Keywords
artificial intelligence - vascular and interventional radiology - large language models
- machine learning - radiology - patient-centered care
Introduction
Over the past decade, the popularity and utilization of artificial intelligence (AI)
within the healthcare industry have experienced a significant surge.[1] This has led to the potential for substantial enhancements in healthcare efficiency
and precision, ultimately resulting in improved patient care, more informed decision-making,
and considerable cost savings. AI facilitates the evaluation and analysis of various
disease conditions, often using complex and rapidly expanding datasets. Many times,
AI achieves remarkable accuracy and depth in these tasks by harnessing cutting-edge
concepts and technologies such as machine learning (ML), neural networks (NNs), and
large language models (LLM).[2]
The discipline of vascular and interventional radiology (VIR) has also seen significant
advancements in recent years, with an increase in the number of procedures performed,
subspecialization, and improved accuracy in interventions.[3] The demonstrated AI applications in various VIR use cases abounds including AI-assisted
endovascular clot retrieval for acute ischemic stroke, predicting responses to transcatheter
arterial chemoembolization in hepatocellular carcinoma, AI-guided ultrasound in echocardiography,
and angiography-based ML algorithms for real-time estimation of fractional flow reserve.[4]
[5] So, as VIR techniques and procedures mature and achieve wider acceptance, it could
prove more useful in multidisciplinary care with the incorporation of AI potentially
leading to improved outcomes for all stakeholders.[6]
In this context, VIR is the frontier of the concept of Imaging 3.0, an initiative
aimed at showcasing the expanded contributions of radiologists beyond conventional
image interpretation.[7] In detailing key aspects, VIR shows its pivotal role in percutaneous, image-guided
procedures such as abscess drainage and needle biopsy. This emphasizes the specialty's
capacity to reduce morbidity, enhance patient outcomes, and offer cost-effective alternatives
to other surgical interventions. Additionally, the significance of transjugular intrahepatic
portosystemic shunts and the central role of interventional radiologists in delivering
cost-effective central venous access services are emphasized, making substantial contributions
to multidisciplinary cancer treatment.
While many of the AI-used cases in VIR primarily fall within the domains of computer
vision and image classification/segmentation in AI, the impact of AI in VIR potentially
extends beyond these conventional boundaries. LLMs are a promising domain in AI with
opportunities for substantial applications in VIR. LLMs are advanced AI models, such
as Open AI's Chat Generative Pre-Trained Transformer (ChatGPT), which are capable
of understanding and generating human-like text.[8] LLMs have emerged as a revolutionary breakthrough in AI, as it pertains to natural
language processing (NLP). Their ability to process and comprehend language has propelled
them into diverse sectors, including finance, marketing, and healthcare.
In general healthcare, the LLM technology is now widely acclaimed for its role in
medical language tasks, including automated report generation and integration with
healthcare systems.[9] It enables the comprehensive extraction of patient data from electronic health records,
laboratory results, and prior imaging studies, significantly enhancing the diagnostic
process and patient care. The literature on the potential application of LLMs in VIR
is, however, sparse relative to the broader discussions of opportunities and concerns
provoked by AI generally but more importantly, the readily accessible LLM technologies
like ChatGPT.
Hence, the aim of this article is to comprehensively explore the potentials of LLMs
in VIR. By reviewing the current abilities of LLMs, the aim is to highlight the potentials
of this technology, outlining its current and prospective contributions to advancing
clinical practices, patient outcomes, and educational activities in VIR. The challenges,
potential future directions, and advancements of LLMs in the field of VIR are also
reviewed.
