PHA · 论文互动解读Interactive paper walkthrough
← 学习中心
Nature · 2026 · Google Research / DeepMind · CC BY 4.0

一个“AI 健康管家”
是怎样炼成的？
The anatomy of a personal health agent

谷歌团队把手环数据、血检报告和大语言模型连在一起，造了一个由三位“AI 专家”组成的团队，并用 1,350 小时的人工评估去检验它。今天我们一起拆开看：它强在哪？风险在哪？和我们有什么关系？
A Google team connected wearable data, blood tests and a large language model into a team of three AI specialists, then tested it with 1,350 hours of human evaluation. Today we take it apart: what works, what is risky, and what it means for us.

← → 或滚动翻页 · 按 S 显示讲者提示 · 右上角切换语言Arrow keys or scroll · press S for speaker notes · switch language top right

开场 1 分钟：不要先讲技术。直接问“谁戴着手表或手环？”举手后再进入下一页投票。
先投票Quick poll

你愿意把自己的手环数据和体检报告，交给 AI 来解读和给建议吗？
Would you let an AI read your wearable data and lab results, and advise you?

讲者可按举手人数逐次点击计数；结果保存在本机浏览器，最后一页会再投一次做对比。The presenter can click once per raised hand; counts stay in this browser and we vote again at the end.

让大家记住自己的选择，最后一页再投一次，看看讲完“好处 + 风险”之后态度有没有变。
第一步：先听用户说什么Step 1: listen to users first

造系统之前，团队先看了大家到底在问什么
Before building anything, they studied what people actually ask

0
条真实健康提问（搜索、论坛等）
real consumer health queries (search, forums)
0
位用户的健康目标问卷
users surveyed on health goals
0
位设计和工程专家的工作坊
UX and engineering experts in a workshop

小测验：下面每个问题属于哪一类需求？Quiz: which need does each question belong to?

例句均来自论文正文（第 2、4 页）。All example questions are quoted from the paper (pp. 2, 4).
让学生喊答案。重点：一句话往往同时属于两类（比如“看懂我的睡眠数据，再告诉我怎么改善”= 数据解读 + 健康建议），这正是后面要用“团队”的原因。
为什么不用一个大模型搞定？Why not one big model?

三种能力，互相拉扯
Three skills that pull in different directions

DS

算得准Precise numbers

看个人数据要写代码、做统计，一个数都不能错。
Personal data needs code and statistics; every number must be right.
DE

说得对Correct facts

医学知识和症状问题，要有可靠依据。
Health knowledge and symptoms need solid evidence.
HC

聊得暖Warm coaching

改变习惯需要倾听、共情、多轮对话。
Changing habits needs listening, empathy and many turns of talk.
论文写道：他们的实验证实，单一的大模型架构很难同时兼顾这些目标，所以改成了“多智能体”分工。
The authors report that a monolithic LLM struggled to balance these objectives, so they split the work across agents.
类比 · 论文也提到受人类健康团队启发，具体比喻是我的Analogy · the paper cites human health teams; this framing is mine
就像医院的多学科会诊（MDT）：检验科看数据，专科医生给判断，健康管理师陪你落实，最后由主治医生汇总成一个方案。
Like a hospital multidisciplinary team: the lab reads the numbers, the specialist interprets, a health coach helps you follow through, and the attending pulls it into one plan.
可以问医学生：MDT 的问题是什么？（慢、贵、协调成本）——这正好为后面的“代价”埋伏笔。
核心结构：一个总调度 + 三位专家Core design: one orchestrator, three specialists

点一下，看一个问题如何在团队里流转
Click to watch one question move through the team

👤 “我想看懂我上周的睡眠数据，并知道怎样改善睡眠。”“I want to understand my sleep data last week and know how I can improve my sleep.”
① 理解需求1 Understand need ② 分派专家2 Assign agents ③ 反思检查3 Reflect ④ 更新记忆4 Update memory
👤 Orchestrator总调度 DS · 数据科学家分析手环睡眠数据 / analyses wearable data DE · 健康领域专家循证睡眠知识 / evidence-based sleep knowledge HC · 健康教练个性化建议 / personalises advice
✅ 总调度把三位专家的结果汇总、反思后，给出一个连贯的回答，并记住这次对话。The orchestrator synthesises and reflects on the three outputs, gives one coherent answer, and remembers the conversation.
例子与四个步骤来自论文第 3、5 页及图 5a；示意图由代码重绘。Example and four steps from pp. 3, 5 and Fig. 5a; diagram redrawn in code.
强调：这不是“三个机器人各说各的”。论文专门对比了“三位专家各自独立回答”的并行版本，后面第二页会看到结果差别。
三位专家各自考了什么试How each specialist was tested

