ClinSeek-Bench
收藏资源简介:
ClinSeek-Bench 是一个用于评估临床推理能力的多模态基准数据集,源自论文《ClinSeekAgent: Automating Multimodal Evidence Seeking for Agentic Clinical Reasoning》。该数据集旨在评估两种配对设置下的模型性能:1) 精选输入设置,模型直接从源基准提供的证据包中回答问题;2) 自动化证据检索设置,模型必须使用ClinSeekAgent工具从原始临床数据中主动检索证据。数据集包含两个主要部分:纯文本电子健康记录(EHR)任务和多模态任务。纯文本部分包含1,800个示例,覆盖45个EHR子任务(包括风险预测和决策制定场景),涉及1,563名独特患者。多模态部分包含989个示例,结合了EHR数据和胸部X光(CXR)图像,其中497个示例源自EHRXQA的CXR问答任务,492个示例源自MedMod的ICU/CXR预测任务,涵盖六个任务组:CXR发现存在性、CXR发现枚举、CXR时序变化比较、24小时失代偿预测、住院死亡率预测和表型预测。由于数据集构建自多个受保护的临床数据源,为保护隐私和遵守许可协议,本版本仅发布重建完整运行时基准所需的元数据文件,不包含原始患者数据、SQLite数据库、胸部X光图像文件或放射报告文本。用户需要根据自身的数据使用协议,从官方来源获取原始临床数据,然后使用提供的元数据和ClinSeekAgent工具在本地重建完整的评估环境。
ClinSeek-Bench is a multimodal benchmark dataset for evaluating clinical reasoning capabilities, derived from the paper *ClinSeekAgent: Automating Multimodal Evidence Seeking for Agentic Clinical Reasoning*. This dataset is designed to evaluate model performance under two paired settings: 1) Curated Input Setting, where models directly answer questions using the evidence pack provided by the source benchmark; 2) Automated Evidence Retrieval Setting, where models must actively retrieve evidence from raw clinical data using the ClinSeekAgent tool. The dataset consists of two main components: a plain-text electronic health record (EHR) task suite and a multimodal task suite. The plain-text subset contains 1,800 examples spanning 45 EHR subtasks including risk prediction and decision-making scenarios, involving 1,563 unique patients. The multimodal subset includes 989 examples that combine EHR data and chest X-ray (CXR) images: 497 examples are sourced from the CXR question answering task of EHRXQA, and 492 examples are from the ICU/CXR prediction task of MedMod, covering six task categories: CXR findings presence detection, CXR findings enumeration, CXR temporal change comparison, 24-hour decompensation prediction, in-hospital mortality prediction, and phenotype prediction. Since the dataset is constructed from multiple protected clinical data sources, to protect patient privacy and comply with licensing agreements, this released version only provides metadata files required to reconstruct the complete runtime benchmark, and does not include raw patient data, SQLite databases, chest X-ray image files, or radiology report texts. Users shall obtain raw clinical data from official sources in accordance with their own data usage agreements, then locally reconstruct the full evaluation environment using the provided metadata and the ClinSeekAgent tool.




