Page 22 - Read Online

Page 47 Talwar et al. Art Int Surg. 2025;5:46-52 https://dx.doi.org/10.20517/ais.2024.81

Such models could understand text, but not create new text. With the release of Generative Pre-trained
Transformer-3 (GPT-3) in 2021, generative large language models (LLMs) have performed well on both
generative and non-generative tasks. Newer LLMs (i.e., GPT-4, Mistral, Claude, Bard, Perplexity) have only
improved in these capacities. Plastic surgeons must understand how NLP tools can be harnessed to improve
our workflows.

In plastic surgery, patient consultations are a critical component of the care process. Consultation requires
effective communication and thorough documentation. The integration of NLP into consultations can
enhance the quality of care and streamline this process. In this manuscript, we review the current state of
clinical NLP integration and provide a perspective for future growth.

NLP is revolutionizing the two overarching domains of documentation and communication, summarized in
Table 1. Examples of NLP tasks related to documentation include information extraction and
summarization, ambient transcription, and coding. NLP tasks related to communication include
understanding patient goals, patient-reported outcomes (PRO), translation, health literacy, and a patient-
facing chatbot. We also discuss ethical considerations, limitations, and challenges of clinical NLP. We are
still in the early stages of clinical NLP development. Plastic surgeons must help guide toward beneficial
applications.

DOCUMENTATION
Information extraction and summarization
Plastic surgery is a highly specialized discipline. Plastic surgeons operate across the whole body.
Unsurprisingly, our experts require very specific knowledge about patients and their medical history. This
includes how well their comorbidities are controlled (i.e., diabetes) and information about their surgical
history (i.e., history of abdominoplasty). Electronic health records (EHRs) contain a wealth of this
information. However, most patient EHRs contain hundreds of clinical documents. Manually searching for
information can be tedious due to the volume and jargon. For example, in a provider note, “PT” can mean
“patient”, “physical therapy”, “posterior tibial artery”, “posterior tibialis”, “prothrombin time”, or “part-
time”.

LLMs that comprehend clinical documentation should be able to understand the context and easily extract
information. The most basic task is “named entity recognition” - deriving the names of patients, medical
procedures, and medications directly written in a document . Plastic surgeons might use named entity
[1]
recognition to identify key surgical information, such as what tissue was resected, what type of mesh/
implant was used, and what flaps were used in the reconstruction.

Summarization, on the other hand, is a more complex and generative task. It requires a holistic
understanding of a text. Here, the generative LLMs could help plastic surgeons by processing several clinical
documents in the EHR to summarize a patient’s surgical history (i.e., all previous breast procedures) or
overall health status. This would help plastic surgeons prepare before or during a consultation to guide
operative planning. A group from Stanford applied eight different LLMs for clinical summarization and
found several untrained LLMs more completely summarized patient history than humans and had fewer
errors . Importantly, their study highlighted the importance of “prompt engineering”, that is, phrasing
[2]
questions so the model can generate useful summaries. Their study also found that GPT-4 had the best
performance compared to other models. Indeed, industry leader Epic © has already announced a
[3]
forthcoming integration of GPT-4 into its platform . This would enable surgeons to use Epic to
instantaneously extract and summarize information for patient care. It would also facilitate powerful