翻开卡片：和“普通版” Gemini 相比
Flip the cards: compared with plain Gemini

DS

数据科学家Data science agent

把“我是不是更健康了？”这种模糊问题，先写成分析计划，再写代码算出来；出错会自己改。
Turns vague questions like “am I getting more fit?” into an analysis plan, then code; it fixes its own errors.

猜猜：普通模型能发现“数据不够/缺失”的比例是多少？Guess: how often did plain Gemini notice missing data?

↻ 点击翻面click to flip

灰 = 普通 Gemini，彩 = 专家智能体grey = base Gemini, colour = agent

DE

健康领域专家Domain expert agent

会查网络、查 NCBI 文献库，把你的手环数据、化验单和医学知识放在一起解读。
Uses web search and NCBI to read your wearable data, labs and medical knowledge together.

猜猜：用户觉得“值得信任”的比例差多少？Guess: how big was the gap in perceived trustworthiness?

↻ 点击翻面click to flip

三项来自三个不同测试；考试题基线已有 81.8%Three separate tests; exam baseline was already 81.8%

HC

健康教练Health coach agent

参考动机式访谈：先问目标、再问困难、听你反馈，最后才给具体建议。
Built on motivational interviewing: goals first, then barriers and feedback, then concrete advice.

猜猜：有没有哪一项普通版得分更高？Guess: did the baseline score higher on anything?

↻ 点击翻面click to flip
31 位用户：6 项能力里 5 项更受欢迎（胜率 61–71%）。唯一例外：进度追踪，普通版得分更高。
专家眼中的差距比用户更大，例如“赋能用户”：专家胜率 69.9%，用户只有 48.4%。
31 users preferred it on 5 of 6 themes (61–71% wins). The one exception: progress measurement, where the baseline scored higher.
Experts saw bigger gaps than users, e.g. empowerment: 69.9% vs 48.4%.
给医生 · 深入一点For clinicians · more detail
DS：173 个单元测试，自我纠错最多 5 次后通过率 79.0%；数据处理错误 25.4% → 11.0%；分析计划中“统计检验”一项无显著差异（基线本就不错）。
DE：2,000 例消费者自述症状的回顾性病例的鉴别诊断 top-1 准确率 46.1%，比此前最佳 DDx 智能体高 4.7 个百分点；医学选择题 83.6% vs 81.8%（P = 0.002）。
注意：可信度（17 位用户、50 个个性化问题）和考试题（2,000 多道）是两个不同测试，不能直接相除比较；在可信度测试中，两者的相关性和数据贴合度相近，医生也独立认为它的总结更有用、引用更可靠。
讲者观点（非论文结论）：用户的信任，或许部分来自引用和个性化的呈现方式，值得在课堂讨论。
DS: 173 unit tests; 79.0% pass after up to 5 self-corrections; data-handling errors 25.4% → 11.0%; no significant difference on the plan's statistical-tests dimension.
DE: differential diagnosis top-1 46.1% on 2,000 retrospective consumer-reported symptom cases (+4.7 points over prior SOTA DDx agent); MCQ 83.6% vs 81.8% (P = 0.002).
Note: trust (17 users, 50 personalised questions) and exam accuracy (>2,000 items) are separate tests and should not be compared directly; in the trust test relevance and data grounding were comparable, and clinicians independently rated its summaries more useful with more credible citations.
Presenter's view (not the paper's): some user trust may come from how citations and personalisation are presented — worth discussing.
数字来自论文第 4–5 页；条形图按原文百分比用代码绘制。Numbers from pp. 4–5; bars drawn in code from reported percentages.
DE 卡片是好的讨论点：考试题基线已高（81.8%→83.6%），而在另一个测试里用户可信度从 38.7% 到 96.9%。提醒这是两个不同测试。问学生：“可信”和“正确”应该怎样分别验证？
团队合体后，打赢了吗？Does the full team win?

三个版本同台：谁被排第一最多？
Three systems head to head: who was ranked first most?

