技能 人工智能 临床病症列表重构与规范化

临床病症列表重构与规范化

v20260803
reconciling-problem-lists
本技能用于处理从临床病历中提取的原始疾病和诊断提及。它负责解决同义词的冲突、排除患者否认或仅为假设的诊断,并为每个概念分配明确的临床状态(活跃、已解决或历史)。最终输出是一个标准化、去重且结构化的问题列表,适用于USCDI或FHIR标准。
获取技能
488 次下载
概览

Reconciling problem lists

A single note mentions the same condition many ways — "DM2," "type 2 diabetes," "diabetes mellitus" — across PMH, HPI, and A&P, some negated, some historical. A usable problem list collapses those mentions into one concept per problem, drops what the patient does not have, and assigns a clinical status (active / resolved / historical). This skill turns OpenMed's per-mention entity stream plus ConText axes into that reconciled, de-duplicated list, shaped for USCDI "Problem" exchange.

When to use

  • After extracting-clinical-entities and resolving-clinical-context, when the user wants a clean problem list, condition reconciliation, or dedup of repeated diagnosis mentions.
  • You need active-vs-resolved-vs-historical status per problem, not just raw mentions.
  • You are assembling a FHIR Condition list or a USCDI Problem element and need one entry per concept.

Quick start

import openmed
from openmed.clinical import resolve_span_context, NEGATED, HISTORICAL, HYPOTHETICAL

note = ("PMH: type 2 diabetes, prior MI 2019 (resolved). "
        "A&P: poorly controlled DM2; denies chest pain.")

ents = openmed.analyze_text(note, model_name="disease_detection_superclinical",
                            output_format="dict")

def normalize(surface: str) -> str:
    # Cheap synonym folding; replace with SNOMED grounding (out-of-process).
    s = surface.lower().strip()
    return {"dm2": "type 2 diabetes", "diabetes mellitus": "type 2 diabetes"}.get(s, s)

problems = {}  # concept -> reconciled record
for e in ents:
    surface = e["word"]
    ctx = resolve_span_context(surface, note)
    if ctx.negation == NEGATED:
        continue                                   # patient does NOT have it -> exclude
    concept = normalize(surface)
    status = ("resolved" if ctx.temporality == HISTORICAL else
              "active")
    if ctx.temporality == HYPOTHETICAL:
        continue                                   # not asserted as present
    rec = problems.setdefault(concept, {"concept": concept, "status": status,
                                        "mentions": 0})
    rec["mentions"] += 1
    # Active anywhere wins over a historical mention of the same concept.
    if status == "active":
        rec["status"] = "active"

problem_list = list(problems.values())
# -> [{"concept": "type 2 diabetes", "status": "active", "mentions": 2}, ...]
# "chest pain" excluded (negated); "MI" -> historical/resolved.

Workflow

  1. Collect Disease/Condition entities from analyze_text across the whole note (or per section if you ran segmenting-clinical-sections).
  2. Attach clinical context per mention with resolve_span_context (or the axes from resolving-clinical-context): negation, temporality, uncertainty.
  3. Exclude what isn't a problem. Drop NEGATED mentions (patient denies / no evidence of) and HYPOTHETICAL mentions (conditional, not asserted). These must never land on the active list.
  4. Cluster synonymous mentions into one concept. Fold surface variants (abbreviations, word order, lexical synonyms) to a single canonical key. Cheap normalization gets you started; SNOMED CT concept grounding is the robust path — run it out-of-process with the user's own license and key on the concept code, not the surface string.
  5. Assign status by aggregating context. A concept that is RECENT/active anywhere (typically A&P) is active; one seen only as HISTORICAL ("history of," "resolved," PMH-only) is resolved/historical. Active wins over historical when the same concept appears both ways.
  6. Emit the reconciled list — one record per concept with status, mention count, and provenance offsets — shaped for USCDI Problem / FHIR Condition.

Hand-off to / from OpenMed

  • From extracting-clinical-entities: consumes analyze_text Disease entities. Run on a sectioned note (segmenting-clinical-sections) for best active-vs-historical signal.
  • From resolving-clinical-context: this skill depends on the negation / temporality / uncertainty axes — reconciliation without them would put "denies chest pain" on the active list.
  • OpenMed calls: from openmed import analyze_text and from openmed.clinical import resolve_span_context, NEGATED, HISTORICAL, HYPOTHETICAL.
  • To FHIR / USCDI: each reconciled problem becomes a Condition with clinicalStatus active/resolved (from temporality) and verificationStatus refuted/provisional (from negation/uncertainty). SNOMED CT codes are user-supplied and grounded out-of-process — OpenMed produces the dedup'd concept and status, not the terminology binding.

Edge cases & gotchas

  • Surface dedup is lossy. "MI" and "myocardial infarction" only fold if your normalizer knows the synonym. Lexical folding handles the easy cases; lean on SNOMED CT grounding for real reconciliation, and never bundle SNOMED — call it out-of-process with the user's credentials.
  • Active beats historical for the same concept. "History of asthma" in PMH plus "asthma exacerbation" in A&P is one active problem, not two entries. Aggregate before assigning status.
  • Don't resurrect resolved problems. A concept seen only as HISTORICAL / "resolved" stays resolved; don't promote it to active just because it appears.
  • Negated and hypothetical are exclusions, not statuses. They never become problem-list entries. Keep them out entirely.
  • Carry provenance. Keep offsets / source sections per problem so a reviewer can trace each entry back to the note text.
  • Local-first, advisory-only. Runs on-device; the reconciled list is decision support for clinician review, not an autonomous diagnosis.

Standards & references

信息
Category 人工智能
Name reconciling-problem-lists
版本 v20260803
大小 6.94KB
更新时间 2026-08-04
语言