Workers' compensation and auto medical bill review data for bill-review AI

This guide is for ML leads at TPAs, bill-review vendors and carriers who are scoping training or evaluation data. It sits alongside the industry-specific operational data hub and is distinct from payer medical-necessity work covered in utilization management review records. Health-plan claims with CARC/RARC remittances are covered on medical coding and claims data.

Why P&C bill review data differs from health-plan claims

P&C bill review is priced by state workers' compensation fee schedules, PPO network contracts and jurisdiction-specific edits, not by a health plan's benefit design. A California physician bill, for example, is priced under the state's Official Medical Fee Schedule, which California's Division of Workers' Compensation updates through dated orders with their own effective dates. Texas, New York, Florida and other states each run separate schedules, ground rules and dispute paths.

That means the label on a line is jurisdiction-dependent and time-dependent. The same CPT code on the same date can carry a different allowance in two states, and the same state can change its conversion factor, NCCI-style edit tables or facility rules mid-year. A model trained without the jurisdiction and fee schedule version will learn noise. Auto PIP adds another layer, since some no-fault states tie medical payments to their workers' compensation schedules while others use a percentage of Medicare or usual-and-customary benchmarks.

The bill and EOR fields that carry the labels

The inputs are the bill as submitted and the labels are what the reviewer returned on the EOR. Ask for both at line level, joined by a stable bill and line identifier.

• Bill header: form type (CMS-1500, UB-04, ADA dental or a pharmacy invoice), billing and rendering provider NPI and taxonomy, place of service or type of bill, jurisdiction state, date of injury, claim type (indemnity, medical-only, PIP), and receipt date.
• Lines: CPT/HCPCS with modifiers, revenue codes for facility bills, NDC for pharmacy, units, billed charge, ICD-10-CM diagnosis pointers, and dates of service.
• Review output: allowed amount per line, reduction amount split by cause (fee schedule, PPO discount, clinical or coding edit, duplicate, unrelated to compensable injury, missing documentation), EOR reason code and free-text explanation.
• Rule provenance: fee schedule name and effective version, PPO network and contract tier, edit library and version, and whether a manual reviewer overrode the automated result.
• Downstream outcomes: reconsideration requests, second-level review results, state medical fee dispute filings and final paid amount.

Many states collect medical bill payment data from claims administrators through the IAIABC EDI medical bill reporting standard [7], often with state-specific variations. If a supplier already reports this way, its extract will use familiar data element numbers, which makes field mapping across suppliers far easier. EOR reason codes are another matter: most bill-review systems use proprietary code sets, sometimes crosswalked to the CARC and RARC lists that health payers use on 835 remittances [1]. Ask for the code dictionary as of each period, because those lists are revised on a recurring schedule [1].

A line-level record a bill-review model can learn from

A usable record keeps the rule, the version and the outcome on every line, so you can train reduction-reason classifiers and evaluate allowance predictions against the schedule in force on the date of service.

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

{
  "bill_id": "B-7Q2X",
  "line_no": 3,
  "jurisdiction": "CA",
  "claim_type": "workers_comp_medical_only",
  "form_type": "CMS-1500",
  "date_of_service": "2026-03-14",
  "date_received": "2026-03-29",
  "provider_npi_token": "PRV_8812",
  "procedure": {"code": "97110", "modifiers": ["GP"], "units": 4},
  "diagnosis_pointers": ["S83.511A"],
  "billed_amount": 260.00,
  "fee_schedule": {"name": "CA OMFS physician services", "version_effective": "2026-03-01"},
  "ppo": {"network_token": "NET_04", "applied": false},
  "edits_applied": [{"library": "vendor_edits", "version": "2026Q1", "edit": "units_cap"}],
  "allowed_amount": 148.20,
  "reductions": [
    {"cause": "fee_schedule", "amount": 74.10},
    {"cause": "units_exceed_limit", "amount": 37.70}
  ],
  "eor_reason_codes": ["FS01", "UN04"],
  "eor_reason_text": "Reduced to fee schedule; units exceed allowed per visit",
  "manual_override": false,
  "reconsideration": {"requested": true, "result": "partial_reversal", "additional_paid": 37.70},
  "final_paid": 185.90
}

The reconsideration block is what turns a reduction dataset into an appeal-triage dataset. Without it, a model can only imitate first-pass review, including the first-pass errors that providers later overturn.

