Deploying a PICO Extractor in Five Steps

A project was undertaken to automate the extraction of PICO (Population, Intervention, Comparator, Outcome) elements from clinical trial descriptions using Natural Language Processing. The process involved benchmarking existing NER models, then fine-tuning several BERT variants, including BioBERT and the more compact BioMobileBERT, on a specialized dataset. Evaluation showed strong performance for Population and Outcome extraction, but challenges with Intervention precision and the superiority of a rule-based approach for Comparators. The final pipeline, favoring the efficient BioMobileBERT model, was deployed as an accessible Streamlit application containerized with Docker. This demonstrates a full lifecycle from model development to deployment for a domain-specific AI task.