Page 11 - AIH-2-2

Artificial Intelligence in Health Role of LLMs in improving patient experience

(especially in the context of health communication Transparency and trust are also concerns, particularly
and when the patient has queries in between follow- in “black box” AI systems where the efficacy may be
ups), ensure that the patient has access to more high but the decision-making process is unclear; here
validated and accurate health information (which “explainability” is essential for physician trust, as the
is frequently a challenge), and increase patient system can establish the reasoning behind a particular
satisfaction. Especially when dealing with a particular decision. Bias is a significant concern because the
subspecialty, it is substantially simpler to curate and data on which the model is trained could have
verify information that will be accessible to the patient. intrinsic gender, racial, ethnic, or other biases that
For diverse patient populations, other advantages the model has not considered and may reflect in the
include multilingual abilities and cross-cultural results. The other concerns, such as cybersecurity and
adaptation. It may also relieve the existing nursing or accountability for the outcomes, are self-evident.
physician assistant team of the burden of addressing • Role of NLP: This has been extensively examined
routine patient queries, allowing them to focus on in 15 of the articles included in our review, with
more urgent tasks. They may also contribute to the a particular emphasis on the efficacy of current
alleviation of the burden on healthcare professionals, NLP systems in a clinical context, for both general
who are often confronted with an increasing amount use and subspecialty use cases. The scope of
20
of administrative and non-medical work, resulting in applications was wide, including performing routine
increased levels of burnout. 58-59 cardiology and predicting future opioid use,
21
54
• Role in teleconsultations: In Saudi Arabia, a study suicidal behavior, substance abuse, neuropsychiatric
18
52
by Alanzi et al. examined the perspectives of 54 disturbances, 48-50 and ICU readmission. These are
healthcare experts regarding the impact of ChatGPT attractive propositions of a screening tool, especially
on teleconsultation. They specifically examined the in settings where mental health services are limited or
positive influence of ChatGPT on the following: unavailable but are crucial for all patients. They could
informational support, diagnostic assistance, act as a triage tool to ensure that high-priority patients
22
communications, enhancing efficiency, cost and receive urgent referrals or care. One article described
time savings, personalizing patient care, multilingual the construction of a clinical LLM, GatorTron,
support, decision-making, documentation, assisting in using >90 billion words of text, of which >82 billion
medical research, and enhancing team collaboration. were deidentified clinical text. This clinical text was
The issues that were addressed included misdiagnosis evaluated based on clinical concept extraction, medical
and errors, issues in personalized care, limited medical relation extraction, semantic textual similarity, natural
context, communication challenges, increased language inference, and medical question answering.
dependence on technology, ethical and legal issues, Scaling up the volume and quality of the training data
liability, and security and privacy. The conclusion that considerably enhanced the output.
the authors arrived at was that despite the enthusiasm • Screening for potential risks: An interesting use of
42
of most healthcare workers about the potential benefits LLMs was demonstrated by Guevara et al., who
of the platform, using it in its present form led to more demonstrated how they could be used to extract data
questions than answers. Having a standardized, secure, related to six specific social determinants of health
optimized platform that capitalized on the capabilities – employment, housing, transport, parental status,
of NLP and machine learning could resolve many relationships, and social support – from narrative
shortcomings in the current working system. text in the electronic health records. This can assist
• Ethical challenges: In their review, Jeyaraman et al. in more effective screening of patients who may
43
succinctly summarize the ethical dilemmas associated require additional assistance and are at higher risk of
with the application of AI in healthcare. Privacy is a complications to help potentially mitigate these in a
55
major concern with the use of big data in healthcare, timely manner. Another study attempted to predict
which can manifest in diverse ways, including privacy the 10-year cardiovascular risk in patients using
breach, misuse of patient data, denial-of-service, and LLMs, revealing comparable efficiency compared with
ransomware attacks in life-threatening situations. conventional nomograms and risk assessment models.
Despite the existence of blueprints for ChatGPT to filter
out patient-identifiable information and comply with 3.4. Contexts of use
regulations such as the Health Insurance Portability The articles included in our scoping review examined
and Accountability Act and General Data Protection the use of LLMs in several discrete settings. The first and
Regulation, the current systems are far from ideal. most prevalent application was a patient-facing interface
60

Volume 2 Issue 2 (2025) 5 doi: 10.36922/aih.4808