Self-Prompting Small Language Models for Privacy-Sensitive Clinical Information Extraction

This research presents a locally deployable framework enabling small language models to extract privacy-sensitive clinical entities from unstructured dental notes through self-generated and refined prompts. The study evaluated open-weight models, achieving high F1 scores with Qwen2.5-14B-Instruct and Llama-3.1-8B-Instruct after supervised fine-tuning and direct preference optimization.