Rapid Case Ascertainment (RCA) — NLP Prototype Notes

These are working notes for a lightweight RCA pipeline that flags probable reportable cancers from pathology text. The page is hidden from the sidebar and sitemap but can be linked from coursework pages.

1) Problem & Objective

Goal: reduce reporting lag by surfacing likely reportable cases within days of pathology sign-out.
Scope: pathology synoptic/diagnostic text → rules/NLP flags → reviewer queue.

Not a replacement for full abstraction—this is triage to accelerate case finding.

2) Minimal Signal Set (first pass)

  • Topography/Morphology cues: site (breast, colon, lung), histology terms (adenocarcinoma, lymphoma, SCC).
  • Behavior: malignant/primary vs in-situ vs benign indicators.
  • Staging fragments: “pT”, “pN”, “pM”, “Stage I–IV”, “AJCC”.
  • Specimen context: biopsy vs resection; exclude consults when no new diagnosis.

Start with a curated lexicon from ICD-O-3 site/histology names (non-copyright summaries) and a small hand-built synonym list.

3) Rule Sketch (deterministic baseline)

  • Positive rule (P1): malignan*|carcinoma|lymphoma|sarcoma AND a site cue (e.g., breast colon lung prostate)
  • Primary hint (P2): presence of “primary” OR absence of “metastatic from …”
  • Exclusions (E): “negative for malignancy”, “benign”, “no evidence of”, “metastatic from [other site]” w/o primary context

Flag case if: (P1 AND P2) AND NOT E

Keep rules transparent so registrars can tune them.

4) NLP Enhancements (incremental)

  • Sectioning: split Synoptic vs Narrative; prioritize Synoptic blocks.
  • Negation/uncertainty: simple patterns (no evidence of, cannot rule out) to drop false positives.
  • Terminology mapping: light mapping of histology variants → canonical tokens (e.g., “adeno ca” → adenocarcinoma).
  • De-duplication: key on MRN + specimen date + site; collapse repeat hits.

5) Reviewer Queue (human-in-the-loop)

Columns:

  • MRN (hashed if needed), specimen date, facility
  • Site cue, histology cue, confidence (rule count), text snippet (first 200–300 chars)
  • Buttons: Confirm, Dismiss, Route to Abstractor

Capture reviewer actions to refine rules.

6) Micro-quality checks

  • Precision sample: 50 flagged reports → % truly reportable
  • Recall proxy: compare to cases captured later in registry for overlapping period
  • Lag metric: median days from report sign-out → reviewer confirmation

Targets for pilot:

  • Precision ≥ 0.85, Recall proxy ≥ 0.75, Median lag ≤ 5 days

7) Privacy & Governance (pilot)

  • Process on secured host; log access; retain minimum necessary text.
  • If exporting examples, remove identifiers and mask rare descriptors.
  • DUA/IRB as appropriate; document data flow.

8) Tiny roadmap

  1. Week 1–2: rule baseline + reviewer queue mock
  2. Week 3–4: negation + de-dup + metrics dashboard
  3. Week 5–6: precision/recall read-out → rule tuning → go/no-go

9) Links & Anchors

  • ICD-O-3 (site/histology) look-up habit for rule vocabulary
  • NAACCR edits (for downstream consistency)
  • SEER Summary Stage vs AJCC notes (context for staging cues)

Footnote

This page is intentionally hidden (nav_exclude: true, sitemap: false).
Link to it from the CPH 705 parent page or from the Modular Analysis page when needed.