Extracting PICO Sentences from Clinical Trial Reports using Supervised Distant Supervision.

Area of Science:

• Medical Informatics
• Computational Linguistics
• Evidence-Based Medicine

Background:

• Systematic reviews are crucial for Evidence-Based Medicine (EBM), requiring comprehensive synthesis of published evidence.
• Identifying Population/Problem, Intervention, Comparator, and Outcome (PICO) elements in trial reports is a critical, time-consuming step.
• Automating PICO extraction can significantly expedite the evidence synthesis process.

Purpose of the Study:

• To develop machine learning models for automatic extraction of PICO-relevant sentences from clinical trial articles.
• To address the challenge of expensive data collection for training such models.
• To propose a novel Supervised Distant Supervision (SDS) method for training models efficiently.

Main Methods:

• Utilized distant supervision (DS) by deriving 'soft' labels from unstructured PICO summaries of existing reviews.
• Developed Supervised Distant Supervision (SDS), a novel approach combining DS with a small amount of direct supervision.
• SDS learns to pseudo-annotate articles using available DS to better exploit large, distantly labeled corpora.

Main Results:

• The proposed SDS method demonstrated improved performance in automated PICO extraction compared to existing methods.
• SDS effectively leverages large, distantly labeled datasets by learning from a small set of directly supervised instances.
• The approach shows promise in expediting the critical PICO identification step in systematic reviews.

Conclusions:

• Supervised Distant Supervision (SDS) offers an efficient and effective solution for automating PICO element extraction.
• This method can significantly reduce the time and cost associated with preparing data for systematic reviews.
• The findings contribute to accelerating evidence synthesis and advancing Evidence-Based Medicine.