Back to 2025 Abstracts
From Months To Minutes: Evaluating An AI-driven Approach To Systematic Reviews In Plastic Surgery
Moreen W. Njoroge, BA1, Thandolwethu Z. Dlamini, MSc
2, Jordan Gornitsky, MD
1, Lily R. Mundy, MD
1, Carisa M. Cooney, MPH
1, Ala Elhelali, PhD
1, Abby D. Liu, BA
1, Thalia Liu, BA
1, Robin Yang, MD, DDS
1, Richard J. Redett, MD
1.
1Johns Hopkins University School of Medicine, Baltimore, MD, USA,
2Massachusetts Institute of Technology, Boston, MA, USA.
Purpose: Systematic reviews are essential for evidence-based plastic surgery but hindered by manual article screening. We developed an artificial intelligence (AI) application to automate systematic review stages to accelerate evidence synthesis while maintaining accuracy.
Methods: We created a multi-agent natural language processing workflow using GPT-3.5 and GPT-4 for article screening. Testing involved four published systematic reviews in plastic surgery, covering digital replantation to facial nerve grading systems. AI workflow performance was compared to human-conducted reviews at title/abstract and full-text screening stages.
Results: The AI processed 5,769 studies across four systematic reviews, completing screening in 1.5-13.1 minutes per review based on complexity. Title/abstract screening achieved 98.4% precision (95% CI: 97.8-98.9%) and 92.5% recall (95% CI: 91.6-93.3%). Full-text evaluation reached 99.3% precision (95% CI: 98.9-99.6%) and 98.6% recall (95% CI: 98.1-99.0%). The AI excelled in identifying quantitative outcomes but showed limitations in nuanced clinical evaluations. Average computational cost was $18.66 per review, significantly below manual review costs.
Conclusions: This AI approach transforms reviews in plastic surgery, reducing time and cost while maintaining accuracy and expediting research implementation in clinical practice. Future development will focus on creating a user-friendly interface for broader accessibility. AI integration could revolutionize evidence-based plastic surgery through frequent updates, enhancing clinical decision-making.
Back to 2025 Abstracts
