Semi-supervised learning from small annotated data and large unlabeled data for fine-grained Participants,...

Area of Science:

• Natural Language Processing
• Medical Informatics
• Clinical Trial Analysis

Background:

• Extracting PICO elements from clinical trial literature is crucial for evidence-based medicine.
• Existing methods lack granularity in distinguishing PICO entity attributes.
• There is a need for advanced models to improve the precision of PICO extraction.

Purpose of the Study:

• To develop a Named Entity Recognition (NER) model, FinePICO, for extracting fine-grained PICO entities.
• To enhance the accuracy and detail of information retrieval from clinical trial abstracts.

Main Methods:

• A semi-supervised learning approach was employed, combining limited annotated PICO data with abundant unlabeled data.
• The FinePICO model was trained on a corpus of 2511 abstracts from four public datasets.
• Performance was evaluated using precision, recall, and F1 scores, with comparisons against baseline models and theoretical bounds.

Main Results:

• FinePICO achieved a precision/recall/F1 score of 0.567/0.636/0.60, significantly outperforming the baseline model (F1: 0.437).
• The model demonstrated generalizability across different PICO frameworks and corpora, consistently outperforming benchmarks (P < .001).
• The approach effectively leveraged unlabeled data to improve the extraction of fine-grained PICO entities.

Conclusions:

• FinePICO offers a feasible and effective semi-supervised method for detailed PICO entity extraction from clinical text.
• The study validates the utility of semi-supervised learning techniques in enhancing PICO extraction.
• Future research can optimize semi-supervised learning algorithms to improve efficiency and reduce computational costs.