Overview of LLMs
LLMs are gaining significant traction as invaluable tools in the field of radiology.[10] These advanced AI models work by transforming text into numerical tokens, thus capturing
contextual information during training.[11] Consequently, they predict and select plausible next tokens based on learned language
patterns, creating the appearance of encyclopedic knowledge and reasoning. ChatGPT
is one of the popular LLMs, developed by OpenAI and made publicly available in November
2022. It is trained on extensive text datasets in various languages and can produce
human-like responses to text input. ChatGPT utilizes the GPT architecture to process
natural language and generate context-based responses. Other publicly available LLMs
include T5, Pythia, and LlaMA.[12]
[13]
[14] For an overview of additional open access LLMs suitable for personal and research
applications, the reader is referred to a detailed Github repository.[15]
The scientific community has shown diverse reactions toward ChatGPT, reflecting the
ongoing debate surrounding the benefits and risks of LLMs and generative AI technologies
in general.[16] On one hand, ChatGPT and other LLMs have demonstrated usefulness in conversational
and writing tasks in medicine, enhancing output efficiency and accuracy.[9] On the other hand, concerns have emerged regarding potential bias in its training
datasets, leading to limitations and factual inaccuracies, a phenomenon referred to
as “hallucination.” Additionally, there are security concerns related to the spread
of misinformation and the possibility of cyber-attacks utilizing LLMs.[17]
The Present and Promising Future of LLMs in VIR
The Present and Promising Future of LLMs in VIR
In a notable experiment, ChatGPT 3.5 demonstrated impressive performance on 376 USMLE
test questions from the June 2022 sample exam, achieving a passing or near-passing
score threshold of 60%, while exhibiting high concordance and insightful responses,
without specialized training.[18] A more recent study indicates that the latest version, ChatGPT 4 performed even
better and demonstrated remarkable medical reasoning.[19]
[20] Additionally, Yan et al[21] introduced RadBERT, a language model fine-tuned for radiology, excelling in NLP
tasks and promising automation in abnormal findings identification and report creation.
These advances could alleviate healthcare workload and burnout, suggesting future
developments in automated radiology reports and broader healthcare applications. It
is fascinating that this technology can achieve such outcomes in its early stages,
and without domain-specific training, more so the LLMs can reason through novel problems
to a remarkable degree without specific training, a phenomenon known as “zero shot.”[22]
Considering these factors, alongside the novel capabilities of LLMs, their potential
applications are noteworthy. While the literature on this emerging subject is still
limited, key areas where LLMs demonstrate promising applications in VIR are highlighted
([Table 1]).
Table 1
Summary of current and future uses of LLMs in VIR
|
S/N
|
Application
|
Specifics
|
References
|
|
1.
|
Supporting clinical-decision-making
|
• Analyzing medical literature, electronic health records, and patient data
|
[23]
|
|
|
• Assisting in disease diagnosis, treatment planning, and prognostic predictions
|
[24]
|
|
|
• Providing evidence-based recommendations and facilitating personalized treatment
strategies
|
[16]
|
|
|
• Continuous learning from new data for evolving recommendations in a dynamic VIR
landscape
|
[9]
[18]
|
|
2.
|
Improving clinical workflow and patient scheduling
|
• Use in radiology report generation, reducing addendum requests and improving reporting
processes
|
[25]
[26]
|
|
|
• Intelligent patient scheduling for risk identification and preventive measures
|
[27]
|
|
|
• Handling administrative duties like patient billings and extracting relevant summaries
from patient records
|
[27]
[28]
|
|
|
• Alleviating healthcare provider workload and reducing risks to patients
|
[29]
|
|
3.
|
Enhancing VIR education
|
• Assisting medical students and trainees in board-style examinations
|
[30]
|
|
|
|
• Synergizing with attending physicians for a comprehensive learning experience
|
[31]
|
|
4.
|
Patient education and patient-centered care in VIR
|
• Simplifying medical reports for patient understanding
|
[32]
|
|
|
• Providing patient education on VIR procedures with potential improvements in accuracy
|
[33]
|
|
|
• Generating patient-friendly explanations of complex medical conditions, treatment
options, and risks
|
[35]
|
Abbreviations: LLMs, large language models; VIR, vascular and interventional radiology.
