achieve higher recall in clinical entity extraction (diagnoses, medications, and time-related entities). The piece is an analysis of the issues related to processing OCR-generated medical text, the implementation and development of hybrid NER pipelines, their institution-agnostic scalability, and research directions, such as multimodal learning, self-supervised pretraining on noisy data, and the orchestration of large-scale healthcare systems with the help of AI.
Author: FNU Sudhakar Abhijeet
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: hybrid
OpenAlex ID: W4413993692