Appeals and reconsiderations as second-level labels

Reconsideration and dispute outcomes are the most valuable and most often missing part of a bill review dataset. They tell you which reductions held, which were reversed, and on what grounds, which is the signal you need to predict overturn risk or prioritize a reconsideration queue.

Request each appeal as its own event linked to the original bill line: request date, provider argument category, added documentation (op notes, invoices for implants, prior authorization), reviewer decision, amount restored and decision date. Where disputes escalate to a state process, such as independent bill review in California [6], capture the final determination separately from the carrier's internal reconsideration. Expect strong class imbalance; most lines are never appealed, so plan sampling and evaluation metrics accordingly.

Fee schedule versions and other failure modes

The most common failure in bill review data is a label that cannot be reproduced because the fee schedule, PPO contract or edit library version was not recorded. Fee schedules change by state on their own calendars, and a single state can carry separate effective dates for physician, hospital outpatient and pharmacy sections.

Other failure modes to check before you train:

• Ambiguous reduction cause. A single "reduced per review" code that bundles fee schedule, PPO and edit reductions makes reason prediction impossible.
• Overwritten amounts. Some systems store only the final allowed amount after a reconsideration, which erases the first-pass label.
• Duplicate bills. Resubmissions and corrected bills need parent-child links, or duplicate denials will look like independent decisions.
• Mixed payers. Group health, PIP and workers' compensation lines on the same provider can be mislabeled if claim type is missing.
• Pharmacy and DME drift. NDC-based pricing and repackaged drug rules often follow different logic than professional services; segment them.
• Vendor code drift. Proprietary EOR codes get retired and reused; without a dated dictionary, label meaning changes silently.

Privacy and de-identification for P&C medical bills

Bill review data contains health information even when the payer is a P&C carrier, so plan for HIPAA-grade de-identification and state confidentiality rules. Workers' compensation insurers are generally not HIPAA covered entities, but the treating providers who generate the bills generally are, and state workers' compensation laws add their own confidentiality limits; have counsel confirm how these apply to the specific supplier and states.

Expect suppliers and your own reviewers to ask which HIPAA method was used. HHS describes two: Safe Harbor, which removes 18 identifier types including those of employers and relatives, and Expert Determination, in which a qualified expert finds the re-identification risk very small [2][3]. Employer identity matters here, because a workers' compensation claim carries the employer by design. Dates of injury and service, small-town providers and rare procedures are also quasi-identifiers; NIST's survey documents how de-identified data has been re-identified [5]. If any bills involve substance use disorder treatment from a federally assisted program, 42 CFR Part 2 applies, with a compliance date of February 16, 2026 for the 2024 final rule [4]. For reviewing the expert's report itself, see how to review a HIPAA Expert Determination report.

Request checklist for a bill review dataset

A good request names the jurisdictions, claim types, form types, label fields and outcome fields you need, plus the version metadata that makes labels reproducible.

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

For the delivery and security side of the deal, use the training data supplier security review checklist.

How SourceX approaches bill review data requests

SourceX sources operational datasets from US companies on request and manages the commercial process, including licensing agreements and ongoing purchases. Bill review records are not held in stock, and a request does not guarantee a match. You describe the data you need, and SourceX looks for US businesses that hold it; every release is approved by the supplying company. You can describe your bill review data needs at any stage of scoping.

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. Personal details such as names, phone numbers and account numbers are removed or replaced before delivery, the method is recorded and a sample is checked, though no method is perfect; health records require HIPAA de-identification by Safe Harbor or Expert Determination. Delivery runs through private, access-controlled workflows after an executed agreement and supplier approval. See also claims administration buyers, insurance claims datasets and healthcare revenue cycle datasets.

Related reading: bodily injury claim valuation data, legal billing and bill-review adjustments, and the AI data hub.

Request workers' comp and PIP bill review data

SourceX sources operational data such as finance workflows and documents from US companies on request, with every dataset rights-reviewed and delivered under a license that defines records, uses, term and delivery. Diligence materials on source, rights, preparation and allowed use are prepared per dataset. Tell SourceX what bill review data you need.

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