Dental claims with narratives and attachments for dental claim review AI

What a dental claims training record should contain

A training record is useful only when the claim line, its evidence and its decision are joined on a stable claim and line key. Dental claims travel as the ASC X12N 837 dental transaction (005010X224A2 with its errata) [6] or on the paper ADA Dental Claim Form, and payer companion guides add their own situational rules on top of the base guide. Buyers training review models should ask for the line-level fields, not a flattened claim summary.

The core fields are the procedure code, tooth number (Universal system 1-32 for permanent teeth and A-T for primary), tooth surfaces (M, O, D, B/F, L/I), oral cavity area or quadrant, date of service, submitted fee, prior placement date for crowns and prosthetics, and the missing-teeth indicator. Paperwork references matter too: the 837D PWK segment carries an attachment control number [6] that ties a claim to images sent separately, and without that key you cannot reliably match an X-ray to its line. Ask suppliers how attachments were linked historically, because practices often sent them through a separate attachment service or by mail.

Narratives, attachments and formats to specify

Narratives and attachments carry the clinical evidence that coded fields leave out, so specify them by type and format. Narratives are short free-text justifications such as "MOD fracture, cusp missing, recurrent decay under existing amalgam" for a crown, or pocket depths and bone loss notes supporting scaling and root planing. They are often abbreviated, copied from templates, or reused across patients, which is itself a signal for documentation-sufficiency models.

Attachments typically include bitewings, periapicals, panoramic images, occasional CBCT volumes, six-point perio charts (probing depths, recession, clinical attachment level, bleeding on probing), and intraoral photos. Formats vary by imaging system: some export DICOM, many export JPEG or PNG from proprietary practice imaging software, and perio charts may arrive as PDFs or screenshots rather than structured values. State the formats you can ingest, the minimum resolution, and whether you need structured perio values or will run OCR yourself.

Adjudication outcomes and label design

Adjudication outcomes are the labels, so define them more finely than paid versus denied. Dental review decisions usually fall into approved, approved with an alternate benefit (for example a posterior composite paid at the amalgam rate), downgraded to a different code, denied for frequency or history limits, denied for missing or insufficient documentation, and pended for more information. Remittances express adjustments with X12 Claim Adjustment Reason Codes, which X12 maintains and updates [1], but internal consultant decision codes are usually more specific and worth requesting alongside them.

Dental consultant comments ("radiograph does not show decay extending to dentin", "bone loss not evident on submitted BWX") are the most valuable text in the record for explanation-generation models. Pre-authorization (predetermination) records follow the same structure and let you train before-service review with fewer paid-claim selection effects, though they carry their own submission biases. For general guidance on remittance-based labels, see our page on claim denial prediction training data, and test whether outcome fields are reliable using verifying outcome labels in operational records.

Watch for label leakage and drift. Fee schedules, frequency rules and alternate benefit policies differ by plan and change by year, so the same image and narrative can produce different outcomes across groups. Ask for plan or policy version identifiers, even if coarse, and split training and evaluation data by time.

CDT code licensing and version control

The CDT Code is the American Dental Association's intellectual property, so plan for licensing before you ship anything that uses its descriptors. The ADA updates the CDT Code annually [5] and, as of October 2026, offers commercial licenses for its use in software, publications and payer systems; check the ADA's current terms for your use. A dataset that contains only code values is easier to handle than one that embeds full descriptors, but a model or product that displays or generates descriptors needs your own license review.

Version control matters for CDT code prediction. Codes are added, revised and deleted annually, so record the CDT version effective on each date of service and map retired codes rather than letting them appear as rare classes. A crown (for example D2740) or scaling and root planing (D4341 or D4342) should be learned against the CDT version in force on its date of service.

De-identifying claims, images and photos

Dental claims and attachments are protected health information when held by covered entities, so the release has to meet HIPAA de-identification through Safe Harbor or Expert Determination [3]. Safe Harbor requires removing 18 identifier types, including all date elements other than year and full-face photographs and comparable images [4]. That rule hits dental data hard: dates of service drive frequency logic, and extraoral photos are routine in orthodontic and prosthetic cases.

Images need pixel-level handling. DICOM's Attribute Confidentiality Profiles state that applying them does not guarantee removal of all identifying information and that they are one part of a broader process [2]. Exported radiographs often carry burned-in patient names, chart numbers or practice names, and JPEG exports may keep identifiers in EXIF or file names. Require text detection and redaction on pixels, exclusion or cropping of extraoral photos, and a reviewer sample check.

Where you need real date intervals, an Expert Determination that permits shifted dates is usually the better route. Our guide to HIPAA Safe Harbor vs Expert Determination for AI training covers the tradeoff, and reviewing an Expert Determination report explains what to check before signing.

Dental claims data request template

A precise request is the fastest way to learn whether suitable data exists. Describe the data and its uses, not the companies you think hold it.

Illustrative example: invented to show structure; it does not describe an available dataset.

request: dental claim review training data
use: train documentation-sufficiency classifier; hold out 20% for eval
claim_scope:
  transaction: X12 837D or ADA claim form images
  service_years: 2021-2025
  code_families: [D2xxx restorative, D4xxx periodontics, D6xxx implants]
  required_fields: [claim_id, line_id, cdt_code, cdt_version, tooth, surface,
                    quadrant_or_area, service_date_or_offset, fee_submitted]
evidence:
  narrative: required where present
  radiographs: [bitewing, periapical, panoramic]
  radiograph_format: DICOM or PNG, min 1000 px on long edge
  perio_chart: structured six-site values preferred
  photos: intraoral only; extraoral excluded
labels:
  outcome: [approved, alternate_benefit, downgraded, denied_frequency,
            denied_documentation, pended]
  consultant_comment: free text
  carc: X12 code on 835 where available
privacy:
  method: Expert Determination with date shifting
  images: pixel-level burned-in text redaction plus sample review
license_needs: model training and internal evaluation; term and delivery to be agreed

How SourceX sources dental claims data

SourceX sources operational datasets from US companies, including support histories, documents and finance and legal workflows, and manages the licensing process and ongoing purchases. Datasets are sourced on request rather than held in stock, so a request does not guarantee a match. SourceX looks for US businesses that hold the data you describe, and every release is approved by the supplying company.

The process runs Find, Assess (data and licensing permissions), Agree (pricing and allowed uses in a license), Transact and Manage, and nothing is contracted until a supplier agrees. Each dataset is rights-reviewed for ownership and consents and delivered under a license that defines records, uses, term and delivery. Health records require HIPAA de-identification by Safe Harbor or Expert Determination, personal details are removed or replaced with the method recorded and a sample checked, and no method is perfect. Diligence materials covering source, rights, preparation and allowed use are prepared per dataset, and delivery runs through private, access-controlled workflows only after an executed agreement and supplier approval. Buyers can describe a dental claims request once they know the fields, formats and labels they need.

For adjacent needs, see CMS-1500 and UB-04 claim form images, utilization management review records, insurance claims AI evaluation, the industry-specific operational data hub, our medical coding and claims datasets, healthcare administration buyers and what AI companies build with dental practices data.

This page is general information, not legal advice. Confirm requirements with counsel for your jurisdiction and use case.

Request dental claims data for claim review AI

SourceX sources operational datasets from US companies on request and serves AI teams wherever they are based. Every dataset is rights-reviewed and licensed with defined records, uses, term and delivery. Describe the dental claims, attachments and outcomes you need.