A span-based model for extracting overlapping PICO entities from randomized controlled trial publications.

Area of Science:

• Natural Language Processing
• Biomedical Informatics
• Information Retrieval

Background:

• Extracting PICO entities is crucial for efficient evidence retrieval in systematic reviews and clinical research.
• Existing methods struggle with identifying overlapping PICO entities, limiting their effectiveness.

Purpose of the Study:

• To introduce PICOX, a novel method for extracting overlapping PICO entities.
• To evaluate PICOX's performance against established baselines on diverse biomedical datasets.

Main Methods:

• PICOX identifies entity boundaries and employs a multi-label classifier for PICO label assignment.
• The method was validated on PICO-Corpus, Alzheimer's Disease, and COVID-19 datasets, using entity-level precision, recall, and F1 scores.
• Ablation studies were conducted to assess the impact of data augmentation.

Main Results:

• PICOX demonstrated superior precision, recall, and F1 scores compared to the EBM-NLP baseline across all tested datasets.
• Micro F1 scores improved significantly, for example, from 45.05 to 50.87 overall and from 77.10 to 80.32 on the COVID-19 dataset.
• PICOX achieved higher recall and F1 scores on the PICO-Corpus and comparable F1 scores with improved precision on the Alzheimer's Disease dataset.

Conclusions:

• PICOX effectively extracts overlapping PICO entities, outperforming leading methods.
• The data augmentation strategy in PICOX is key to minimizing false positives and enhancing precision.