Castor Catalyst · Case studies

Catalyst in practice: three studies, three kinds of source

Version 2.2 · Updated 21 October 2026

Three studies, customers unnamed, that get source data into the electronic data capture (EDC) system: a direct-to-participant registry, a site-upload study with paper patient-reported outcomes (PROs), and a post-market glaucoma study.

Each study is labelled with its evidence type; none are peer-reviewed.

Registry with no site1 hour

from enrollment to data in the EDC

What happens when there is no site?
Paper questionnaires0.8%

reviewer override rate on AI-extracted fields

What happens when sites upload paper questionnaires?
Glaucoma study0–1 queries

on primary efficacy data, down from up to 15

What about handwritten worksheets and medication data?
Case 1 · Direct-to-participant registry

Medical record data in the EDC without a site visit

Anonymised customer case study, 2025

The problem

A registry needed clinical data from participants' medical records, with no brick-and-mortar site to abstract it. Retrieving and abstracting records by hand was slow and labour-intensive.

How it works

  1. Participants consent to the study and authorise patient-mediated retrieval of their own records through Fast Healthcare Interoperability Resources (FHIR).
  2. Records arrive as United States Core Data for Interoperability (USCDI) v1 data classes.
  3. The records are mapped to study data domains. FHIR data is mapped mostly deterministically; AI checks the mapping and reads free-text notes.
  4. A reviewer sees each data point next to its FHIR resource ID and approves or overrides it under the review policy the sponsor or CRO sets.
  5. Values cleared under the review policy go to the EDC, with no personal identifiers.

In the current workflow, records a participant authorises are redacted by the Castor Catalyst redaction engine in the Source Gateway before the AI pipeline uses them.

Fictional example · values invented for this page
In this fictional example, one laboratory Observation supplies one eCRF value. The reviewer sees the value next to the FHIR resource it came from.

  • Why it works

    • FHIR data is structured, so it is mapped mostly deterministically; AI checks the mapping and reads free-text notes.
    • Each value keeps the ID of the resource it came from, so a reviewer can trace it back to its source.

  • What it means for sites

    • Where participants can authorise retrieval, no site has to abstract anything.
    • Participants submit their own records, so the site does not approve redactions for them.
    • The hospital that holds the records needs no integration project, interface build or IT cost.
  • 1 hour

    from enrollment to data in the EDC

  • Under 4 hours

    of human effort, down from 40 hours

Case 2 · Paper patient-reported outcomes

Sites upload paper questionnaires; reviewers check each answer under the review policy the sponsor or CRO sets

Production telemetry, 2026

The problem

Participants completed PRO questionnaires on paper at the site. Every answer had to be typed into the EDC a second time.

How it works

  1. Sites upload the completed questionnaires in bulk, by participant and time point. For the site, the task is to log in and upload; there is no case report form to complete for these questionnaires.
  2. On upload, Catalyst checks the participant ID and time point in the page header against the study. If the check fails, a person can override it, and the override is recorded in the audit trail.
  3. Catalyst extracts each answer and maps it to the electronic case report form (eCRF).
  4. Reviewers check answers against the highlighted source, under the review policy the sponsor or CRO sets, before they are committed to the EDC. Today every extracted value goes to a Castor reviewer. Where a page is unclear, such as two answers circled or a missing assessment date, the reviewer follows the study's written review rules and leaves a note on the value.
Questionnaire page from the study · redacted · evidence boxes added
A paper quality-of-life questionnaire page with the protocol, site and participant numbers and the time point redacted in black. A blue box marks each circled answer, and a label next to it shows the value for the eCRF, such as 2, A little, or 1, Not at all.
A questionnaire page from the study, with the protocol, site and participant numbers and the time point redacted for this page. Each box marks a circled answer; the label shows the value extracted for the reviewer to confirm before it goes to the eCRF. The page header, redacted here, carries the participant and time point that Catalyst checks on upload.

  • Why it works

    • Questionnaires are fixed forms: the same pages, the same header and the same answer grid for every participant, so each circled answer maps to one eCRF field.
    • The page headers repeat on every page, so the participant and time point can still be checked when the cover page is missing.
    • Every override a reviewer makes is stored in the audit trail.

  • What it means for sites

    • No second entry of every answer into the EDC.
    • Sites log in with the Castor account they already use; upload access is a permission in the study's user roles.
    • Training on the upload flow takes about 30 minutes.
  • 0.8%

    reviewer override rate on AI-extracted fields: how often a reviewer corrected the AI. An override rate, not an accuracy rate.

  • 99.8%

    submission success rate for values committed to the EDC

Case 3 · Post-market glaucoma study

Handwritten pressure worksheets and medication data for the primary endpoint

Sponsor-reported results

The problem

The pressure readings and medication data that drive the primary efficacy endpoint took about 80% of the sponsor's data-cleaning time.

Catalyst was scoped to the handwritten intraocular pressure (IOP) worksheet collected at every scheduled visit, and to IOP and IOP-lowering medication data, from the worksheets and, at sites on the EMR route, from the electronic medical record (EMR). Sites use one of two routes.

