Page 8 - AIH-2-2

Artificial Intelligence in Health Role of LLMs in improving patient experience

1. Introduction facilitated the automated composition of scientific articles
and research, prompting concerns that they may adversely
Large language models (LLMs) employ artificial impact critical thinking and reasoning. Finally, progress in
10
intelligence (AI) to simulate responses that are comparable the field of LLMs has been driven by for-profit companies,
to those of humans. These models are typically trained on which significantly impacts access and restricts the research
1
vast amounts of data, which are frequently accessible on areas that can be explored.
internet, ensuring that the model can respond to queries
based on keywords, thereby generating the relevant This scoping review aims to explore the current status
information. The rapid development of LLMs has been of LLMs in improving provider–patient experience and
considerably stimulated by the enhanced efficacy of interaction efficiency. The exact context of their use
natural language processing (NLP) models, computational may vary; some may be patient-facing, while others
2,3
power, and increased access to large datasets. OpenAI may be physician-facing. Their objective may be to train
introduced the Generative Pre-trained Transformer physicians or other healthcare professionals or to at
(GPT)-1, followed by other models from companies least partially replace them for patient communication.
such as Meta and Google, expanding the use of LLMs This communication may also be intended to offer
4
into the public domain. With the release of ChatGPT-4, technical information or emotional support and solace.
Llama, and ChatBot BARD, LLMs have gained popularity Through this work, we aim to gain a more comprehensive
and are largely used, as they can now be integrated into understanding of the diverse role that LLMs currently
independent software as plugins. 5 occupy in this field and to examine the parameters on
which they are being evaluated. This understanding will
LLMs have permeated practically every scientific enable the development of more relevant, accurate, and
discipline, with the primary appeal being the ability better-performing LLMs, as well as the identification of
to assimilate large amounts of data and facilitate user their deficiencies and inconsistencies, so that they can be
navigation. This has eliminated the complexity of search addressed in subsequent iterations.
terminologies and strategies. Simple conversational
prompts are now available to assist users in their search 2. Methods
for pertinent information, which is presented to them in a
structured, coherent manner. However, there are numerous 2.1. Overview
concerns regarding their application in healthcare, which The Preferred Reporting Items for Systematic Reviews and
remain a barrier to their widespread adoption, similar to Meta-Analyses extension for Scoping Reviews checklist
other disciplines. First, the accuracy and efficacy of these was employed. The scoping review methodology adopted
11
models in a clinical setting are contingent upon the veracity was consistent with the approach proposed by Tricco et al.
and authenticity of the data on which they are trained. Large This included identifying the research question, isolating
datasets frequently present inaccurate data, which can and selecting articles, charting, collating, summarizing
misinform or misguide users, which is a significant concern. the data, and reporting the results. Objectivity may be a
6
When utilized in the healthcare sector, it is crucial to have challenge when evaluating the adherence of the selected
curated and authenticated data on which these models are articles to the research question and the inclusion/
trained. Companies, such as Google have achieved this with exclusion criteria, as the research question is sufficiently
applications, such as Med-PaLM. Another area of concern broad. The articles to be included were carefully selected
is “hallucination” or “confabulation,” where LLMs generate after applying the objective exclusion criteria illustrated in
false outputs or unsubstantiated responses to questions. Figure 1 (articles published before 2015, those in a language
The precise cause of this phenomenon remains uncertain; other than English, and those using LLMs for manuscript
however, detecting these is a crucial part of refining them preparation). After these exclusions, approximately
for medical applications. Determining the single source 84 manuscripts remained, which were independently
7
of truth for complex queries for which data have been evaluated by two researchers (ABV and NS). Of these, 47
aggregated from millions of sources will remain a significant manuscripts met our inclusion criteria and were finally
challenge. Using guardrails to filter inputs or outputs for included for consideration. The detailed PubMed search
8
these healthcare-related LLMs is essential to ensure that the strategy is provided in Supplementary File 1.
responses are factually accurate, contextually pertinent, and
consistent in nature. Infringement of copyrights is another 2.2. Identifying the research question
major concern; these LLMs may generate text that is both Our study aimed to understand the current role of
plagiaristic and infringes upon copyrighted texts. In the LLMs in improving provider–patient experience and
9
realm of healthcare research and education, LLMs have interaction efficiency. Our primary research question was

Volume 2 Issue 2 (2025) 2 doi: 10.36922/aih.4808