MYCIN 医疗专家系统

1970 年代中期，斯坦福大学的研究者面对一个现实问题：血液感染和脑膜炎的诊断复杂，而感染病专家稀缺。Edward Shortliffe 的博士论文项目 MYCIN，想做的事很直接——把专家的诊断推理编码成程序，让计算机给出治疗建议。和当时很多实验程序不同，MYCIN 面向的是一个真实得不能再真实的场景：病人该用什么抗生素。

MYCIN 的核心是一堆「如果…那么…」的规则。如果感染是脑膜炎，且病人是儿童，且培养结果是某类细菌，那么就考虑某种治疗方案。每条规则还带着一个置信度数字——因为医学判断从来不是非黑即白。程序诊断时顺着规则推理，每走一步都记下「为什么」：这个结论来自哪几条规则，依据了什么检查结果。系统因此能回答用户的追问：你为什么怀疑这种细菌？

这套机制在今天看起来朴素，在当时却是了不起的工程。MYCIN 最常被引用的评估，是它与斯坦福感染病专家的对比：给出同一批病例，专家的治疗建议与 MYCIN 的推荐被另一组专家盲评，结果 MYCIN 的建议接受度与人类专家相当。这个结果在 1970 年代引发了巨大兴奋——计算机看病，似乎真的可行。

但 MYCIN 从没有被正式投入使用。原因不在推理能力，而在系统之外：医疗责任怎么算？程序出错谁来负责？知识库谁来持续更新？医院是否愿意为维护买单？这些问题没有答案，MYCIN 最终停留在研究层面。这个结局本身成了 AI 史上最常被引用的教训之一：技术可行与落地部署之间，隔着工程、制度与信任的三道门。

MYCIN 的遗产更多在方法层面。它定义了规则库、推理引擎、置信度传播、解释机制的经典架构，也把「知识工程」变成了 1980 年代 AI 的主流词汇。后来的专家系统商品——比如 XCON——在企业里获得实际应用，走的正是 MYCIN 趟出来的路线。而当 2010 年代深度学习重新入场时，医疗 AI 又绕回到 MYCIN 当年撞过的问题：准确率重要，责任与信任同样重要。

今天，大模型可以在几秒钟内给出诊断建议式的回答，看起来比 1976 年的规则系统聪明得多。但 MYCIN 留下的那个问题——程序给出建议时，谁来为它负责——并没有随着技术升级而消失。它像一枚提前几十年立好的界碑，提醒每一个做医疗 AI 的人：推理做得多好，只回答了一半问题。

In the mid-1970s researchers at Stanford faced a real problem: blood infections and meningitis are hard to diagnose, and infectious-disease specialists are scarce. Edward Shortliffe's doctoral project, MYCIN, set a direct goal—encode expert diagnostic reasoning in a program that could recommend treatment. Unlike many experimental programs of the time, MYCIN confronted a genuinely real scenario: which antibiotic should the patient receive?

MYCIN's core was a pile of if-then rules. If the infection is meningitis, and the patient is a child, and the culture indicates a certain bacterium, then consider a certain therapy. Each rule carried a confidence number, because medical judgment is never black and white. As the program reasoned down its rules, it recorded the "why" at every step—which rules produced this conclusion, what test results supported it. The system could therefore answer follow-ups: why do you suspect this organism?

The mechanism looks plain today, but the engineering was substantial for its time. MYCIN's most cited evaluation compared it with Stanford infectious-disease experts: given the same cases, another panel blindly rated MYCIN's therapy recommendations against the specialists', and MYCIN's were judged acceptable about as often as the humans'. In the 1970s that result stirred enormous excitement—a computer that appeared to practice medicine.

MYCIN was never put into actual clinical use, and the reasons lay outside its reasoning ability. Who bears medical liability? Who is responsible when the program errs? Who keeps the knowledge base current, and will a hospital pay to maintain it? With no answers, MYCIN stayed a research project. Its ending became one of the most-cited lessons in AI history: between technical feasibility and real deployment sit three gates—engineering, institutions, and trust.

MYCIN's legacy is mostly methodological. It defined the classic architecture of rule bases, inference engines, certainty propagation, and explanation, and made "knowledge engineering" a buzzword of 1980s AI. Later commercial expert systems—such as DEC's XCON—found real enterprise use by walking the road MYCIN had paved. When deep learning re-entered medicine in the 2010s, it ran back into the same wall MYCIN hit in 1976: accuracy matters, but so does responsibility and trust.

Today a large model can produce a diagnosis-flavored answer in seconds, looking far cleverer than a 1976 rules engine. But the question MYCIN raised—who answers for a program's advice—did not disappear as technology improved. It stands like a boundary stone planted decades early, reminding everyone building medical AI that how well the reasoning performs only answers half the problem.