AutoMedatAGENT · 医学数据静默智能体

摘要 / ABSTRACT ABSTRACT 面对 2.16TB、两万余份检查的医院原始影像,AutoMedatAGENT 以静默智能体的方式完成发现、身份核对、格式转换、标准化、质检与特征提炼六道工序;每一步可审计、可回退,身份冲突等关键决策交由人工复核。为某慢性病大队列产出科研就绪数据与 2560 维影像特征,特征提取成功率 97.9%。 Facing 2.16 TB of raw hospital imaging across 20,000+ exams, AutoMedatAGENT runs a six-stage cleaning line — discovery, identity resolution, conversion, standardization, QC and feature extraction — as a silent resident agent. Every step is audited and reversible; critical decisions such as identity conflicts are escalated to human review. Built for a large chronic-condition cohort, it delivers research-ready data and 2,560-d imaging features at a 97.9% extraction success rate.

The problem: a drawer full of mixed socks

难题:一抽屉混在一起的袜子

Hospital scans (CT / MRI) arrive as thousands of folders copied from machines. Folder names can't be trusted, one patient has several IDs, and sometimes two patients are mixed in one folder. Doing this by hand for 20,000+ exams is soul-crushing.

医院的 CT/MRI 影像是从机器里一盘盘拷出来的成千上万个文件夹:文件夹名字不可信、一个病人有好几个编号、有时两个病人混在同一个文件夹里。两万多份检查靠人肉整理,足以压垮任何研究生。

What it actually does: a 6-stage cleaning line

它做的事:一条六道工序的清洗流水线

Drop new data into a folder and the line starts by itself. Every run is logged and reversible.

把新数据丢进一个文件夹,流水线自动开工。每一步都有记录、可追溯、可回退。

AutoMedatAGENT system architecture — **Fig. 1** 系统总览:编排层(策略守则 · LLM 路由 · 人工复核)监督六道工序,产出科研就绪数据。 **Fig. 1** System overview: the orchestration layer (policy, LLM routing, human review) supervises a six-stage pipeline that emits research-ready data.

What makes it an "agent", not just a script

为什么说它是「智能体」而不只是脚本

It works like a careful night-shift intern: it watches for new data, plans the steps, and knows exactly what it may decide alone — and when it must raise its hand and ask a human.

它像一名靠谱的值夜班实习生:盯着新数据、自己规划步骤,并且清楚哪些事可以自己拍板,哪些必须举手请示人类。

✓ Allowed to decide alone

✓ 可以自己拍板

Running approved steps · retries & timeouts · skipping duplicate files · writing reports

执行已批准的工序 · 重试与超时处理 · 跳过重复文件 · 写运行报告

✋ Must ask a human first

✋ 必须请示人类

Deleting data · overwriting results · identity conflicts between patients · the deep-learning phase

删除数据 · 覆盖正式结果 · 病人身份冲突 · 深度学习特征提取阶段

These boundaries aren't vibes — they're written down as a policy document (workflow_policy.json) the agent reads. Large language models (DeepSeek, GLM) help classify inputs and summarize reports, but the pipeline runs fine without any model at all.

这些边界不是口头约定,而是一份写好的《自主性守则》(workflow_policy.json),智能体自己会读。大语言模型(DeepSeek、智谱 GLM)帮忙识别输入类型、读 PDF 报告;但即使一个模型都不配,流水线照跑不误。

In numbers

用数字说话

2.16 TBof raw imaging tamed原始影像数据

21,451CT exams quality-checkedCT 质检条目

97.9%feature extraction success特征提取成功率

92/minexams processed (RTX 3090)每分钟处理(RTX 3090)

Built for a large chronic-condition cohort: images, clinical spreadsheets and PDF reports all flow through the same audited pipeline.

为某慢性病大队列而建:影像、临床表格、PDF 报告全部流经同一条可审计的流水线。