跳到正文

医疗报告结构与来源检查

整理病例、诊断、试验和研究报告的文档结构,使用本地检查工具核对来源事实与报告字段,结果供专业人员审阅。

开源精选医疗业务39 个文件50,135 Stars浏览次数 2K-Dense-AI/scientific-agent-skills MIT版本 a1043ef
开始使用

适用环境:Requires Python 3.11+ for optional dependency-free local scripts. No credentials or network needed for data processing; network access is needed only to review current public guidance.

输入资料
获准使用且符合原文范围的资料 经过核查的来源事实与报告类型 阅读本节完整说明
交付内容
结构化报告草稿 来源与文档检查记录 阅读本节完整说明
  1. 获取完整包,保留配套文件的目录结构。
  2. 按照原文准备所需工具、依赖与输入资料。
  3. 阅读对应使用步骤与输入输出说明,在自己的环境中验证结果。

使用提示摘自原文,配套文件保留来源版本;实际运行条件以各包说明为准。

正文

获取原文件

references/privacy_and_deidentification.md

This reference documents a review process. It is not legal advice, a technical de-identification service, or evidence of HIPAA compliance.

HHS framework

HHS guidance under 45 CFR 164.514(b) describes two methods:

  1. Expert Determination — a person with appropriate knowledge and experience applies generally accepted statistical and scientific principles and documents that re-identification risk is very small under the anticipated conditions.
  2. Safe Harbor — specified identifiers are removed and the covered entity has no actual knowledge that remaining information could identify an individual.

A local script cannot perform Expert Determination, establish “no actual knowledge,” or decide whether an organization is a covered entity or business associate.

Safe Harbor identifier categories

The responsible privacy professional must review the exact regulation and HHS guidance. The categories include:

  1. names;
  2. geographic subdivisions smaller than a state, subject to the specific ZIP-code rule;
  3. date elements more specific than year directly related to an individual, plus the age rule for persons over 89;
  4. telephone numbers;
  5. fax numbers;
  6. email addresses;
  7. Social Security numbers;
  8. medical record numbers;
  9. health-plan beneficiary numbers;
  10. account numbers;
  11. certificate or license numbers;
  12. vehicle identifiers and serial numbers;
  13. device identifiers and serial numbers;
  14. URLs;
  15. IP addresses;
  16. biometric identifiers;
  17. full-face photographs and comparable images;
  18. other unique identifying numbers, characteristics, or codes.

Removal of obvious patterns is insufficient. Initials, partial identifiers, metadata, free text, rare events, small cells, unusual dates, images, and combined quasi-identifiers may still identify a person.

Minimum necessary

HHS states that covered entities generally take reasonable steps to limit uses, disclosures, and requests for PHI to the minimum necessary for the purpose. HHS also lists exceptions, including certain treatment disclosures, disclosures to the individual, authorized uses/disclosures, uses/disclosures required for HIPAA administration, HHS enforcement, and uses/disclosures required by law.

Do not apply the phrase mechanically. The responsible privacy/legal reviewer determines scope, exceptions, authorization, waiver, limited-data-set rules, and any more protective law or policy.

Local process

  1. Define purpose, recipient, authority, jurisdiction, and data class.
  2. Exclude fields not needed for the purpose.
  3. Keep source records in the authorized system; use authorized opaque locators and integrity hashes in the draft workspace. A locator, value hash, or aggregate can still disclose information: hashing an identifier is not a Safe Harbor substitute, and predictable values can be guessed. Include the manifest itself in privacy review.
  4. Select Safe Harbor, Expert Determination, or a documented synthetic/aggregate-data rationale through the responsible reviewer.
  5. Review structured fields, free text, attachments, images, headers, filenames, metadata, and linked data.
  6. Review combinations and small-cell/rare-case risk.
  7. Record actual-knowledge review or Expert Determination documentation as applicable.
  8. Verify access controls, storage, transmission, retention, and deletion under organizational policy.
  9. Obtain privacy/legal/institutional approval for the intended disclosure.
  10. Re-review after every content, recipient, or purpose change.

What the checklist does

assets/deidentification_process_checklist.json and scripts/check_deidentification.py verify that required process fields are documented. They deliberately:

  • do not scan patient free text;
  • do not output detected identifiers;
  • do not label a document COMPLIANT, SAFE, or DEIDENTIFIED;
  • do not substitute for Expert Determination or legal review;
  • block missing process declarations and retain a review-required status even when the declarations are complete.

The strongest successful result is PROCESS_DOCUMENTED_REVIEW_REQUIRED.

Publication consent, research consent, HIPAA authorization, IRB/Privacy Board waiver, and permission to use an image are distinct. Do not infer one from another or generate a stock assertion.

Record only the status and local documentation reference verified by the responsible human. Never store a signed consent form or direct identifier in this skill’s assets, tests, or example manifests.

Incident handling

If real PHI is unexpectedly present:

  1. stop processing;
  2. do not echo, copy, transform, or upload it;
  3. preserve only the minimum operational information needed under policy;
  4. notify the authorized privacy/security contact through the institution’s process;
  5. do not independently determine breach status or notification duties.

CONTACT HAONAN

联系haonan

医疗 AI · 产品与独立开发

刘浩南

关注医疗 AI 从 demo 到真实落地,聚焦产品化、商业化与院内工作流效率提升。

12年医疗产研经验,大湾区头部医疗AI公司前产品总监,独立开发者

医疗 AI 产品化医疗 AI 商业化院内工作流效率
刘浩南的微信联系二维码
微信扫一扫,联系 haonan