feat(report-anonymization): radiology-report de-identification skill (NeMo Anonymizer) - #44

report-anonymization skill

De-identifies English radiology reports by replacing detected PHI with bracketed role tokens ([PATIENT], [DOCTOR], [DATE], ...) using NeMo Anonymizer (GLiNER-PII detection + LLM validate/augment). Packaged as an agent-callable Medical AI Skill around the upstream nemo-anonymizer package.

Engineering / research use only — not a regulatory de-identifier and not for clinical/patient-facing use.

Contents (skills/report-anonymization/)

• SKILL.md, README.md, BENCHMARK.md, skill_manifest.yaml, requirements.txt
• scripts/anonymize_reports.py — wrapper: Redact strategy with a strict GLiNER label set, per-stage telemetry, and a run_report.json. Opt-in --mode rewrite and --iterative-redaction (novel critic-hinted loop) are off by default.
• validators/output_schema.json, evals/evals.json, references/, fixtures/
• data/synthetic_reports_100_w_PHI.csv — 100-case synthetic test set (no real PHI)

Results (100 synthetic reports, gpt-5.5 as LLM judge)

• With skill: 291/300 report-runs fully redacted (94-99/100 per backend)
• NVIDIA-default de-identifier (raw anonymizer run, gpt-oss-120b): 85/300 (40-45/100)
• Unaided (no doc): produced no working output

Quick test (needs NVIDIA_API_KEY)

python skills/report-anonymization/scripts/anonymize_reports.py \
  skills/report-anonymization/data/synthetic_reports_100_w_PHI.csv \
  --output-dir runs/quicktest --num-records 5

Notes / feedback wanted

• Draft for testing + feedback.
• Is the opt-in iterative-redaction extension in scope for this repo, or should it be dropped to keep the skill aligned with upstream NeMo Anonymizer?
• The eval harness that produced the results above lives separately; happy to add it if useful.