Supporting Clinical Decision-Making
LLMs, equipped with sophisticated NLP methods, possess the capability to analyze extensive
volumes of medical literature, electronic health records, and patient data. By processing
and understanding the intricate patterns and nuances within this information, LLMs
can assist the interventional radiologist in making more informed and precise decisions
regarding disease diagnosis, treatment planning, and prognostic predictions.
Shen et al[23] have shown that ChatGPT can use large knowledge bases to swiftly answer questions
about the most suitable imaging study for specific clinical scenarios. A recent study
assessed the performance of two LLMs, ChatGPT and Glass AI, in predicting optimal
neuroradiology imaging modalities compared with an experienced neuroradiologist.[24] Both LLMs scored similarly at 1.75 and 1.83, respectively, out of a maximum possible
3 points, while the neuroradiologist outperformed with a score of 2.20. ChatGPT showed
greater variability, suggesting room for improvement with targeted medical text training,
unlike Glass AI, which has more precise training on medical literature. Furthermore,
ChatGPT showed promising prospects to enhance diagnostic accuracy, streamline workflow,
and improve patient care by providing evidence-based recommendations and facilitating
personalized treatment strategies.[16]
Additionally, the ability of LLMs to continuously learn from new data ensures that
their recommendations evolve with the dynamic landscape of VIR. The regenerative attribute
of LLMs makes it amenable to keep pace with the prolific medical devices industry
stocking the cath laboratories. Evidence indicates improved performance on clinical
tasks when LLMs are trained on domain-specific clinical data.[9]
[18] Based on this, there is some optimism that LLMs trained on VIR specific data will
offer clinical decision support utility for the interventional radiologist, particularly
in suggesting relevant procedures and appropriate treatment modalities including device
choice and compatibility. However, it is crucial to acknowledge that the expertise
of a trained interventional radiologist remains indispensable for interpreting and
verifying the outputs from LLMs.
Improving Clinical Workflow and Patient Scheduling
AI has the potential to improve the interventional radiologist's daily practice in
various ways. For instance, structured reporting has been shown to lead to a reduction
in addendum requests for insufficient documentation, indicating a more comprehensive
and clear reporting process.[25] Recognizing this unique need, recent studies have explored the use of LLMs in radiology
report generation (R2Gen). R2GenGPT is an emerging innovative framework for R2Gen,
which leverages LLMs for automated radiology reporting.[26] It demonstrates state-of-the-art performance and reduced computational complexity.
Incorporation of LLMs similar to R2GenGPT as adjuncts for generating structured VIR
report holds promise for further improving the clinical practice workflow.
Another significant aspect is intelligent patient scheduling, where AI can identify
patients at high risk and take precautions to avoid potentially preventable morbidities
or mortalities, while also reducing the chances of missing necessary care through
smart scheduling and patient selection.[27] Though the core of these algorithms may be based on complex supervised learning
models or advanced ML techniques other than LLMs, LLMs still possess the ability to
be the front-end conversational interface. This interface would be able to take the
output of the back-end AI models and present it in an intelligible and interactive
manner.
Furthermore, LLMs can handle administrative duties such as patient billings, and extract
relevant summaries from a patient's records such as problem lists, clinical notes,
laboratory data, pathology reports, vital signs, prior treatments, and prior imaging
reports.[27]
[28] These summaries provide the interventional radiologist with crucial contextual information
for clinical uses. Patel and Lam[29] demonstrated the utility of ChatGPT in creating discharge summaries, allowing the
clinicians to focus on more clinical commitments. Employing LLMs for laborious tasks
like these also potentially reduces risks to the patient.
Enhancing VIR Education
The potential of LLMs in VIR education for medical students and trainees is promising.