对手：① 单一智能体 PHIA；② 三位专家各自独立回答的“并行版”；③ 有总调度的完整 PHA。45 位用户 + 3 位健康专家参与比较。
Contenders: (1) single-agent PHIA; (2) a parallel version where the three agents answer independently; (3) the full PHA with an orchestrator. 45 end users and 3 health experts compared them.

普通用户：回答质量排第一End users: ranked #1 on response quality

健康专家：回答质量排第一Experts: ranked #1 on response quality

相对基线，PHA 在用户评分上平均高 4.6%，在专家评分上平均高 17.1%。专家比普通人更能看出差别。
Versus baselines, PHA scored on average 4.6% higher with users and 17.1% higher with experts. Experts noticed differences users did not.
给医生 · 评估规模For clinicians · evaluation scale
全文共 11 项基准任务、超过 7,000 条人工标注、1,350 小时人力（用户 651 小时 + 专家 699 小时）。数据来自 WEAR-ME 研究（n = 1,165），以 Fitbit 设备为主。用户与专家的排名均为图 5b、5c 中的“回答质量”排名；3 位专家审阅的是对话记录，没有亲自与系统对话。
11 benchmark tasks, >7,000 annotations, 1,350 h (651 user + 699 expert). Data from the WEAR-ME study (n = 1,165), mainly Fitbit. Both user and expert bars are response-quality rankings (Figs. 5b, 5c); the 3 experts rated conversation transcripts rather than chatting themselves.
这页的“启发”是：谁来评价，结论就可能不同。患者满意 ≠ 医学上更好，反之亦然。
最值得停下来的一页The slide to pause on

医生审查了它写的健康总结：多少可能有害？
Clinicians reviewed its health summaries: how many could cause harm?

83% 无/极小风险 · none/minimal 15% 轻度 · mild

红色那一小段：2% 中度到重度风险。论文作者自己也在讨论部分强调了这一点。The red sliver: 2% moderate-to-severe potential harm. The authors flag this themselves in the Discussion.

思想实验 · 我的引申，按比例换算，不是论文结论Thought experiment · my extrapolation, not a finding
如果有 10,000 份这样的总结，按同样比例，约有 1,500 份轻度风险、200 份中重度风险。
If there were 10,000 such summaries, at the same rates about 1,500 would carry mild and 200 moderate-to-severe potential harm.

注意：原始审查只基于 30 位参与者的总结，样本小，比例不确定性很大。Caution: the review covered summaries for only 30 participants, so these rates are uncertain.

论文提出的安全措施Safeguards the paper proposes
定位是研究原型，不做诊断或分诊，不能替代专业医疗。未来要：识别高风险问题并拒答或明确转交医生；检索限定在 WHO、PubMed、NICE、CDC 等可靠来源并标出处；代码要系统校验；遵守 HIPAA、GDPR 等隐私规范；关注公平性（不同设备、语言、读写水平、人群）。
A research prototype, not for diagnosis or triage, and not a replacement for care. Future versions should detect high-risk queries and refuse or hand off to a clinician; restrict retrieval to curated sources (WHO, PubMed, NICE, CDC) with provenance; validate generated code; meet HIPAA/GDPR; and evaluate fairness across devices, languages, literacy and groups.
问医生学员：一个药物的严重不良反应率 2% 你会怎么评估？AI 的 2% 应该用同样标准吗？谁来负责？
代价与局限Costs and limits

比单一智能体更慢、更贵，但比并行版更省
Slower and costlier than a single agent, but leaner than the parallel team

单一智能体 PHIA：3.0 次模型调用Single-agent PHIA: 3.0 LLM calls
36.4 s
完整 PHA：6.5 次模型调用Full PHA: 6.5 LLM calls
205.3 s
并行版：7.0 次模型调用Parallel multi-agent: 7.0 LLM calls
244.4 s

平均每轮回答；在论文合成的对话上测得（Methods“Computational cost”）。Average per response round, measured on the paper's synthesized conversations (Methods, Computational cost).

