多模态大语言模型辅助颅内动脉瘤血管内治疗决策的专家一致性评估
本研究回顾性纳入59例接受血管内治疗的未破裂颅内动脉瘤患者,由多学科神经血管团队、独立高年资介入神经放射医生、多模态大语言模型GPT-5及第一年神经介入专科培训医师分别基于相同临床资料与3D-DSA图像提出治疗策略。结果显示GPT-5与专家组治疗方式选择的中度一致性κ=0.64,优于培训医师κ=0.15。专家偏好评分中GPT-5方案中位评分4分,高于培训医师3分,差异p<0.001。
要点
- GPT-5在未破裂颅内动脉瘤血管内治疗方式选择上与专家组一致性κ=0.64,显著优于低年资培训医师κ=0.15。
- GPT-5与专家主要在策略规划层面存在分歧,在器械尺寸选择与着陆区评估等规则化亚任务上表现相似。
- 该模型当前仅作为辅助标准化决策与教学工具,需前瞻性多中心及完整DICOM流程验证后方可临床整合。
核心结论
- 多模态大语言模型GPT-5生成的颅内动脉瘤血管内治疗策略在真实临床决策场景中与专家启发式决策模式具有中等一致性,且显著优于培训医师水平。
- 模型并非自主临床推理,而是再现了神经介入实践中编码化的专家启发式规则,可作为标准化决策和培训支持的辅助工具。
研究设计与样本
- 回顾性单中心研究,纳入59例接受血管内治疗的未破裂颅内动脉瘤患者。
- 四位决策者分别评估每例病例:执行手术的多学科神经血管团队作为背景临床参照、独立高年资介入神经放射医生(EVT经验超过10年)、多模态大语言模型GPT-5、第一年神经介入专科培训医师。
- 所有评估者接受相同的临床摘要及6-10幅标准化3D-DSA投影图像,盲法专家对非参照方案采用4分等级量表评分。
关键影像表现/征象
- 所有决策者基于相同的6-10幅标准化3D-DSA投影图像及临床摘要进行评估。
- GPT-5与专家组的分歧主要出现在血管内治疗方式层面(SAC/BAC/FD),而在器械尺寸选择与着陆区估计等规则化亚任务上与专家表现相似。
诊断标准与关键数值
- GPT-5与专家组治疗方式选择的一致度Cohen κ=0.64,培训医师κ=0.15。
- 专家偏好评分显示GPT-5方案中位评分4分,培训医师中位评分3分,p<0.001。
- 4分等级量表反映手术可行性与启发式偏好的符合程度,而非临床正确性。
临床应用与阅片要点
- 多模态大语言模型可作为标准化治疗决策模式的辅助工具,尤其在低年资医师培训中提供参照。
- 在器械尺寸选择和着陆区估计等规则化任务上模型表现与专家接近,可用于报告或预案初稿中辅助操作参数选择。
- 阅片时需注意模型策略规划层面的分歧多集中于治疗方式选择层面,不能替代人工决策与监督。
局限性
- 回顾性单中心研究,样本量59例。
- 需要前瞻性多中心验证,并整合完整DICOM工作流程及人在环监督后方可临床整合。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
英文原文(摘要)· 点击展开
BACKGROUND: Endovascular treatment planning for intracranial aneurysms requires integrating vascular geometry, branch preservation, device feasibility, and patient-specific risk tolerance. Because multiple strategies may be clinically acceptable for the same anatomy, treatment selection reflects expert heuristics rather than a single deterministic solution. Whether multimodal large language models (LLMs) can reproduce such expert decision patterns from angiographic inputs remains incompletely characterized.
METHODS: In this retrospective single-center study, 59 patients with unruptured intracranial aneurysms treated with endovascular therapy were analyzed. Four decision agents reviewed each case: (i) the multidisciplinary neurovascular board that performed the index procedure (contextual clinical reference), (ii) an independent interventional neuroradiologist (> 10 years EVT experience), (iii) a multimodal LLM (GPT-5), and (iv) a first-year neurointerventional fellow included as a trainee-level comparator. All evaluators received identical clinical summaries and 6-10 standardized 3D-DSA projections. The blinded expert scored all non-reference proposals on a 4-point ordinal scale that reflects procedural feasibility and alignment with heuristic preferences rather than clinical correctness. Agreement with board decisions for modality-level strategies (SAC/BAC/FD) was quantified using Cohen's kappa (κ).
RESULTS: GPT-5 demonstrated moderate concordance with board-selected treatment modalities (κ = 0.64), within the range of expert-level variability reported in intracranial aneurysm management and higher than agreement observed for the trainee-level comparator (κ = 0.15). Expert preference-based scoring indicated that most GPT-5 proposals were judged as either preferred strategies or clinically acceptable alternatives (median score 4 vs 3 for the fellow; p < 0.001). Divergence between evaluators occurred primarily at the strategic planning level, whereas similar performance was observed for rule-based subtasks such as device sizing and landing-zone estimation.
CONCLUSIONS: In this retrospective case series, a multimodal LLM generated endovascular treatment strategies that frequently aligned with expert heuristic decision-making patterns when evaluated in a real-world clinical decision-making context. Rather than representing autonomous clinical reasoning, the model appears to reproduce codified expert heuristics embedded in neurointerventional practice. Multimodal LLMs may therefore serve as adjunct tools for standardizing decision patterns and supporting trainees. Prospective multicenter validation incorporating full DICOM workflows and human-in-the-loop oversight is required before clinical integration.
引用
- Demir, Yasar, Agackaya, et al. Evaluating expert decision-pattern alignment in endovascular planning for intracranial aneurysms using a multimodal large language model. Neuroradiology. 2026. PMID: 42347983. doi:10.1007/s00234-026-04084-8
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。