OpenMed 临床领域标签目录Clinical Domain Label Catalog完整解读零样本 NER 标签体系、风险分级与离线覆盖率门禁【免费下载链接】openmedLocal-first healthcare AI: clinical NER HIPAA PII de-identification that runs 100% on-device. 2,200 medical models, 21 languages, Apple MLX Python, no cloud, no patient data leaving your network. Apache-2.0项目地址: https://gitcode.com/GitHub_Trending/ope/openmed导读本文基于 OpenMed 仓库中的 docs/clinical-domains.md 临床领域标签目录系统讲解 OpenMed 零样本 NER / 临床信息抽取的领域标签体系——从 40 领域分组的显示标签display label、规范标签canonical label、政策类别与风险等级、编码系统提示system hints到 fixture 驱动的离线覆盖率评估门禁与benchmark domain-coverage命令的完整用法。读完本文你将能够理解 OpenMed 各类临床模型输出的标签如何被归一化到统一规范词汇表知道哪些领域有配套合成 fixture 可供离线评测并能独立运行覆盖率门禁验证标签映射的完整性。注意本文目录仅用于离线评估与评审工作流规划不构成临床指导不得用于推断诊断、治疗、编码或报销决策。目录体系概览从显示标签到规范标签的归一化OpenMed 的临床领域标签目录是一份从零样本领域标签映射zero-shot domain label map与规范标签元数据canonical label metadata生成的目录。整份目录由若干领域分组domain组成每个领域分组下列出若干标签Label每个标签都绑定 5 列结构化元数据列含义取值示例Label显示标签即零样本标签映射中实际出现的标签名Disease、VaccineName、ZygosityCanonical Label归一化后的规范标签UPPER_SNAKE_CASEDISEASE、VACCINE_NAME、ZYGOSITYCategory政策类别policy classCLINICAL_CONCEPT、DIRECT_IDENTIFIER、QUASI_IDENTIFIERRisk Level风险等级low、medium、highSystem Hints建议的编码系统terminology hintsICD-10-CM、SNOMED、HPO、RxNorm、LOINCFixture Path配套合成 fixture 文件相对仓库根目录tests/fixtures/clinical/immunization.jsonl或Not shipped这份目录的数据源头有两个其一是零样本领域标签映射 openmed/zero_shot/data/label_maps/defaults.json每个领域分组的显示标签列表与之完全一一对应其二是规范标签元数据 openmed/core/labels.py提供 canonical label、policy class、risk level、system hints 的定义。为什么需要规范标签Canonical Label归一化从 openmed/core/labels.py 的模块注释可以看到OpenMed 的不同 PII / 生物医学 NER 模型家族使用不同的标签命名约定英语 / 多语种 SuperClinical 模型使用小写snake_case如first_name、date_of_birth葡萄牙语模型暴露 52 个大写标签如FIRSTNAME、DATEOFBIRTHprivacy-filter 模型家族输出 BIOES 标签如B-NAME、I-EMAIL、E-ADDRESS、S-PHONE生物医学 NER 家族输出大写与 snake_case 混合的概念如DISEASE、simple_chemical、gene_or_gene_product。因此labels.py定义了单一的CANONICAL_LABELS规范词汇表统一为UPPER_SNAKE_CASE以及normalize_label帮助函数把上述所有输入形式映射到规范名称。下游代码匿名化、映射表、配置只应基于规范标签工作——这正是本文目录中 Canonical Label 一列存在的意义它让目录中的显示标签、模型原始输出与下游策略处理共享同一套词汇。领域分组全表40 临床与非临床领域目录按领域分组呈现下面按文档顺序完整列出各领域及其标签元数据Category、Risk、System Hints、Fixture 均以文档为准。基础生物医学与临床领域Biomedical / ClinicalBiomedical对应 defaults.json 的biomedical键Disease → DISEASElowICD-10-CM/SNOMED/HPO、Drug → DRUGlowRxNorm/SNOMED、Gene → GENElowSNOMED/HPO、Organism → ORGANISMlowSNOMED全部Not shipped。ClinicalProblem → PROBLEMmediumICD-10-CM/SNOMED/HPO、Treatment → OTHERlow、Test → LAB_TESTlowLOINC/SNOMED、BodyPart → BODY_SITElowSNOMED全部Not shipped。注意Problem在目录中被标记为medium风险。从 openmed/core/labels.py 的注释可以看到设计意图自由文本问题/适应证文本与结果值保持 medium 风险是为了让罕见疾病与独特测量值对风险工具可见而它们的政策类别仍为CLINICAL_CONCEPT因此不会进入默认的 PII 脱敏范围。