EvidenceMap: a three-level knowledge representation for medical evidence computation and comprehension.

Area of Science:

• Biomedical Informatics
• Natural Language Processing
• Clinical Trial Research

Background:

• Medical evidence from randomized controlled trials (RCTs) is often unstructured, hindering efficient retrieval and synthesis.
• Existing frameworks like PICO (Participant, Intervention, Comparator, Outcome) lack the granularity for detailed evidence representation.
• Developing computable formats for medical evidence is crucial for advancing evidence-based medicine.

Purpose of the Study:

• To develop a computable representation for medical evidence from RCT abstracts.
• To create a gold-standard dataset of annotated RCT abstracts.
• To build a natural language processing (NLP) pipeline for transforming free-text RCT evidence into a structured format.

Main Methods:

• Developed EvidenceMap, a 3-level hierarchical representation (Entity, Proposition, Map).
• Annotated 229 disease-agnostic and 80 COVID-19 RCT abstracts by two independent annotators.
• Trained an NLP pipeline using the annotated corpus and evaluated its performance using F1 scores (0.84 and 0.86).

Main Results:

• Annotated corpora yielded 12,725 entities and 1,602 propositions.
• EvidenceMap reduced user time by 51.9% compared to reading raw abstracts.
• High user ratings for EvidenceMap's representation of enrollment (4.85), study design (4.70), and results (4.20).

Conclusions:

• EvidenceMap extends the PICO framework into a computable, hierarchical structure for medical evidence.
• The NLP pipeline effectively transforms free-text RCT evidence into this structured representation.
• EvidenceMap facilitates efficient evidence retrieval, synthesis, and comprehension of clinical trial findings.