Professor of Anatomy Siddhartha Medical College, India
Purpose: Large language models (LLMs), such as ChatGPT, offer a unique opportunity for clinical decision support by providing rapid access to evidence-based guidelines, procedural recommendations, and patient-specific risk stratification. In interventional radiology (IR), where decisions are often complex and time-sensitive, LLMs could serve as valuable tools to assist practitioners. However, the feasibility of using these models in real-world clinical settings, particularly in IR, has not been thoroughly evaluated. This study aims to assess the feasibility, accuracy, and potential utility of ChatGPT in providing clinical decision support for interventional radiology procedures using simulated cases.
Materials and Methods: A set of 100 simulated interventional radiology cases was created, covering a range of procedures such as percutaneous drainage, venous access, and tumor embolization. Each case included relevant patient data (e.g., medical history, imaging results, and laboratory values). ChatGPT was prompted to provide procedural recommendations, risk assessments, and treatment plans based on the case data. The AI-generated responses were compared to expert clinician decisions to evaluate accuracy and relevance. Metrics for assessment included the accuracy of recommendations, the inclusion of critical procedural details, and the appropriateness of risk stratification.
Results: ChatGPT provided accurate and clinically relevant recommendations in 85% of the simulated cases (n=85/100), closely aligning with expert decisions. In 12% of the cases, the model offered partial information that required expert augmentation, while 3% of cases included recommendations that were inconsistent with established clinical guidelines. Common areas where ChatGPT excelled included identifying suitable procedural approaches and suggesting standard risk mitigation strategies. However, limitations were observed in nuanced cases that required deeper clinical insight, such as complex comorbidities or rare conditions.
Conclusion: ChatGPT demonstrated promising feasibility as a clinical decision support tool for interventional radiology, performing well in the majority of simulated cases. While the model provided accurate recommendations in most instances, a small percentage of cases revealed limitations, underscoring the importance of expert oversight. Further development and integration into clinical workflows could enhance the utility of LLMs in IR, particularly when used as a supplemental tool to assist clinicians with routine and complex decision-making.
.jpg)
