D007 检验与生命体征标准化｜个人思维图谱小站

正文

SKILL.md

v0.2.1 发布合同（优先于历史版本说明）

用已识别目录kind纠正lab/vital类型，保存classification_lineage，不把生命体征再按lab传下去。保留unit_system_raw与unit_interface_raw，冲突来源单位仍review，不臆造换算。

详见 references/repair-v0.2.1.md。完整替换D001–D014并重新生成中间产物，不混用历史伪ID、清单或staging编号。测试通过不表示临床生产批准。

v0.2.0 修复合同（优先于旧版说明）

补齐本次来源数据中明确的项目名称和 meq/L 单位别名；不推导新的临床换算公式。按项目/单位兼容性保留已知原单位。重复候选必须限定同患者、同就诊等身份；不同就诊重用 specimen_id 不自动合并。

具体运行界限见 references/repair-v0.2.md；任何自测通过都不等于临床生产批准。

Mission

Turn D006 lab and vital facts into comparable, traceable measurement facts while preserving the original clinical observation.

D007 answers:

What measurement is this, at a local canonical-name level?
Is the result numeric, comparator-valued (<, >=) or qualitative?
Is the unit recognized?
Can the result be safely converted to the local preferred unit using a reviewed deterministic rule?
Can the reference range be parsed and converted under the same rule?
Is the result low/normal/high from the stated reference range?
Which specimen/accession identity belongs to the result?
Are two results strong duplicate candidates, same-specimen different results, or merely possible repeats?

D007 does not answer:

the LOINC code;
the OMOP concept ID;
whether a duplicate candidate should be deleted or merged;
whether a lab result is clinically significant or diagnostic;
missing/ambiguous units by guessing from typical practice.

Required upstream input

Required:

D006 context_resolution_manifest.json with overall_status=ready_for_domain_normalization (or no_facts);
finalized D006 contextualized_facts.jsonl.

Never read raw pre-D003 patient sources.

Core measurement contract

Each D006 lab/vital fact is enriched with:

{
  "measurement": {
    "canonical_key": "glucose",
    "canonical_name": "Glucose",
    "name_resolution": "exact_alias",
    "value": {
      "raw_text": "202",
      "relation": "eq",
      "numeric_value": "202",
      "raw_unit": "mg/dL",
      "recognized_unit": "mg/dL",
      "normalized_value": "11.210456",
      "canonical_unit": "mmol/L",
      "conversion_status": "converted",
      "conversion_rule_id": "glucose-mgdl-to-mmoll"
    },
    "reference_range": {},
    "interpretation": {},
    "identity": {
      "specimen_id": "...",
      "accession_id": "..."
    },
    "normalization_status": "normalized",
    "review_status": "auto_resolved"
  }
}

Critical rule: do not guess units

If a glucose value is 11.2 with no unit, D007 must not assume mmol/L.

Correct:

unit_missing -> needs_review

Incorrect:

11.2 -> mmol/L because that is common locally

Critical rule: measurement-specific conversion only

D007 is not a generic dimensional-conversion engine for clinical data.

For example:

glucose mg/dL <-> mmol/L is allowed because the analyte identity and reviewed factor are explicit;
creatinine mg/dL <-> umol/L is allowed;
an unknown analyte in mg/dL must not be converted to mmol/L by guessing molecular mass;
HbA1c % <-> mmol/mol is intentionally not converted in v0.1.

Read references/unit-conversion-policy.md before extending conversions.

Workflow

1. Prepare deterministic normalization

python "$SKILL_DIR/scripts/prepare_measurement_normalization.py" \
  --context-manifest /path/to/d006/context_resolution_manifest.json \
  --output-root ./runs/haonan-d007

Outputs:

auto_normalized_measurement_facts.jsonl
duplicate_candidates.jsonl
measurement_review_queue.csv
agent_measurement_overrides.jsonl
measurement_normalization_manifest.json

2. Review only unresolved measurement identity

Codex may review a queued fact only to select a canonical_key that already exists in assets/measurement_aliases.json.

Do not invent a new conversion formula during a patient-data run. Do not copy patient values into the override file.

Example:

{
  "fact_id": "fact_x",
  "canonical_key": "glucose",
  "review_status": "agent_confirmed",
  "reason_codes": ["resolved_abbreviation_from_source_context"]
}

If the raw unit remains missing/unrecognized after identity review, the fact still remains needs_review.

3. Finalize

python "$SKILL_DIR/scripts/finalize_measurement_normalization.py" \
  --run-dir /path/to/d007/<package-id>

4. Validate

python "$SKILL_DIR/scripts/validate_measurement_normalization.py" \
  --manifest /path/to/d007/<package-id>/measurement_normalization_manifest.json

Result parsing

Supported deterministic forms include:

5.2
<5
<=5
>=10

Qualitative values such as positive, trace, pending are preserved but not coerced to numbers in v0.1.

Reference ranges

D007 parses simple ranges such as:

3.9-6.1
135 to 145
<5
>=10

If the measurement value is converted under an allowlisted rule, numeric reference-range bounds are converted under the same rule.

D007 does not manufacture population-specific reference ranges.

Interpretation

If an explicit upstream lab flag is available, preserve it separately. If a numeric value and parseable reference range are available, D007 may derive:

low | normal | high | unknown

Derived interpretation is data normalization, not diagnosis.

Duplicate and repeat handling

D007 never removes facts.

It may emit:

strong_duplicate_candidate
same_specimen_different_result
possible_repeat_measurement

A strong duplicate requires the same patient, same canonical measurement, shared specimen/accession identity, equivalent normalized values and compatible time evidence.

A different value from the same specimen/accession is not automatically a duplicate; it may represent a corrected/repeated result.

Final merge decisions belong to D010.

Plausibility warnings

D007 uses deliberately broad data-quality bounds for selected vital signs. A violation adds a warning flag but never deletes or clinically interprets the value.