AI Trends

Mirobody Organizes Years of Health Data Locally on an Ordinary Laptop

Ayi AI NotesUpdated Nov 20 16:55

AI summary

The author ran the open-source health engine Mirobody on an ordinary laptop with 16GB of memory and connected years of lab reports and Apple Health data to Claude Code or Cursor. It can standardize reports from different hospitals offline and trace data back to the original documents, but processing blurry images using only a CPU may take several minutes. When connected to a cloud-based AI client, readings are still sent to the model provider. The tool only organizes health records; it cannot diagnose illnesses or prescribe medication.

Why read this

AI score 71

Drawing on hands-on experience running it locally, the author explains its health data standardization and source-document traceability capabilities, while also addressing CPU-only performance, privacy, and the boundaries of its medical use.

Topics

Tags

Quoted post

WY

@akokoi1

English text is being prepared.

View original

ChatGPT Health这类 AI 健康应用体验都做得很好,注册就能用。但很多人一直有个顾虑,体检报告是一个人最私密的数据,不想把它们全交给云端。 今天试了下在我这台5年前的Mac(M1 Pro)上,用 Theta Health 开源的健康数据引擎 Mirobody(Apache-2.0)搭了一套 100% 本地架构。应用、PostgreSQL 数据库和本机 llama.cpp 模型服务(配置 MiniCPM5-2B + GLM-OCR)全在本机,没有数据离开电脑。 分享下真实部署体感和实测结果: 1️⃣ 部署与踩坑 用 Docker 跑应用和数据库,主机直接用 llama.cpp 起 llama-server 调用 Metal GPU(无显卡的机器直接 COMPOSE_PROFILES=local-cpu ./deploy.sh 也能跑,16 GB 内存起步,只是纯 CPU 回答可能需要几分钟;Windows 需要在 WSL 2 里跑)。 实测遇到1个小细节:国内直连 Hugging Face 拉本地模型权重会 SSL 报错,启动 llama-server 前加个 HF_ENDPOINT=https://t.co/Yg4ljWINoo就能正常下载。 2️⃣ 数据标准化 看历年体检报告最头疼的,就是不同医院写法不同、单位也不一样,连不成线。 我先在终端用一条命令测了它的离线标准化引擎(不用模型也不用 Key): uvx --python 3.12 mirobody resolve "LDL cholesterol" 血红蛋白 ヘモグロビン "Total Kolesterol" 血脂 实测输出: - 血红蛋白 和日文 ヘモグロビン 同归 LOINC 718-7 - LDL cholesterol 归为 LOINC 13457-7 - 血脂 直接返回 unresolved——因为它是个大类不是具体检查,宁可留空也不靠大模型瞎猜。 这套规则和词表完全离线、确定性输出(LOINC 2.83 共 63,416 个编码、1,023,273 条名称,331 个 UCUM 单位,常规体检覆盖测试 347/347),同样输入在任何机器上结果都一样。 3️⃣ 本地网页端实测 为了保护隐私,我直接用了自带的虚拟演示账号: - 指标页里每一项数据都挂着标准 LOINC 编码(见图 3); - 切换到关爱圈里共享的 Mom (demo) 档案,点开总胆固醇(LOINC 14647-2):两份不同时期的报告单位不同,一份 198 mg/dL 被自动换算成 5.12 mmol/L,和 5.3 mmol/L 连成了同一条趋势线,下面既保留了报告原值,也能直接点开原文件核对(见图 1)。 - 开源版自带的「关爱圈」设计很实用:家人各自管自己的档案、决定给谁看,不自己登录的长辈也可以先由你代管一份。 4️⃣ 它和 ChatGPT Health / 蚂蚁阿福差在哪? 两者其实不冲突,定位互补(详细对比看图 4): - ChatGPT Health、蚂蚁阿福是开箱即用的云端健康助手; - Mirobody 是给想把数据掌控在自己手里的人用的开源底座,支持 PDF、照片、表格等 23 种文件,不限地区,模型、提示词和工具都能按自己需求换。 (如果暂时不想本地部署,国内网页版 https://t.co/CFokHfEXtL 和微信小程序「Mirobody」也基于同一套引擎;海外版叫 Theta:https://t.co/lkcPXDrn56) 如果需要自己部署,可以直接把这个GitHub链接丢给AI:https://t.co/qTnWiuhM0h

View original post

Body · AI translation

A friend of mine recently lost his father to pancreatic cancer before the age of 60. It was only a few months from diagnosis to his passing. Everyone really must pay attention to their health. Since you’ve gotten past the Great Firewall, you can look into some highly informed health practices from overseas. Take Bryan Johnson, who spends $2 million a year fighting aging: he is an absolute beast who has pushed his muscle mass and bone density into the top 2% worldwide and reversed his skin age by nine years. He personally eliminated a number of popular internet protocols, including SLU-PP-332, which disrupted his sleep; plasmalogens, which produced no change in his biomarkers; and methylene blue, which reduced his tolerance for hypoxic training. He candidly said that the vast majority of the health benefits still come from a few free, simple habits: going to bed at the same time every day, stopping eating four hours before bed, putting away his phone one hour before bed, and getting sunlight after waking up. He also does a 20-minute dry sauna at 79 to 100 degrees every day, which directly reduced the microplastics in his blood and semen by 85%. What truly sets this wealthy longevity enthusiast apart is that whenever he tries a new habit, he monitors his blood and sleep metrics for feedback and immediately abandons it if the data does not improve.

We ordinary people can recreate this kind of personal controlled experiment at home without wasting money. I also installed Mirobody—the top-ranked open-source health engine on GitHub shared by our buddy wy—on my own computer. It can run offline on an ordinary laptop with 16GB of memory and does not require a dedicated graphics card.

In just two minutes, you can connect your entire family’s historical lab reports and Apple Health data to Claude Code or Cursor.

It has 63,416 international medical standard codes and 1.02 million multilingual aliases built in. Drop in PDFs and photos of lab reports from different hospitals, and its local terminology database will automatically convert the units and generate continuous line charts in 0.4 seconds. Every number in its answers can be clicked to jump directly back to a photo of the original report from that year for verification.

Before you get started, however, let me lay out the pitfalls and limitations so you do not waste your time: The psychedelic mushrooms and metformin cycles mentioned in Bryan Johnson’s article are extreme personal experiments. Do not impulsively follow his example and take medications indiscriminately.

Mirobody only organizes health records and traces data back to the original documents. It absolutely does not diagnose illnesses or issue prescriptions; you still need to consult a doctor to determine whether your results are normal.

When it encounters an obscure hospital abbreviation or a broad umbrella term such as blood lipids, it will honestly leave the field blank and mark it as awaiting confirmation. You will need to confirm it manually with one click; it will not force a match.

When running a local model entirely on the CPU without a graphics card, it can take several minutes to read blurry photos taken with a phone. Before storing real reports belonging to you and your parents, be sure to disable the demo-data setting in the configuration. Once you connect it to a cloud-based AI client, any readings you query will still be sent to the model provider.

I’ve posted Bryan Johnson’s checklist of six free daily habits in the comments. Feel free to grab it.

Source

Ayi AI Notes