Page 63 - Read Online

Page 477 Bala et al. Art Int Surg. 2025;5:476-89 https://dx.doi.org/10.20517/ais.2025.44

75% to 88.9%. Chi-square tests confirmed statistical significance (P < 0.05) for most of the observed
improvements.

Conclusion: This study demonstrates that combining zero-shot generative AI with retrieval augmentation can
effectively convert unstructured patient reports into structured data. This approach offers a scalable and
interpretable method for adverse event monitoring in mesh implant surgeries and supports data-driven evaluation
of patient-reported outcomes.

Keywords: Mesh implantation, adverse events, Retrieval-Augmented Generation (RAG), hernia repair, GenAI

INTRODUCTION
Hernia mesh implants have become a cornerstone of modern surgical practice for repairing inguinal,
umbilical, incisional, and stress urinary incontinence (SUI)-related hernias. These devices are designed to
reinforce weakened tissue and reduce recurrence rates, thereby improving surgical outcomes and patient
[1]
recovery . However, despite their benefits, mesh implants are frequently associated with postoperative
complications such as chronic pain, infection, erosion, migration, bowel obstruction, and urinary
[2]
dysfunction . These adverse events can severely affect patients’ quality of life, increase healthcare costs, and
raise long-term safety concerns. As awareness of these risks grows, there has been an increased focus on
post-market surveillance systems to capture, analyze, and respond to real-world patient experiences in a
timely manner.

The U.S. Food and Drug Administration’s (FDA) Manufacturer and User Facility Device Experience
[3]
(MAUDE) database provides an invaluable resource for this purpose . It contains thousands of patient-
generated narratives detailing mesh-related complications, offering granular insights into device
performance in clinical practice. However, these reports remain underexploited for regulatory science and
safety monitoring. The primary challenge lies in their format: narratives are unstructured free-text with
inconsistent terminology, varying linguistic styles, and significant subjectivity . Extracting clinically
[4]
meaningful, structured insights from such data is challenging but essential to support evidence-based
decision making for both regulators and clinicians.

Traditional natural language processing (NLP) approaches have attempted to address this challenge,
typically using rule-based extraction systems, keyword searches, or supervised machine learning models
[5-7]
trained on annotated corpora . However, these methods face persistent limitations in handling linguistic
variability, negation, synonymy, and implicit mentions of adverse events . Additionally, they depend
[8]
heavily on expensive, domain-specific labeled datasets, which restrict their scalability and adaptability across
different medical contexts . As a result, adverse events that are described implicitly or in non-standard
[9]
terminology may remain undetected, leading to underestimation of device-related complications in
surveillance data.

In recent years, large language models (LLMs) such as Meta’s LLaMA 2 have shown remarkable promise in
addressing these limitations. With their ability to perform zero-shot prompting, LLMs can execute tasks
based solely on natural language instructions, eliminating the dependence on large annotated datasets [10-13] .
This is particularly advantageous in domains such as mesh implant surveillance, where specialized labeled
corpora are scarce or unavailable . However, despite these advantages, LLMs are prone to factual
[14]
inconsistency and hallucinations when operating without external knowledge grounding, posing significant
risks in safety-critical domains where accuracy and interpretability are paramount.