Starved by scarcity: A three scarcities framework and decision roadmap for medical named entity recognition in low-resource languages

Named entity recognition (NER) is a foundational task in medical natural language processing (NLP), enabling the extraction of clinically relevant information from unstructured text and supporting downstream applications such as information retrieval and clinical decision support. While pre-trained language models such as BioBERT, ClinicalBERT, and PubMedBERT have achieved state-of-the-art performance for English medical NER, progress in low-resource languages remains fragmented and uneven, even though healthcare systems worldwide generate vast amounts of textual data whose value depends on the availability of language technologies. Existing surveys provide valuable overviews but largely treat low-resource settings as a single category, overlooking important differences in the constraints practitioners face. To address this gap, we conducted a systematic review of 46 studies on medical NER in low-resource languages published since 2020. Based on the synthesized evidence, we introduce the three scarcities framework, which conceptualizes low-resource medical NLP as the interaction among data, model, and infrastructure scarcities. We also propose a practitioner-oriented decision roadmap that maps appropriate modeling and data augmentation strategies to different resource configurations. Our review develops a structured taxonomy of approaches, including cross-lingual transfer, in-domain pre-training, annotation projection, back-translation, and large language model-based synthetic data generation, and shows that the effectiveness of techniques depends less on the method itself than on the underlying scarcity profile. We also identify persistent weaknesses in evaluation practices, including inconsistent reporting and limited use of statistical significance testing.