Recently, ChatGPT demonstrated impressive performance on a radiology-board style examination,
correctly answering 69% of questions.[30] It excelled in lower-order thinking questions but faced challenges with higher-order
thinking questions, particularly those related to describing imaging findings, calculations,
classifications, and applying concepts. Another study compared ChatGPT-4 and Bard
(developed by Google) in responding to questions from the American College of Radiology's
Diagnostic Radiology In-Training (DXIT) examination. ChatGPT-4 exhibited a higher
overall accuracy of 87.11% compared with Bard's 70.44%. Despite occasional failures
in addressing questions accurately, the authors expressed cautious optimism, suggesting
that LLMs like ChatGPT-4 could serve as valuable study tools for trainees in the future.
Consequently, the VIR-specific trained LLM can assume a crucial role in the learning
curve of the VIR trainee within a learner-centered collaborative training framework.
Within this framework, the LLM acts as an immediate repository, delivering a trove
of updated literature, procedural guidelines, and case studies to enrich the learning
experience. Proficient in evaluating lower-order thinking questions, it also becomes
an invaluable tool in gauging the trainee's foundational knowledge. As the trainee
confronts higher-order challenges, the attending physician and LLM synergize, addressing
complexities and filling knowledge gaps. The LLM's identified limitations in imaging,
procedural descriptions, and calculations are mitigated by the attending physician's
expertise, creating a dynamic feedback loop for comprehensive learning. Such a personalized
adaptive learning pathway aligns with the findings of Duong et al,[31] demonstrating the potential benefits of AI in achieving “precision education” within
the field of radiology. The unique strengths and weaknesses of the trainee are harnessed
to achieve superior learning experience facilitated by the personalized integration
of LLMs thus heralding a new dawn in enhanced VIR education.
Patient Education and Patient-Centered Care in VIR
The integration of LLMs into VIR education extends beyond the training of the workforce.
It holds some promise in enhancing patient-centered care through patient education
as well. An exploratory case study conducted by radiologists revealed promising results
in utilizing ChatGPT to simplify medical reports while maintaining factual accuracy,
completeness, and patient safety.[32] Scheschenja et al[33] also explored the viability of using LLMs, specifically ChatGPT-3 and ChatGPT-4,
for patient education in VIR. The authors designed hypothetical questions about common
VIR procedures, comparing the accuracy of responses from the two models. While both
models provided accurate information on general procedure details, preparation, risks,
and postinterventional aftercare, ChatGPT-4 demonstrated better overall accuracy than
ChatGPT-3 in answering questions related to Port Implantation, PTA, and TACE procedures.
Recognizing the complexities associated with ensuring language clarity and response
coherence, they concluded still that the LLMs exhibit feasibility for safe and relatively
accurate patient education in VIR, with GPT-4 showing incremental improvements.
Other authors have reported similar results after investigating ChatGPT's performance
on VIR knowledge.[34]
[35] McCarthy et al[35] evaluated the LLM's efficacy in delivering educational content on VIR, comparing
it to standard material from the Society of Interventional Radiology Patient Care
Web site. Despite occasional inaccuracies and a tendency to produce lengthy and somewhat
complex responses, ChatGPT was generally deemed a reliable source for most VIR procedures.
By leveraging the capabilities of LLMs in this manner, interventional radiologists
can generate patient-friendly explanations of complex medical conditions, treatment
options, and potential risks associated with VIR procedures. This empowers patients
with accessible information, fostering a better understanding of their health conditions
and treatment plans. Moreover, the LLM's ability to produce human-like text can enhance
communication between healthcare professionals and patients, fostering a more empathetic
and transparent doctor–patient relationship. Hopefully, further improvements will
minimize instances of incorrect information and lead to safer patient education.
Challenges, Ethics, and Recommendations in LLM Implementation for VIR
Challenges, Ethics, and Recommendations in LLM Implementation for VIR
Although the incorporation of LLMs into VIR holds great promise, it is also fraught
with numerous challenges and ethical considerations.[36] These include reduced human involvement, potential harm resulting from LLM reasoning
weaknesses, limited availability of comprehensive datasets, the risk of biases leading
to healthcare disparities, and cost constraints in low-resource settings. An overdependence
on AI has the potential to diminish human involvement in decision-making processes.[37] To counteract this trend, it is essential for LLMs to augment human expertise rather
than replace it, emphasizing the continued centrality of interventional radiologists.
