Skip to main content

Focus curation on the documents that matter

Use Information Extraction findings to rank a patient's documents and, by default, to narrow which of them Data Curation actually processes. This page covers the two controls that do the work, how to source the Information Extraction results they depend on, and how to check afterward which notes were selected and which were filtered out. It stops once you have a curation job whose results and evidence you can review.

Who this is for​

Anyone who has to curate patients whose charts run to hundreds of documents each:

  • Cancer registrars and oncology abstractors working case by case against a registry standard
  • Data curators configuring curation jobs for a study or registry project
  • Clinical subject-matter experts who define what an extraction profile should look for
  • The platform administrator who confirms that Information Extraction and the ranking and terminology services are configured in the deployment

Why it matters​

Take a registry curation job over an oncology cohort where each patient carries several hundred notes. The facts you need sit in a handful of them: the receptor status in a pathology report, the stage in a staging summary or an oncology consult note, a recurrence hint in one imaging report. Everything else is refill messages, phone notes, and routine follow-up visits that were never going to contain the answer.

Sending the full chart to Data Curation spends compute on those documents anyway. Worse, it hands reviewers a result set where most rows are empty. A reviewer who works through long stretches of nothing gets slower, and accuracy on the rows that do matter drops with it. The cost is not only runtime, it is attention spent in the wrong place.

Ranking changes what reaches the reviewer. Information Extraction runs first and reports where the relevant findings are, curation concentrates on those notes, and the evidence behind each curated field comes from documents that were selected for a reason you can inspect. PJI structures the documents and surfaces the evidence; the clinical judgment about what the record means stays with the clinicians and abstractors reviewing it.

What you gain​

A curation job that prioritized the informative notes per patient, with the ranking detail persisted so you can see which documents were kept, which were filtered out, and why. When Augment is enabled, the linked Information Extraction results also sit alongside the curated fields as additional evidence, so a reviewer checking a value has both the curation evidence span and the extraction finding that pointed at that document.

The result survives scrutiny because nothing is hidden. Filtering is a recorded decision, not a silent drop: you can open the ranking view for a patient and read the set that curation worked from. Where the filter was too aggressive, you can see that too, lower the threshold, and rerun. What you cannot get from this is a result that needs no review. Filtering and augmenting improve focus and evidence quality, and the curated fields still need a human to validate them.

Before you start​

  • A dataset containing the target patients and documents. Data Curation starts from a dataset, not from a raw source. See Build a dataset from a cohort or ingestion
  • The registry, ontology, or curation template that matches the abstraction goal. A template that does not match the documents produces sparse results whether or not you filter
  • Either a completed Information Extraction job for the same dataset and NLP pipeline, or a clear extraction goal if you plan to create a new run during Augment
  • Confirmation that Information Extraction and the supporting deployment services, such as terminology and ranking, are configured in your environment. Both controls on this page depend on them
  • Familiarity with the base curation flow in Start Data Curation from a dataset

Step 1: Start the curation job and set the extraction scope​

  1. Create or select the dataset that contains the target patients and documents.
  2. Start a new Data Curation job.
  3. Select the cancer registry, ontology, or curation template that matches the abstraction goal.
  4. In Configuration, choose Patient-level extraction when you want one consolidated result per patient across documents. This is the setting that makes per-patient ranking meaningful, since the job is reasoning across the patient's whole document set rather than each note alone.

Step 2: Turn on Augment with Information Extraction​

  1. Open Reference Results in the job configuration.
  2. Enable Augment with Information Extraction. Information Extraction findings are then passed into Data Curation as additional evidence, and they appear alongside the curated fields during review.

Augment is the parent control. Filter documents only appears once Augment is enabled, which is why the two are easy to confuse: enabling Augment alone changes the evidence available to curation, and it is the filter underneath that changes which documents curation reads.

Step 3: Decide how Filter documents behaves​

Filter documents opens under Augment with Information Extraction and is checked by default. Leave it enabled so curation focuses on the most relevant notes using the Information Extraction findings. Adjust the ranking threshold if needed: lower keeps more documents, higher keeps only stronger matches. 0.5 is a balanced default.

You can also augment without filtering. The two combinations behave differently:

ConfigurationDocuments sent to Data CurationIE findings as additional evidenceEffect on the run
Augment with Information Extraction on, Filter documents on (the default)Only the notes that meet the ranking thresholdYesLess noise in the result set, shorter run, and a persisted record of what was filtered out
Augment with Information Extraction on, Filter documents offEvery document in the datasetYesEvidence is still enriched, but the run may take longer because nothing is narrowed
Augment with Information Extraction offEvery document in the datasetNoThe standard curation run. Filter documents is not available, since it opens under Augment

Turn Filter documents off when you need curation to read the complete chart, for example on a case where an unranked note may still carry the fact, and you want the Information Extraction findings only as extra evidence.

Step 4: Point Augment at an Information Extraction job​

Choose one of the two sourcing options:

  • Use existing job reuses a completed Information Extraction run for the same dataset and selected NLP pipeline. Pick this when a suitable run already exists and its extraction goal matches what you are curating.
  • Create new job starts a new Information Extraction run as part of the curation setup. With create new, Information Extraction runs first, and when it finishes, Data Curation starts automatically using those results. You do not start the curation stage by hand.

If you are creating a new zero-shot or multitask extraction and no suitable profile exists, prepare one in Information Extraction before you configure Augment:

  1. Open Custom Profiles and select a multitask zero-shot model.
  2. Use the prompt-assisted flow to describe what should be extracted, for example biomarkers, staging terms, or recurrence clues.
  3. Review the generated entities, refine the names and descriptions, then save the custom profile. Vague prompts produce noisy or incomplete profiles, so this review is not optional.
  4. Return to the curation job and select that profile when creating the new Information Extraction job under Augment.

Step 5: Run the job, then review what was selected​

  1. Complete job naming and review the configuration, then start the automation.
  2. Monitor progress. With create-new Augment, wait for the Information Extraction phase to complete before the document filtering and curation stages advance.
  3. Review the curated fields with their evidence once the job finishes.
  4. Use Document Analysis or the per-patient ranking views to see which notes were selected and which were filtered out.
  5. If a document you expected to carry a fact was filtered out, lower the ranking threshold or revisit the extraction profile, then rerun. A filter tuned for one abstraction goal will not suit a different one.

Limitations​

  • Augment with Information Extraction and Filter documents depend on Information Extraction completing successfully, and on deployment configuration such as terminology or ranking services. Where those services are not configured, the controls are not usable.
  • Filtering can exclude clinically useful notes if the threshold is set too high, or if the Information Extraction profile does not match the curation goal. The filter ranks against what the profile looks for, not against what the registry template asks for.
  • You can augment without filtering. Information Extraction results still enrich the evidence, but all dataset documents are sent to Data Curation and the run may take longer.
  • Prompt-generated multitask entities should be reviewed before production use. A vague prompt produces a noisy or incomplete profile, and that profile then drives both the evidence and, when filtering is on, the document set.
  • Filtering and augmenting improve focus and evidence quality. They do not replace human review of curated fields.