基因组学领域Genomic / Genomic VariantGenomicVariant → VARIANT_DESCRIPTOR、Gene → GENE、Transcript → OTHER、Phenotype → OTHER全部 low 风险、Not shipped。Genomic Variant对应genomic_variant键首个带 fixture 的领域GeneSymbol → GENE_SYMBOL、VariantDescriptor → VARIANT_DESCRIPTOR、ProteinChange → PROTEIN_CHANGE、Zygosity → ZYGOSITY、AlleleFrequency → OTHER、ClinicalSignificance → CLINICAL_SIGNIFICANCE、ReferenceTranscript → OTHER全部 low 风险fixture 均为 tests/fixtures/clinical/genomic_variant.jsonl。这些规范标签在 openmed/core/labels.py 中定义为 issue #906 的临床基因组变异提及概念。非临床通用领域Finance / Legal / News / Ecommerce / Cybersecurity / Chemistry / Organism / Education / Social这部分领域多为跨域抽取场景如财务文本、新闻、电商、网络安全通告其标签大多归类为CLINICAL_CONCEPT/OTHER且低风险属于标签归一化的兜底处理但其中两个领域具有真正的标识符语义NewsPerson → PERSONhighDIRECT_IDENTIFIER、Organization → ORGANIZATIONmediumQUASI_IDENTIFIER、Location → LOCATIONmediumQUASI_IDENTIFIER、Date → DATEmediumQUASI_IDENTIFIER全部Not shippedSocialPerson → PERSONhighDIRECT_IDENTIFIER、Handle → OTHER、Hashtag → OTHER、URL → URLhighDIRECT_IDENTIFIER全部Not shippedGeneric与 News 相同的Person/Organization/Location/Date四件套风险分级一致。其余如FinanceCompany → ORGANIZATION 为 medium/QUASI_IDENTIFIER、Legal、Ecommerce、Cybersecurity、Chemistry其中Unit → UNITLOINC/SNOMED、OrganismSpecies → SPECIES、Education均为Not shipped且大多为 low/OTHER。专科临床领域Public Health / Disease / Pharmaceutical / Oncology / Anatomy / Genomics / Chemical / Species / Protein / Pathology / Hematology / Cardiology / Microbiology / Dermatology / Ophthalmology / RadiologyPublic HealthCondition → CONDITIONICD-10-CM/ICD-10-CN/SNOMED注意这里额外包含中国 ICD-10-CN 提示、Intervention → PROCEDURE、Outcome → OTHER、Population → OTHER全部Not shippedDiseaseDisease → DISEASE、Condition → CONDITION、Pathology → PATHOLOGY全部Not shippedPharmaceuticalChemical → CHEMICAL、Drug → DRUG、Medication → MEDICATIONRxNorm/CN-DRUG/SNOMED含中国药品编码 CN-DRUG全部Not shippedOncology18 个标签Oncotree 风格如SimpleChemical → CHEMICAL、Cancer → CANCER、GeneOrGeneProduct → GENE_OR_GENE_PRODUCT、Tissue → TISSUE、PathologicalFormation → PATHOLOGY等全部 low 风险、Not shippedAnatomyOrgan → ORGAN、Tissue → TISSUE、Anatomy → ANATOMYGenomicsGeneOrGeneProduct → GENE_OR_GENE_PRODUCT、Gene → GENE、Protein → PROTEIN、DNA → DNA、RNA → RNA、CellLine → CELL、CellType → CELLChemicalSimpleChemical/Chemical → CHEMICAL、Drug → DRUG、Medication → MEDICATIONSpeciesOrganism → ORGANISM、Species → SPECIESProteinGeneOrGeneProduct → GENE_OR_GENE_PRODUCT、Protein/ProteinComplex/ProteinEnumeration/ProteinFamilyOrGroup/ProteinVariant → PROTEINPathologyDisease/Condition/Pathology三件套HematologyCancer → CANCER、Disease → DISEASE、CellLine → CELLCardiologyCardiacFinding/ECGFinding/EjectionFraction/CardiacProcedure/CardiacDevice → OTHER、Anatomy → ANATOMYMicrobiologyMicroorganism → MICROORGANISMSNOMED/LOINC、Antibiotic → ANTIBIOTICRxNorm/SNOMED、Susceptibility → SUSCEPTIBILITYLOINC/SNOMED、SpecimenSource/CultureResult → OTHERDermatologySkinLesion/Morphology/Distribution → OTHER、Anatomy → ANATOMYOphthalmologyEyeFinding/VisualAcuity/IntraocularPressure → OTHER、Anatomy → ANATOMYRadiology首个专科带 fixture 领域Finding → FINDING、ImagingModality → IMAGING_MODALITY、Anatomy → ANATOMY、Laterality → LATERALITY、Measurement → MEASUREMENTLOINC/SNOMED、Impression → FINDINGfixture 为 tests/fixtures/clinical/radiology_finding.jsonl。Radiology 概念在 openmed/core/labels.py 中标注为 issue #1971 引入。FHIR 对齐的护理记录领域Anesthesia / Nutrition Diet / Endocrinology / Gastroenterology / Nephrology Renal / Immunization / Allergy Intolerance / Pulmonology / Pediatrics Growth / Procedures / Nursing Observation这一组领域是目录中最重的部分——它们大多拥有已提交的合成 span fixture且部分领域带有 FHIR 资源映射对齐说明Anesthesiaissue #952AnesthesiaType → ANESTHESIA_TYPE、AnestheticAgent → ANESTHETIC_AGENTRxNorm/SNOMED、AirwayManagement → AIRWAY_MANAGEMENT、ASAClass → ASA_CLASS、MonitoringModality/IntraoperativeEvent → OTHERfixture 为 tests/fixtures/clinical/anesthesia.jsonlNutrition Dietissue #951DietType → DIET_TYPE、NutritionTarget → NUTRITION_TARGET、Supplement → OTHER、FeedingRoute → FEEDING_ROUTE、IntakeFinding → OTHER、NutritionalStatus → NUTRITIONAL_STATUS、FluidRestriction → OTHERfixture 为 tests/fixtures/clinical/nutrition_diet.jsonlEndocrinologyissue #895GlycemicMeasure → GLYCEMIC_MEASURE、ThyroidFunctionMeasure → THYROID_MEASURE、HormoneLevel → HORMONE_LEVEL、InsulinRegimen → INSULIN_REGIMEN、MetabolicFinding → CONDITIONICD-10-CM/ICD-10-CN/SNOMED、EndocrineGland → BODY_SITEfixture 为 tests/fixtures/clinical/endocrinology.jsonlGastroenterologyissue #894EndoscopicFinding → ENDOSCOPIC_FINDING、GISymptom → GI_SYMPTOM、BowelPrepQuality → GI_SCORE、BiopsySite → BODY_SITE、GIScore → GI_SCORE、LesionMorphology/PolypDescriptor → POLYP_DESCRIPTORfixture 为 tests/fixtures/clinical/gastroenterology.jsonlNephrology Renalissue #892RenalFunctionMeasure → RENAL_FUNCTION_MEASURE、CKDStage → CKD_STAGE、DialysisModality → DIALYSIS_MODALITY、UrineFinding → URINE_FINDING、FluidStatus/RenalReplacementAccess/ElectrolyteDisturbance → OTHERfixture 为 tests/fixtures/clinical/nephrology_renal.jsonlImmunization见下节 FHIR 对齐说明fixture 为 tests/fixtures/clinical/immunization.jsonlAllergy Intolerance见下节 FHIR 对齐说明fixture 为 tests/fixtures/clinical/allergy_intolerance.jsonlPulmonologyissue #893SpirometryMeasure → SPIROMETRY_MEASURE、OxygenSupport → OXYGEN_SUPPORT、RespiratoryFinding → RESPIRATORY_FINDING、DyspneaGrade → DYSPNEA_GRADE、LungAuscultation → RESPIRATORY_FINDING、PFTInterpretation → OTHER、AirwayDevice → AIRWAY_MANAGEMENTfixture 