论文自己承认的局限Limitations the authors state

• 只证明了短期对话质量，没有证明长期行为改变或健康结局改善
• 部分评分由 AI 打分员完成，可能受格式、长度影响
• 数据以 Fitbit 为主，换设备未必适用
• 缺少老年人、残障人士、不同语言和读写水平人群的验证
• Shows short-term conversation quality, not long-term behaviour change or health outcomes
• Partly scored by LLM autoraters, sensitive to format and length
• Fitbit-centric data; other devices untested
• Not yet tested in older adults, people with disabilities, other languages and literacy levels

好消息The good news

系统代码以 Apache 2.0 开源；作者验证了 DS、DE 两个智能体的架构换用其他大模型也有效。其他团队可以复现和挑错。
The system is open-sourced under Apache 2.0, and the authors showed the DS and DE architectures also work with other LLMs, so others can reproduce and critique it.
PHA 一次回答要 3 分多钟，但比三位专家各自回答的并行版（约 4 分钟）还快——总调度按需派活，省掉了不必要的调用。问：在门诊、在凌晨 3 点焦虑的患者面前，这个速度能接受吗？
启发Takeaways

这篇论文能教给我们什么？
What can we learn from it?

以下为讲者的引申与观点，不是论文结论The presenter's reflections, not the paper's conclusions
  1. 先研究人，再造工具。他们先分析了 1,370 条真实提问，才决定系统长什么样。做医疗 AI，起点是患者真正的问题。
    Study people before building tools. They analysed 1,370 real questions before designing anything. Medical AI should start from patients' real questions.
  2. 分工协作值得尝试。在这项研究里，有统筹的专家团队胜过单一智能体，但作者也说这个配置并非“普遍最优”。
    Division of labour is worth trying. Here a coordinated team beat a single agent, though the authors say their configuration is not “universally optimal”.
  3. “让人觉得可信”和“正确”要分开检验。这篇论文分别测了知识准确性、用户信任和医生评价，三者并不总是同步。评价医疗 AI，不能只看用户觉得它靠不靠谱。
    Test “feels trustworthy” and “is correct” separately. The paper measured knowledge accuracy, user trust and clinician judgement separately, and they do not always move together. Judge medical AI by more than how convincing it feels.
  4. 评价者决定结论。专家看到的优势远大于普通用户。医疗 AI 的评估既要有患者，也要有专业人员。
    Who evaluates shapes the verdict. Experts saw far bigger gains than users. Evaluate medical AI with both patients and professionals.
  5. AI 的位置：帮你准备就医，而不是替你看病。论文也建议未来系统帮用户记录症状、整理问题、找到合适的就医路径。
    AI's role: prepare you for care, not replace it. The paper also suggests future systems help users log symptoms, prepare questions and navigate care.
第 5 点的后半句来自论文讨论部分；前面是我的总结。
课堂讨论Discussion

留给大家三个问题
Three questions for you

1. 2% 的总结可能造成中重度伤害——这个比例可以接受吗？出了问题，由谁负责？
1. 2% of summaries carried moderate-to-severe potential harm. Is that acceptable? Who is accountable?
2. 如果 AI 健康管家让人“更放心”，会不会让一些人推迟去看医生？
2. If an AI health agent makes people feel reassured, could some delay seeing a doctor?
3. 在你的工作或生活里，哪一件事最适合交给这样的“AI 团队”？哪一件绝对不行？
3. In your work or life, what would you hand to such an AI team — and what never?

再投一次：现在你愿意吗？Vote again: would you now?

课前Before
课后After
来源与说明Source and notes

Heydari AA, Gu K, Srinivas V, Yu H, Paruchuri A, et al. The anatomy of a personal health agent. Nature (2026).

DOI: 10.1038/s41586-026-11161-8 · Open access, CC BY 4.0

• 所有数字均来自论文正文与图 5，逐条可追溯（见讲者的事实表）。
• 图表和示意图均由代码按原文数字重绘，未复制原图，未使用 AI 生成图片。
• 虚线框内的类比、思想实验和启发是讲者引申，不是论文结论。
• 代码：github.com/Google-Health/consumer-health-research（Apache 2.0）。
• 本课件仅用于教学讨论，不构成医疗建议。
• All numbers come from the main text and Fig. 5 and are traceable (see the presenter's fact sheet).
• Charts and diagrams are redrawn in code from reported numbers; no original figures copied; no AI-generated images.
• Dashed boxes (analogies, thought experiment, takeaways) are the presenter's own, not the paper's findings.
• Code: github.com/Google-Health/consumer-health-research (Apache 2.0).
• For teaching only; not medical advice.
← 返回学习中心 · Agent 工程