Page 26 - AIH-1-2

Artificial Intelligence in Health LLMs-Healthcare: Application and challenges

In addressing the above challenges in dermatological features in skin images, there are inherent challenges.
diagnostics, Zhou et al. introduced SkinGPT-4, an Several challenges associated with deploying SkinGPT-4
18
innovative interactive dermatology diagnostic system include ensuring consistent diagnostic accuracy across
underpinned by an advanced visual LLM. This study various skin conditions, safeguarding patient privacy
was mainly focused on tackling the prevalent issues in while managing sensitive health data, and integrating the
dermatology, such as the shortage of specialized medical technology seamlessly into existing healthcare systems. In
professionals in remote areas, the intricacies involved addition, despite SkinGPT-4’s high diagnostic accuracy,
in interpreting skin disease images accurately, and the continuous human oversight in medical diagnosis and
demanding nature of creating patient-friendly diagnostic treatment planning remains critical to complement the
reports. SkinGPT-4, utilizing a refined version of MiniGPT-4, AI’s capabilities with professional medical judgment
trained on an extensive dataset that included 52,929 images and ensure optimal patient care outcomes. In addition,
of skin diseases, both from public domains and proprietary advancements might focus on developing models that can
sources, along with detailed clinical concepts and doctors’ adapt to new, emerging skin conditions and leveraging
notes. This comprehensive training on skin-related disease telemedicine to extend dermatological care to remote
images endowed SkinGPT-4 to articulate medical features areas, thus promoting health-care equity.
in skin disease images using natural language and make
precise diagnoses. The functionality of SkinGPT-4 allows 4. Neurodegenerative disorders
users to upload images of their skin conditions, after Neurodegenerative disorders are characterized by the
which the system autonomously analyzes these images. gradual deterioration of specific neuron groups, differing
It identifies the characteristics and categorizes the skin from the non-progressive neuron loss seen in metabolic
conditions, performs an in-depth analysis, and provides or toxic conditions. These diseases are categorized by their
interactive treatment recommendations. A notable aspect primary symptoms (such as dementia, parkinsonism, or
of SkinGPT-4 is its local deployment feature, combined motor neuron disease), the location of neurodegeneration
with a solid commitment to maintaining user privacy, within the brain (including frontotemporal degenerations,
making it a viable option for patients seeking accurate extrapyramidal disorders, or spinocerebellar degenerations),
dermatological assessments. To ascertain the efficacy of or the underlying molecular abnormalities. Dementia is a
19
SkinGPT-4, the study conducted a series of quantitative broad category of brain diseases that cause a long-term and
evaluations on 150 real-life dermatological cases. Certified often gradual decrease in the ability to think and remember,
dermatologists independently reviewed these cases to affecting daily functioning. Alzheimer’s disease (AD) is the
validate the diagnoses provided by SkinGPT-4. Among most common cause of dementia, characterized by memory
the 150 cases, a commendable 78.76% of the diagnoses loss, language problems, and unpredictable behavior.
rendered by SkinGPT-4 were validated as either accurate LLM such as Google Bard and ChatGPT have emerged
or relevant by the dermatologists, breaking down into as valuable tools for predicting neurodegenerative
73.13% that firmly aligned and another 5.63% that disorders. A study by Koga et al. evaluated these
20
agreed. The outcomes of this evaluation underscored models’ predictive accuracy using cases from Mayo
the accuracy of SkinGPT-4 in diagnosing skin diseases. Clinic conferences. The researchers extracted 25 cases of
While SkinGPT-4 is not positioned as a replacement for neurodegenerative disorders, from among the cases in
professional medical consultation, its contribution to the Mayo Clinic brain clinicopathological conferences,
enhancing patient comprehension of medical conditions, as their sample pool. These clinical summaries were then
improving communication between patients and doctors, utilized for training and testing the models. The diagnoses
expediting dermatologists’ diagnostic processes, and offered by each model were compared against the official
potentially fostering human-centered care and health-care diagnosis provided by medical professionals. Findings
equity in underdeveloped regions is significant. from the study highlighted that ChatGPT-3.5 aligned with
32% of all the physician-made diagnoses, Google Bard
3.1. Challenges associated with utilizing LLMs in with 40%, and ChatGPT-4 with 52%. When assessing the
dermatology
accuracy of these diagnostic predictions, ChatGPT-3.5
The introduction of SkinGPT-4 by Zhou et al. marks a and Google Bard both achieved a commendable score of
18
significant advancement in dermatological diagnostics, 76%, while ChatGPT-4 led the pack with an impressive
addressing challenges such as dermatologist shortage, and accuracy rate of 84%. The evident proficiency exhibited by
simplifying skin disease image interpretation and patient- LLMs, specifically ChatGPT and Google Bard, highlights
friendly report generation. Despite its innovative approach their considerable potential in revolutionizing diagnostic
and the training on an extensive dataset to articulate medical processes in neurodegenerative disorders.

Volume 1 Issue 2 (2024) 20 doi: 10.36922/aih.2558