D008 用药标准化与动作判定｜个人思维图谱小站

正文

获取原文件

SKILL.md

v0.2.1 发布合同（优先于历史版本说明）

本模块既有运行逻辑保持；本次与其他模块统一发布0.2.1，并执行原有回归。不得仅凭版本号推断新增临床能力。

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）

兼容 D005 raw_* 字段并保留零值；明确接收 doses_per_24_hrs，表示来源每日次数，不臆造时间间隔。prescription 明确属于 medication_order，不能推断已给药。

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 medication facts into explicit, traceable medication events without inventing drug identity or exposure.

D008 answers:

What drug text was explicitly mentioned?
Can dose, dose unit, route, frequency, duration and PRN be parsed deterministically?
Does the evidence explicitly describe start, continue, hold, stop, resume, increase, decrease, replace, or only a neutral medication mention?
What kind of medication evidence is this: home medication, order, administration, discharge plan, medication list, or narrative mention?
Does the action require manual review because the wording is ambiguous?

D008 does not answer:

RxNorm, ATC, SNOMED or OMOP codes;
whether an ordered medication was actually administered;
whether a home medication was still active during admission;
whether two mentions should be merged into one longitudinal medication state;
whether the medication is clinically appropriate.

Required upstream input

Required:

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

Optional but recommended for narrative medication facts:

D004 structured document package, so D008 can recover a local de-identified evidence window using provenance offsets.

Never read raw pre-D003 patient sources.

Core output contract

Each medication fact is enriched with:

{
  "medication": {
    "drug_text": "二甲双胍",
    "drug_text_normalized": "二甲双胍",
    "dose": {
      "value": "500",
      "unit": "mg",
      "form": null,
      "normalization_status": "converted_from_g"
    },
    "route": {
      "code": "PO",
      "source_text": "口服"
    },
    "frequency": {
      "code": "BID",
      "times_per_day": 2,
      "interval_hours": null,
      "source_text": "bid"
    },
    "duration": {
      "value": null,
      "unit": null
    },
    "prn": false,
    "action": {
      "code": "hold",
      "confidence": 1.0,
      "cue": "暂停"
    },
    "evidence_class": "narrative",
    "exposure_assertion": "not_equivalent_to_administration",
    "normalization_status": "normalized",
    "review_status": "auto_resolved"
  }
}

Critical rule: action is not exposure

These statements are different facts:

长期服用二甲双胍
开具二甲双胍医嘱
给予二甲双胍500mg
暂停二甲双胍
出院继续二甲双胍

D008 must not flatten them into medication=metformin, active=true.

Instead preserve the evidence class and action.

Evidence classes

D008 uses the following local classes:

home_medication — medication reported before the current encounter;
medication_order — an order/prescription exists, not proof of administration;
medication_administration — documented administration/MAR event;
discharge_medication — discharge medication plan/list;
medication_list — list entry without stronger exposure semantics;
narrative — medication mention in narrative text;
unknown — source semantics are not sufficiently clear.

A source class must never be silently upgraded. In particular:

order != administration
medication list != administration
home medication != current inpatient administration

Medication action vocabulary

Allowed action codes:

start
continue
hold
stop
resume
increase
decrease
replace
prn_use
unknown

hold means temporarily withheld/suspended. stop means explicitly discontinued. resume means a previously held/stopped medication was explicitly restarted. replace records a replacement/switch statement, but D008 does not infer the target pair unless both facts are explicitly represented.

Critical rule: do not infer an action from the drug name alone

Examples:

二甲双胍 500mg bid

Without source semantics or an action cue, action is unknown, not continue.

长期服用二甲双胍 0.5g bid

May be continue only as a home medication context, not proof of inpatient administration.

暂停二甲双胍

→ hold

停用二甲双胍

→ stop

恢复二甲双胍

→ resume

Negative medication mentions

If D006 already resolved a medication mention as negative/refuted, D008 must not turn it into a positive exposure.

Example:

患者未使用二甲双胍

D006:

polarity = negative

D008 preserves the parsed drug text but sets:

exposure_assertion = negative_mention

and does not assign start/continue merely because a medication name appears.

Dose normalization

D008 only performs reviewed, deterministic mass conversions:

1 g = 1000 mg
1 mg = 1000 ug

Examples:

0.5 g -> 500 mg
500 mg -> 500 mg

D008 does not guess drug-specific strength equivalence and does not convert tablets/ampoules to mass unless the evidence explicitly states the strength.

1片 remains one tablet; it does not become 500 mg from model memory.

Route normalization

D008 may normalize explicit route aliases such as:

口服 / po -> PO
静脉注射 / iv -> IV
静滴 / ivgtt -> IV_INFUSION
肌注 / im -> IM
皮下 / sc -> SC
吸入 -> INHALATION
外用 -> TOPICAL

Missing routes remain unknown.

Frequency normalization

Reviewed aliases include:

qd / qd. / 每日一次 -> QD
bid / 每日两次 -> BID
tid / 每日三次 -> TID
qid / 每日四次 -> QID
qhs / 睡前 -> QHS
qod / 隔日一次 -> QOD
q4h -> Q4H
q6h -> Q6H
prn / 必要时 -> PRN

Frequency normalization is semantic normalization, not a dosing recommendation.

Workflow

1. Prepare deterministic medication normalization

python "$SKILL_DIR/scripts/prepare_medication_normalization.py"   --context-manifest /path/to/d006/context_resolution_manifest.json   --output-root ./runs/haonan-d008   --structured-doc-root /optional/path/to/d004

Outputs:

auto_normalized_medication_facts.jsonl
medication_review_queue.csv
agent_medication_overrides.jsonl
medication_normalization_manifest.json

2. Review unresolved cases only

Codex may review queued medication facts using only de-identified D004 evidence.

Allowed review actions:

choose an action already defined in assets/action_rules.json;
choose an evidence class already defined by D008;
confirm route/frequency aliases already present in assets;
mark the fact needs_review.

Codex must not:

invent a drug code;
invent a dose not present in evidence;
infer administration from an order;
infer a current medication list from one narrative mention;
copy patient prose into the override file.

Example override:

{
  "fact_id": "fact_123",
  "action": "hold",
  "evidence_class": "narrative",
  "review_status": "agent_confirmed",
  "reason_codes": ["explicit_pause_cue_in_deidentified_window"]
}

3. Finalize

python "$SKILL_DIR/scripts/finalize_medication_normalization.py"   --run-dir /path/to/d008/<package-id>

4. Validate

python "$SKILL_DIR/scripts/validate_medication_normalization.py"   --manifest /path/to/d008/<package-id>/medication_normalization_manifest.json

Longitudinal boundary

D008 deliberately outputs events rather than a final medication state.