12,445 papers · continuously updated · last export: 10 Aug 2026livingmeta.ai
← Browse all papers
AI evidence extraction

Automating in High-Expertise, Low-Label Environments: Evidence-Based Medicine by Expert-Augmented Few-Shot Learning

Rong Liu, Jingjing Li, Marko Zivkovic, Ahmed Abbasi · MIS Quarterly · 2025

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10
Relevance
2/4
Quality (LMQS)
D
Evidence
2
Citations
3.36
FWCI
Top 10%
Impact

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.25300/misq/2024/18573

Methodology & findings

Study design

Computational design science research combining theory-guided design (Compositionality Theory), artifact development, benchmarking analysis against multiple baseline models (traditional classifiers, LLMs, semi-supervised learning, few-shot learning methods), ablation studies, and mixed-methods evaluation (quantitative metrics on three datasets + qualitative expert interviews + downstream case study integration with SR company workflow)..

Primary method

Theory-guided design science (grounded in Compositionality Theory from cognitive science). Systematic requirements analysis based on task complexity identification, artifact instantiation and evaluation following computational design genre (Padmanabhan et al., 2022; Rai, 2017), with socio-technical evaluation encompassing benchmarking, ablation analysis, downstream case study integration, and qualitative expert consultation.

Main result

FastSR achieves an average F1@3 of 70% for sentence classification, exceeding SVM by 40%, GPT-4 by 33%, CNN by 11%, fine-tuned BioBERT by 8%, and ProtoNet by 7%. For sequence tagging, "FastSR obtained the highest performance, with an F1word@3 (F1BERT@3) of 61% (67%), about 15% higher than LSTM+CRF and 4% than ProtoNER." The study demonstrates that "FastSR significantly reduces time and cost due to its efficient automation and superior performance compared to benchmarking models. With single screening, FastSR achieves a 65% of time reduction relative to a double-screening manual SR, leading to a cost saving of $73,500 for a project with 550 articles."

Research paradigm

Design Science Research / Computational Design

Author conclusions

"In conclusion, we propose a customized FSL solution with several novel extensions to integrate human annotation logic into a computational framework that requires minimal labeling effort. Our comprehensive multi-faceted evaluation showcases FastSR's advantages in data extraction and SR automation compared to baseline models. Our results have important implications for designing computational artifacts in high-expertise, low-label environments." The authors further conclude that "FastSR significantly reduces time and cost due to its efficient automation and superior performance compared to benchmarking models. With single screening, FastSR achieves a 65% of time reduction relative to a double-screening manual SR, leading to a cost saving of $73,500 for a project with 550 articles."

Risk of bias

Selection bias: Limited to medical SR domain; generalizability to other domains not tested; Data contamination risk: ALMs (GPT-4) potentially contaminated by overlapping pre-training data with EBM-NLP dataset; Labeling bias: PICO element definitions disease-specific and evolving; depends on expert consensus; Class imbalance: Multiple PICO elements with varying prevalence; may bias model learning; Sample selection bias: WD dataset consists of only 222 labeled sentences from a single pharmaceutical company's 1,500-hour project; limited generalizability to other disease domains; Limited labeled data: FSL approach with K=5 support samples may not capture full complexity of PICO element variation; Annotation quality: Kappa scores (0.85 for WD, 0.87 for COVID, 0.67 for EBM-NLP) indicate moderate inter-rater agreement, particularly lower for public dataset; Data contamination risk: Authors acknowledge GPT-4 performance gap reduced significantly on public EBM-NLP dataset, suggesting potential pre-training data overlap; Class imbalance: Countered with one-vs-rest approach but may still affect minority class performance; Evaluation context bias: Authors had access to SR company domain experts for WD/COVID but public dataset (EBM-NLP) was crowd-labeled with lower agreement; Potential domain overfitting to medical literature; Limited to predefined PICO classes; Data contamination risk in LLM benchmarking (GPT-4); Inter-annotator agreement variability across datasets (Kappa: 0.85-0.87 for WD/COVID, 0.67 for EBM-NLP)

Limitations

  • "First, we have chosen to concentrate on demonstrating generalizability within the medical SR context, leaving the exploration of its application to other domains for future work
  • Second, FastSR assumes predefined PICO classes established by medical professionals, but future work could focus on developing methods to automatically identify new, disease-specific PICO classes
  • Third, PICO classes often have a hierarchical structure..
  • Future work could explore prototype designs that might better reflect this hierarchy to accommodate small or rare subclasses
  • Fourth, while ALMs underperform in data extraction tasks compared to BERT-based models, they might be useful for downstream tasks like evidence synthesis."

Open questions raised

  • Generalizability beyond medical SR domain to other high-expertise, low-label contexts
  • Automatic identification of new, disease-specific PICO classes
  • Support for hierarchical PICO class structures
  • Integration of LLMs for downstream evidence synthesis tasks
  • Replacement of individual ML components as new technologies emerge
  • Application beyond medical SR context to other high-expertise, low-label domains
Data: Wilson Disease (WD) dataset: 222 labeled training sentences from 550 screened articles, 110 test articles with 12,320 sentences (proprietary, from global life sciences company); COVID-19 dataset: 266 articles merged from https://www.kaggle.com/allen-institute-for-ai/CORD-19-research-challenge (143 articles) and https://www.kaggle.com/mlconsult/summary-page-covid-19-risk-factors; EBM-NLP dataset: 5,000 abstracts from RCTs, publicly available at https://github.com/bepnye/EBM-NLP; WD (Wilson Disease) dataset: 222 labeled sentences from 550 articles, 27 PICO elements grouped into 6 classes; 110 test articles. Proprietary dataset from pharmaceutical company (GLS); availability not stated but authors indicate sharing 27/137 elements (see Endnote OA-1); COVID-19 dataset: Obtained from open research challenge (L. Wang et al., 2020); 266 articles merged from https://www.kaggle.com/allen-institute-for-ai/CORD-19-research-challenge and https://www.kaggle.com/mlconsult/summary-page-covid-19-risk-factors. Custom annotated by research team with 30 years combined medical experience; EBM-NLP dataset: Public dataset, 5,000 abstracts of RCT studies; 200 expert-annotated abstracts reserved for testing. Available at https://github.com/bepnye/EBM-NLP; Wilson Disease (WD) dataset - proprietary, obtained from global life sciences company; COVID-19 dataset - merged from Kaggle sources (CORD-19 research challenge and COVID-19 risk factors dataset); EBM-NLP dataset - public, available at https://github.com/bepnye/EBM-NLPCode: Supplementary materials provided at https://osf.io/tsdza/?view_only=1c0507e9e88d4bce8f7b63c530880e78 (includes 137 PICO elements of Wilson Disease SR project, data annotation process, metric definitions, model implementation, and error analysis); Additional analyses and implementation details provided at https://osf.io/tsdza/?view_only=1c0507e9e88d4bce8f7b63c530880e78; Additional analyses and supplementary materials at https://osf.io/tsdza/?view_only=1c0507e9e88d4bce8f7b63c530880e78Extracted from: pdfAgreement 45%

Explore related topics

Related papers