Route 1: the paper worksheet

  1. The site uploads the completed worksheet under the participant and time point.
  2. On upload, Catalyst checks whether a document is a paper worksheet or an EMR export. For a worksheet, it also checks the participant ID and time point. If the site has already entered data for that visit in the EDC, the upload is blocked, so existing data is not overwritten.
  3. Catalyst extracts every pressure reading per tonometer. Medications are standardised against the study's list, so medications are counted by active ingredient.
  4. A reviewer confirms the values against the highlighted source, including the study eye, which an EDC edit check also covers. Today every extracted value goes to a Castor reviewer.
  5. The EDC applies the protocol's derivation to the readings. The paper workflow was planned over eight weeks.
Worksheet from the study · redacted · evidence boxes added
A handwritten IOP worksheet with the study name, participant number and visit redacted in black. Blue boxes mark the study eye, OS; three handwritten medications, Rocklatan, Rhopressa and dorzolamide-timolol, labelled with their active ingredients; and two Tono-Pen readings, 18 and 19 mmHg.
The study's IOP worksheet, with the study name, participant number, visit and form date redacted for this page. The boxes mark the values extracted for the reviewer to confirm before they go to the eCRF: the study eye, each medication by active ingredient (Rocklatan becomes netarsudil and latanoprost), and the two Tono-Pen readings, 18 and 19 mmHg.

Route 2: the EMR export

  1. Before the first EMR upload for a participant, the site selects the study eye in the EDC.
  2. The site exports the visit note from the EMR, such as Epic, as a PDF and uploads it under the participant and time point.
  3. Catalyst checks whether the upload is a paper worksheet or an EMR export. If it cannot tell, the site selects the document type and continues the upload. For an EMR export, Catalyst takes the study eye from the EDC; if none is selected, the upload stops, and the site selects it and uploads again.
  4. For this study's EMR route, Catalyst reads the IOP readings per device for the study eye and the current visit only. Readings in prior-visit tables and history sections, T-max, target IOP and pachymetry are not taken, and a same-visit recheck replaces the first reading.
  5. Medications from the note's medication list and narrative are mapped to the study's list of IOP-lowering medications. Entries that are struck through, stopped, or named only as an allergy or a reaction are left out, and the reviewer sees each one that was left out, with the reason. Today every extracted value goes to a Castor reviewer.
Synthetic clinic notes · fictional patients, no real patient data · evidence boxes added
Two synthetic Epic-style clinic notes, one above the other, both with the left eye (OS) as study eye and the patient banner blacked out. In the top note, the left-eye Goldmann reading of 19 is boxed in blue as the study eye value; T-max, target IOP, corrected IOP and pachymetry, the right-eye reading and the maximum pressure row are boxed in grey as skipped. In the bottom note, Cosopt and Xalatan, both used in the left eye, are boxed in blue and labelled Dorzolamide Timolol and Latanoprost; an Alphagan allergy, a stopped Lumigan and a timolol reaction are boxed in grey as skipped.
Two synthetic clinic notes in the shape of an Epic export, made for testing, one above the other; in both, the study eye is the left eye (OS), and the patient banner is blacked out for illustration. Blue marks values extracted for the reviewer to confirm before they go to the eCRF: the study eye's Goldmann reading of 19, and Cosopt and Xalatan, used in that eye, as their study medications. Grey marks what is skipped: T-max, target and corrected pressure, the fellow eye's reading, the maximum pressure, an allergy, a stopped drug and a reaction.

  • Why it works

    • The worksheet is the study's own form, used across sites, so each value has a fixed place on the page.
    • The tonometer is inferred from where the readings sit on the worksheet.
    • The derivation runs as one calculation in the EDC, so the rule is applied the same way at every site.
    • On the EMR route, the study eye comes from the EDC, so only that eye's readings are taken.

  • What it means for sites

    • Sites upload the worksheet they already fill in at the visit, or, on the EMR route, the visit note they already have in the EMR.
    • On the EMR route, the site selects the study eye in the EDC before the first upload.
    • No second entry of pressure readings and medications into the EDC.
    • For an unscheduled visit, the site creates the visit in the EDC and uploads against it.
    • If handwriting is illegible or a value is missing, the reviewer queries the site from the review screen.
  • 0–1 queries

    on primary efficacy data, down from up to 15

  • 5 working days

    from upload to a locked data point, versus about three weeks

  • No SDV visits

    for the Catalyst form in the sponsor's monitoring plan (SDV: source data verification)

“It goes from upload to a ready-to-analyse, locked data point in five business days.”

Sponsor of a post-market glaucoma study, October 2026
What sites can upload

The documents a study sets up, as PDF

Each study's document types are set up during configuration, as in the paper questionnaire and glaucoma studies above.

  • Document types a study can set up

    • Paper source worksheets, including handwritten ones
    • Patient-reported outcome questionnaires
    • EMR exports and screenshots, such as visit and clinic notes
    • Scanned pages and phone photos of these documents
    • Other documents, such as discharge letters, clinic letters or lab reports, where the study sets them up as a document type

    Sites upload only the document types set up for their study.

  • File format

    • PDF only. Save phone photos and screenshots into a PDF before upload.
    • Up to 30 MB and 1,000 pages per PDF.
    • On studies where Castor redacts uploads for the site, up to 20 pages per PDF.
Next step

See it on the study's own documents

Castor benchmarks the extraction on the sponsor's documents and scopes the plan.

  1. Share the study

    A protocol, the forms in scope and sample source documents. Synthetic samples are welcome.

  2. Talk to Castor about Catalyst

    Leave contact details in the form below.

  3. How Catalyst works

    End to end, including how it compares with manual chart review: read How Catalyst works.