为 tests/fixtures/clinical/pulmonology.jsonlPediatrics Growthissue #896GrowthParameter → GROWTH_PARAMETERSNOMED/LOINC、GrowthPercentile/GrowthZScore → GROWTH_PERCENTILE、DevelopmentalMilestone → DEVELOPMENTAL_MILESTONE、FeedingHistory → NUTRITIONAL_STATUS、PediatricFinding → CONDITIONfixture 为 tests/fixtures/clinical/pediatrics_growth.jsonlProceduresProcedure/Surgery/DiagnosticProcedure → PROCEDURE、Device → DEVICE、Approach → OTHER全部Not shippedNursing Observationissue #910IntakeOutput → INTAKE_OUTPUTLOINC/SNOMED、LineDrainTube → LINE_DRAIN_TUBE、RiskScore → NURSING_RISK_SCORE、MobilityStatus/PainScore → OTHER、CareIntervention → CARE_INTERVENTION、SkinAssessment → BODY_SITEfixture 为 tests/fixtures/clinical/nursing_observation.jsonl。FHIR 资源对齐说明Immunization 与 Allergy Intolerance目录中两个领域带有显式的 FHIR 对齐段落属于提取元数据层面的设计需要特别注意其边界声明Immunization 领域显示标签是为规划的 OM-138 FHIR Immunization 导出器塑形的——VaccineName映射到vaccineCode、DoseNumber映射到protocolApplied.doseNumber[x]、AdministrationRoute映射到route、AdministrationSite映射到site、VaccineLot映射到lotNumber、AdministrationDate映射到occurrence[x]、VaccineSeries映射到protocolApplied.series。文档明确声明这只是提取元数据不创建导出器、推荐、剂量或排程逻辑。对应规范标签定义在 openmed/core/labels.pyissue #897。Allergy Intolerance 领域显示标签为规划的 OM-138 FHIR AllergyIntolerance 导出器塑形——Allergen映射到code、ReactionManifestation映射到reaction.manifestation、ReactionSeverity映射到reaction.severity、Criticality映射到criticality、AllergyType映射到type、OnsetContext映射到onset。文档明确声明这只是描述性提取元数据不是禁忌症检查、药物-过敏交叉核对、推荐或临床决策逻辑。对应规范标签定义在 openmed/core/labels.pyissue #865。编码系统提示System Hints的语义System Hints 列给出每个标签建议归属的编码系统用于规划术语落地grounding而非实际绑定术语数据。其取值范围在 openmed/core/labels.py 中定义为RxNorm、LOINC、ICD-10-CM、ICD-10-CN中国 ICD-10、CN-DRUG中国药品编码、HPO人类表型本体、SNOMED。可以观察到以下规律疾病/表型类Disease、Cancer、Pathology、Problem→ICD-10-CM, SNOMED, HPO药品/化学品类Drug、Medication、AnestheticAgent、VaccineName、Allergen→RxNorm, SNOMED涉及中国场景时追加CN-DRUG检验/测量类Test、Measurement、Susceptibility、GrowthParameter→LOINC, SNOMED解剖/结构类BodyPart、Anatomy、Organ、Tissue、BiopsySite→ 仅SNOMED目录中大量OTHER标签的系统提示为SNOMED, ICD-10-CM, HPO, RxNorm, LOINC五系统全集表明这是未细分时的兜底提示。风险分级与政策类别脱敏策略如何参考本目录每个标签的Category与Risk Level直接来自 openmed/core/labels.py 中的政策元数据政策类别POLICY_LABELS共四类DIRECT_IDENTIFIER直接标识符、QUASI_IDENTIFIER准标识符、SENSITIVE_ATTRIBUTE敏感属性、CLINICAL_CONCEPT临床概念风险等级RISK_LEVELS三档low、medium、high_label_metadata还根据政策类别推断标签种类kind政策类别为CLINICAL_CONCEPT时记为BIOMEDICAL否则记为PII。本目录中的临床领域标签几乎全部为CLINICAL_CONCEPT / low——它们不会进入默认 PII 脱敏而 News/Social/Generic 领域的Person → PERSONhigh / DIRECT_IDENTIFIER、Organizationmedium / QUASI_IDENTIFIER、Locationmedium、Datemedium、URLhigh则是真正的隐私风险标签。这与 openmed/core/labels.py 中LABEL_METADATA的逐标签定义一致如PERSON: DIRECT_IDENTIFIER, highDATE: QUASI_IDENTIFIER, mediumLOCATION: QUASI_IDENTIFIER, medium。