COMPARE: Clinical Optimization with Modular Planning and Assessment via RAG-Enhanced AI-OCT: Superior Decision Support for Percutaneous Coronary Intervention Compared to ChatGPT-5 and Junior Operators

Wei Fang, Chiyao Wang, Wenshuai Ma, Hui Liu, Jianqiang Hu, Xiaona Niu, Yi Chu, Mingming Zhang, Jingxiao Yang, Dongwei Zhang, Zelin Li, Pengyun Liu, Jiawei Zheng, Pengke Zhang, Chaoshi Qin, Wangang Guo, Bin Wang, Yugang Xue, Wei Zhang, Zikuan Wang, Rui Zhu, Yihui Cao, Quanmao Lu, Rui Meng, Yan LiarXiv 2025 · DOI · arXiv

Abstract

Percutaneous Coronary Intervention (PCI) guided by Optical Coherence Tomography (OCT) is essential for optimizing patient outcomes, yet its efficacy is often limited by the high degree of operator dependency in image interpretation. This paper introduces CA-GPT, a specialized AI system that integrates automated vascular imaging analysis with a Retrieval-Augmented Generation (RAG) framework. By grounding a large language model (LLM) in a curated knowledge base of over 100,000 annotated cases and clinical guidelines, CA-GPT provides real-time, evidence-based decision support. In a retrospective study of 96 patients, CA-GPT significantly outperformed both general-purpose AI (ChatGPT-5) and junior physicians, achieving a median agreement score of 5/5 with senior experts and demonstrating superior accuracy in critical tasks such as stent sizing.

Introduction

Percutaneous Coronary Intervention (PCI) is a cornerstone of modern cardiology, used to restore blood flow in obstructed coronary arteries. To ensure the long-term success of these procedures, clinicians increasingly rely on Optical Coherence Tomography (OCT), which provides high-resolution, 3D visualizations of the vessel wall. However, the interpretation of OCT images is notoriously difficult and highly "operator-dependent." Junior physicians often face a steep learning curve, and even experienced practitioners may vary in their assessments of plaque morphology and vessel dimensions.

While general-purpose large language models (LLMs) like ChatGPT-5 have shown potential in medical contexts, they are prone to "hallucinations"—generating factually incorrect but plausible-sounding advice. Furthermore, they lack the domain-specific grounding required for high-stakes surgical planning. To address these gaps, the authors developed the CA-GPT system, a modular framework designed to combine the precision of automated image analysis with the sophisticated reasoning of a medical-grade AI.

Methodology

The Modular Architecture

CA-GPT utilizes a dual-layer "small- and large-model" framework to bridge the gap between raw data and clinical decision-making:

  1. The Small Model Layer: This layer consists of 13 specialized algorithms designed for automated image processing. These tools perform tasks such as segmenting the vessel lumen, characterizing plaque types (e.g., calcified vs. lipid-rich), and calculating OCT-derived Fractional Flow Reserve (OCT-FFR) to assess the physiological significance of a lesion.
  2. The Large Model Layer (CA-GPT): The central reasoning engine is built upon DeepSeek-R1 (14B parameters). This model processes the quantitative outputs from the small model layer to generate clinical recommendations.
  3. Retrieval-Augmented Generation (RAG): To ensure safety and accuracy, the system employs a RAG framework. Instead of relying solely on pre-trained weights, the AI queries a curated knowledge base containing over 100,000 annotated PCI cases and the latest international expert consensus documents. This ensures that every recommendation is "evidence-traceable."

Experimental Design

The system was evaluated using a retrospective dataset of 96 patients (160 lesions) from Tangdu Hospital. The performance of CA-GPT was compared against ChatGPT-5 and junior physicians (1–5 years of experience). A "gold standard" was established by a panel of senior experts with over 10 years of experience. The primary endpoint was a "total agreement score" (0 to 5) across ten metrics covering pre-PCI planning and post-PCI assessment.

Results

Superior Decision Support

CA-GPT demonstrated a clear advantage in clinical decision-making. In pre-PCI planning, CA-GPT achieved a median agreement score of 5/5, significantly higher than the 3/5 achieved by ChatGPT-5 and the 4/5 achieved by junior physicians.

Precision in Stent Selection

One of the most critical aspects of PCI is selecting the correct stent diameter and length. In this task, CA-GPT achieved 90.3% agreement with senior experts. In contrast, junior doctors reached only 72.2% agreement, and ChatGPT-5 lagged behind at 63.9%. This suggests that CA-GPT could effectively mitigate errors caused by human inexperience.

Performance in Complex Cases

The performance gap was most pronounced in complex clinical scenarios, such as patients presenting with Acute Coronary Syndrome (ACS) or severely calcified lesions. In these high-pressure situations where human judgment is most likely to falter, CA-GPT’s data-driven approach remained consistent and highly aligned with expert consensus.

Discussion

Domain Specificity vs. Model Size

The findings highlight a critical lesson in medical AI: domain-specific grounding is more important than raw parameter count. Despite being significantly smaller than ChatGPT-5, the specialized CA-GPT was far more accurate. This underscores the value of the RAG framework in providing a safety mechanism against hallucinations by tethering AI outputs to validated medical evidence.

Efficiency and the "Junior Doctor Gap"

Beyond accuracy, the system offers substantial efficiency gains, reducing the time required for complex image interpretation from several minutes to mere seconds. By serving as a digital "co-pilot," CA-GPT has the potential to elevate the performance of junior operators to near-expert levels, potentially standardizing the quality of care across different hospital settings.

Limitations and Future Directions

While promising, this study was retrospective and conducted at a single center. Future research must involve prospective, multi-center trials to validate the system's performance in live surgical environments. Additionally, while the system shows high agreement with expert decisions, further study is required to determine if these AI-assisted decisions lead to improved long-term clinical outcomes, such as reduced rates of Major Adverse Cardiovascular Events (MACE). Future iterations will focus on human-AI hybridization, exploring how best to integrate AI recommendations with a surgeon’s clinical intuition.

Summary updated

BibTeX
@misc{fang2025compare,
  author = {Wei Fang and Chiyao Wang and Wenshuai Ma and Hui Liu and Jianqiang Hu and Xiaona Niu and Yi Chu and Mingming Zhang and Jingxiao Yang and Dongwei Zhang and Zelin Li and Pengyun Liu and Jiawei Zheng and Pengke Zhang and Chaoshi Qin and Wangang Guo and Bin Wang and Yugang Xue and Wei Zhang and Zikuan Wang and Rui Zhu and Yihui Cao and Quanmao Lu and Rui Meng and Yan Li},
  title = {{COMPARE: Clinical Optimization with Modular Planning and Assessment via RAG-Enhanced AI-OCT: Superior Decision Support for Percutaneous Coronary Intervention Compared to ChatGPT-5 and Junior Operators}},
  year = {2025},
  eprint = {2512.10702},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2512.10702},
}
Figure