The assessment of the generative capabilities of LLMs, particularly in the context
of VIR, heavily relies on the availability of comprehensive datasets. The scarcity
of medical data from VIR operating suites poses a significant obstacle to effective
data collection.[38] Also, LLMs encounter challenges such as hallucinations and weak numerical reasoning.[36]
[39] When applied in patient care without due caution, these issues can lead to severe
harm or even fatal consequences, underscoring the urgency of developing improved mitigation
techniques.
To address these concerns, recent research has introduced effective strategies. These
include the integration of external tools such as code interpreters, retrieval augmentation,
knowledge graphs, and other mathematical tools.[40]
[41]
[42] These measures aim to enhance the reliability and safety of LLM applications in
VIR, ensuring that they contribute positively to healthcare without compromising patient
well-being.
Also, an issue of utmost concern revolves around bias and fairness, and addressing
these are pivotal concerns in VIR. LLMs acquire knowledge from the data they are trained
on. If this data contains inherent imbalances and biases against certain races or
group of peoples, there is a risk of replicating these biases in its AI predictions.[32]
[43] This, in turn, could result in unfair outcomes, potentially worsening existing healthcare
disparities. So, there must be emphasis on fairness-aware ML and transparent development,[44] where data are well balanced and representative of all groups, and any patient data
are well protected to ensure privacy and security.
Moreover, the widespread adoption of LLMs faces challenges due to cost and resource
constraints, particularly in low-resource settings.[45] Addressing this, efficient transfer models and leveraging cloud resources can enhance
the accessibility of LLMs.[46] Fostering public–private collaboration can also help distribute costs and resources,
facilitating broader adoption.
Given these challenges, additional research is needed to assess the performance of
LLMs in clinical VIR. These investigations should be organized around the various
phases of clinical interactions: preoperative, perioperative, and postoperative care.
Evaluation criteria should include conventional AI metrics like specificity, sensitivity,
and F1-score. Moreover, it is crucial to also employ metrics tailored to LLMs, such
as BLEU, ROUGE, BERT Score, and LLM-EVAL.[47]
[48]
[49]
Ultimately, rigorous validation, ongoing monitoring, and collaboration with medical
experts and VIR specialists are crucial on all these bases. Creating a regulatory
framework that spans the multiple disciplines is crucial for the secure integration
of LLM into VIR practice. This ensures transparency and accountability without impeding
advancements. Addressing these challenges and ethical concerns can maximize LLMs'
potential to improve healthcare outcomes.
Conclusion
The incorporation of LLMs into the domain of VIR signifies a promising frontier poised
to enhance the discipline. Although the current landscape indicates that the widespread
implementation of LLMs in VIR may be premature, their potential holds the promise
of improving various aspects of the practice. These advanced AI tools have the capacity
to improve clinical decision-making, streamline workflow, enhance medical education,
and facilitate patient-centered care.
Moreover, full integration of these technologies into the clinical workflow of VIR
necessitates further exploration of multi-modal AI. This involves leveraging the text
and language capabilities of LLMs in conjunction with computer vision AI models, recognizing
the inherently visual nature of VIR. Additional research in this direction is crucial
to unlock the full spectrum of benefits and possibilities that LLMs can bring to the
field.
Effective use of LLMs in VIR also requires recognizing challenges and ethical considerations,
such as AI over-reliance, potential misinformation, and the need for rigorous validation.
Collaboration among radiologists, AI researchers, and regulators is essential for
balancing LLMs' potential with patient safety. Unlocking LLMs' full potential in VIR
also requires training and refining for domain nuances, implementing robust frameworks,
and adhering to ethical standards. This fosters a new era of medical practice, blending
human expertise with advanced AI for patient-centered care and innovation.