从源码结构看这份分级还可支撑多个下游机制例如 NDPA 敏感数据类别映射openmed/core/labels.py 中NDPA_SENSITIVE_CLASS_LABELS将GENE_SYMBOL、VARIANT_DESCRIPTOR、CKD_STAGE、GLYCEMIC_MEASURE等临床规范标签归入NDPA_HEALTH_STATUS健康状态敏感类同时目录风险等级也可作为评审工作流中风险评估的输入见 openmed/core/review_workflow.py。Fixture 支撑矩阵哪些领域可离线评估目录的Fixture Path列是关键实战信息所有标记为Not shipped的领域仅有标签映射而无合成语料只有下列领域带有已提交的合成 span fixture均为仓库内真实存在的文件领域Fixture 文件相对仓库根目录Genomic Varianttests/fixtures/clinical/genomic_variant.jsonlRadiologytests/fixtures/clinical/radiology_finding.jsonlAnesthesiatests/fixtures/clinical/anesthesia.jsonlNutrition Diettests/fixtures/clinical/nutrition_diet.jsonlEndocrinologytests/fixtures/clinical/endocrinology.jsonlGastroenterologytests/fixtures/clinical/gastroenterology.jsonlNephrology Renaltests/fixtures/clinical/nephrology_renal.jsonlImmunizationtests/fixtures/clinical/immunization.jsonlAllergy Intolerancetests/fixtures/clinical/allergy_intolerance.jsonlPulmonologytests/fixtures/clinical/pulmonology.jsonlPediatrics Growthtests/fixtures/clinical/pediatrics_growth.jsonlNursing Observationtests/fixtures/clinical/nursing_observation.jsonl这一矩阵与覆盖率门禁实现 openmed/eval/domain_coverage.py 中的CLINICAL_DOMAIN_FIXTURE_NAMES映射完全一致anesthesia、endocrinology、gastroenterology、genomic_variant、immunization、nephrology_renal、nursing_observation、nutrition_diet、pediatrics_growth、pulmonology、radiology共 11 个领域其注释明确写道其余打包的标签映射是 legacy 或在 docs/clinical-domains.md 中被显式标记为 Not shipped不属于本临床 fixture 门禁范围。Fixture 文件结构示例以 tests/fixtures/clinical/immunization.jsonl 第一条记录为例可以看到合成语料的组织方式{text: Tdap vaccine, dose 1 of 2, was administered intramuscularly in the left deltoid on 2026-03-14 from lot A123B as series 1 of 2. ..., metadata: {synthetic: true, disclaimer: ...}, entities: [{label: VaccineName, start: 0, end: 12, text: Tdap vaccine}, {label: DoseNumber, start: 14, end: 25, text: dose 1 of 2}, {label: AdministrationRoute, start: 44, end: 59, text: intramuscularly}, {label: AdministrationSite, start: 67, end: 79, text: left deltoid}, {label: AdministrationDate, start: 83, end: 93, text: 2026-03-14}, {label: VaccineLot, start: 103, end: 108, text: A123B}, {label: VaccineSeries, start: 112, end: 125, text: series 1 of 2}]}每条记录包含text合成文本、metadata含synthetic: true标记与免责声明和entitieslabel start/end 偏移 表面文本。tests/fixtures/clinical/radiology_finding.jsonl 采用结构化 gold 形式finding、laterality、size_value、location等字段 provenance_spans偏移覆盖Finding/Laterality/Measurement/Anatomy等标签。所有 fixture 均为合成数据不包含真实患者数据可安全用于离线评测。离线覆盖率评估clinical_domain_coverage 门禁门禁定位与运行方式文档给出的标准验证命令为在仓库根目录执行openmed benchmark domain-coverage --json --output domain-coverage.json该命令由 openmed/cli/main.py 注册为benchmark子命令domain-coverage其 help 文本为 Gate shipped clinical label maps against synthetic span fixtures.用合成 span fixture 门禁已发布的临床标签映射。门禁对应的评估套件名为clinical_domain_coverage实现在 openmed/eval/domain_coverage.py 中其职责是验证每个已发布的显示标签在合成 fixture 中有非空 span且 fixture 标签都能解析到规范标签目录。命令行参数domain-coverage子命令支持以下参数源自 openmed/cli/main.py参数类型说明--domain DOMAIN可重复append只检查指定领域可重复传入以选择离线覆盖范围--label-map PATHPath覆盖默认零样本标签映射本地 JSON label map--fixture-dir PATHPath覆盖默认合成临床 JSONL fixture 目录--output PATHPath输出机器可读的覆盖率摘要文件--markdown-output PATHPath输出聚合 Markdown 覆盖率报告--jsonflag以 JSON 格式输出文档示例命令使用默认的标签映射路径与 fixture 目录在 openmed/eval/domain_coverage.py 中定义标签映射默认指向openmed/zero_shot/data/label_maps/defaults.jsonfixture 目录默认指向tests/fixtures/clinical。失败条件与证据边界门禁在以下任一情况判定失败缺失 fixturemissing fixture某领域应有 fixture 但找不到文件孤儿标签orphan label标签映射中存在显示标签但无法解析到规范标签目录CANONICAL_LABELS无效偏移invalid offsetfixture 中实体偏移越界或格式非法无 span 标签label with no span某显示标签在 fixture 中没有任何非空 span。门禁的证据边界是其关键设计报告只包含领域名、标签、偏移与计数fixture 文本永不输出openmed/eval/domain_coverage.py 的模块注释明确说明这是泄漏优先leakage-first设计——避免把合成语料内容复制进报告或失败消息。数据结构上CoverageOffset 只暴露line/start/end/length而不暴露表面文本LabelCoverage 聚合 label、canonical_label、span_count、fixture_count 与 offsetsCoverageIssue 记录 domain、label、line、reason 等文本无关的失败证据。实战验证标签映射完整性的完整流程结合上文一套可复现的离线验证流程如下全量运行门禁仓库根目录openmed benchmark domain-coverage --json --output domain-coverage.json定向检查单个领域如免疫接种并生成可读报告openmed benchmark domain-coverage --domain immunization --markdown-output immunization-coverage.md使用自定义标签映射与 fixture 目录例如本地新增领域评测openmed benchmark domain-coverage \ --label-map ./my-label-map.json \ --fixture-dir ./my-fixtures \ --json --output my-coverage.json解读报告确认每个CLINICAL_DOMAIN_FIXTURE_NAMES中的领域都有非空 span 计数span_count 0且每个显示标签都解析到目录中声明的 canonical label若出现孤儿标签或缺失 fixture则对应领域需要补充或修正标签映射 / 合成语料。使用边界与声明本目录的使用边界在文档开头的免责声明与各领域注释中反复强调必须严格遵守目录仅用于离线评估与评审工作流规划不构成临床指导不得用于推断诊断、治疗、编码或报销决策Immunization / Allergy Intolerance 的 FHIR 对齐仅描述提取元数据的字段映射意图不创建导出器、不提供推荐/剂量/排程逻辑、不是禁忌症检查或药物-过敏交叉核对Fixture 均为合成数据报告中不输出任何 fixture 文本目录从打包的零样本领域标签映射与规范标签元数据生成若改动 defaults.json 或 labels.py 中的规范标签定义应重新运行domain-coverage门禁验证一致性。延伸阅读本目录所依赖的规范标签词汇表与归一化逻辑见 openmed/core/labels.py零样本领域标签映射源文件见 openmed/zero_shot/data/label_maps/defaults.json覆盖率门禁实现见 openmed/eval/domain_coverage.py合成 fixture 目录见 tests/fixtures/clinical。零样本 NER 的上手方式可参考 docs/zero-shot-ner.md 与 docs/zero-shot-howto.md。【免费下载链接】openmedLocal-first healthcare AI: clinical NER HIPAA PII de-identification that runs 100% on-device. 2,200 medical models, 21 languages, Apple MLX Python, no cloud, no patient data leaving your network. Apache-2.0项目地址: https://gitcode.com/GitHub_Trending/ope/openmed创作声明：本文部分内容由AI辅助生成（AIGC），仅供参考